Improve performance and gain room to grow by easily migrating to a modern OpenShift environment on Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC processors and high-speed 100GbE Broadcom NICs

We deployed this modern environment, then migrated database VMs from legacy servers and saw performance improvements that support consolidation.

Transactional databases are the backbone of many business operations, powering ecommerce and order fulfillment, human resources and payroll, and a host of other activities. If your company is running these kinds of workloads on server infrastructure that is several years old, you might believe that performance is adequate and that you have little reason to consider upgrading to new servers with modern processors, networking, and a Red Hat^® OpenShift^® container-based environment. In fact, by continuing to use this older gear, you could be incurring higher than necessary operating expenditures by maintaining and powering more servers than you need to perform a given volume of work. You could also be risking downtime with aging hardware that is likelier to break down. By upgrading to a modern environment, you could mitigate these issues and future-proof your infrastructure. A 2019 Forrester Consulting report recommended that organizations refresh their servers at least every three years to maximize agility and productivity.^[1] The report states not only that modern servers allow organizations to adopt more emerging technologies at a faster rate, but also “modern hardware has a profound impact on business benefits such as better customer experience, employee productivity, and innovation.”^[2]

We explored the process of migrating VMs from a legacy environment and conducted testing to quantify the resulting improvements in network and database performance. We started with a legacy environment consisting of MySQL^™ virtual machines (VMs) running on a cluster of three Dell^™ PowerEdge^™ R7515 servers with 3^rd Generation AMD EPYC^™ processors and 25Gb Broadcom^® NICs. We then deployed a modern OpenShift container-based environment comprising three Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC processors and high-speed 100Gb Broadcom NICs. While the primary application of OpenShift is typically for containerized workloads, we used OpenShift Virtualization, which presents a familiar VM layer to administrators while utilizing the containerized technology on the underlying layer. Both environments used a Dell PowerStore 1200T for external storage that the servers accessed using iSCSI. We measured database performance using the HammerDB TPROC-C benchmark.

We found that the modern cluster environment of Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC processors and high-speed 100Gb Broadcom NICs outperformed the legacy cluster environment, delivering 44 percent greater database performance. These improvements mean that companies that upgrade can enjoy savings by meeting their workload requirements with fewer servers to license, maintain, power, and cool. Selecting 100Gb Broadcom NICs also positions companies well to take advantage of increasingly popular network-intensive technologies such as artificial intelligence (AI).

The benefits of containerization and Red Hat OpenShift Virtualization

Many organizations choose containers for DevOps due to their easy scalability and portability. Because a container encapsulates an application as well as everything necessary to run that application, it’s simple to move the container from development to test and production environments, adding instances of the application by replicating the container. Containers can also be useful for microservices, data streaming, and other use cases.^[3]

Containers aren’t necessarily ideal for every use case, however, and for some infrastructures, IT teams may wish to incorporate both containers and VMs. Red Hat OpenShift Virtualization, which we used in our testing, enables organizations to run both VMs and containers on the same platform by bringing VMs into containers.^[4] This lets IT reap the benefits of both containers and VMs with the efficiency benefit of relying on one management tool, rather than having to maintain two distinct infrastructures.

About our testing

We explored the process of deploying a modern data center environment and migrating VMs to it from a legacy environment. We also measured the database performance the VMs achieved in both environments:

Legacy environment

• Three Dell PowerEdge R7515 servers with 3^rd Generation AMD EPYC 7663 56-core processors and Broadcom Advanced Dual 25Gb Ethernet NICs
• External storage using Dell PowerStore 1200T over iSCSI
• VMware^® vSphere^® 8 

 Modern environment

• Three Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC 9554 64-core processors and Broadcom NetExtreme-E BCM57508 100GB NICs
• External storage using Dell PowerStore 1200T over iSCSI
• Red Hat OpenShift 4.14

 Figure 1 presents a diagram of our test configuration. In addition to our test server clusters, we needed three servers to host infrastructure VMs, workload client VMs, and the OpenShift control node VMs. We configured a Dell PowerEdge R7525 to serve as the host for our infrastructure VMs for services such as AD, DHCP, and DNS, as well as HammerDB client VMs. We also configured a Dell PowerEdge R7625 to host additional HammerDB client VMs. For the OpenShift environment, we deployed a Dell PowerEdge R540 to host the OCP control nodes. We virtualized the control nodes to reduce the number of servers needed for the test bed.

Figure 1: Our test configuration. Source: Principled Technologies.

To test the MySQL database performance of each environment, we used the TPROC-C workload from the HammerDB benchmark. HammerDB developers derived their OLTP workload from the TPC-C benchmark specifications; however, as this is not a full implementation of the official TPC-C standards, the results in this paper are not directly comparable to published TPC-C results. For more information, please visit https://www.hammerdb.com/docs/ch03s01.html.

Each VM had a single MySQL instance with a TPROC-C database. We targeted the maximum transactions per minute (TPM) each environment could achieve by increasing the user count until performance degraded.

What we learned

Finding 1: Deploying OpenShift in the modern environment was easy

For our environment, the OpenShift installation process using the Red Hat Assisted Installer to install an OpenShift Installer-Provisioned Cluster was straightforward and simple. We started by setting up the prerequisites for the environment, which included a VM for Active Directory, DNS, and DHCP. We created a domain for our private network and added the API and ingress routes as DNS A records. Next, we set up a VM as a router so that our OpenShift environment could access the internet from our private network. Finally, we created three blank VMs to serve as our OpenShift controller nodes. Once we had met the pre-requisite requirements, we logged into the Red Hat Hybrid Console and navigated to the Assisted Installer to create the cluster.

The Assisted Installer streamlined the process by walking us through configuration menus for storage, network, and access to the cluster. We started the cluster creation by assigning it a name, providing the domain, and selecting an OpenShift version. From there the installer guided us through the process of providing an installer image using the SSH public key of the server running the installer. After downloading the ISO, we booted each of the controller and worker nodes into the image and the Assisted Installer discovered each node. After discovering the controller and worker nodes, the installer walked us through the rest of the configuration process and then began the installation. The Assisted Installer made the process very simple with only six configuration tabs to advance through, and with our total install time after configuration taking around three hours. Once the installation was complete, each node rebooted into the OpenShift OS and the Assisted Installer provided us with a cluster console fully qualified domain name (FQDN) to connect to and manage the cluster from. For detailed steps on the OpenShift deployment process, see the science behind the report.

Finding 2: Migrating VMs from the legacy VMware environment to the modern OpenShift environment was easy

Migrating a VM from the VMware environment to OpenShift was also a straightforward process and quick to set up. While the actual migration time will vary depending on VM size and hardware speed, the setup consists of only a few steps and took us less than 10 minutes. We first installed the Migration Toolkit for Virtualization from the OpenShift OperatorHub. We then entered the IP address and credentials for the vCenter as a new provider. Next, we created a NetworkMap and a StorageMap to connect the respective resources between the environments. We then created a new migration plan to map the VMs to a namespace in OCP. We ran the migration plan on a single VM, and confirmed that we were able to enter the VM console once the migration was complete. For detailed steps on the process of migrating VMs from the legacy environment to the modern environment, see the science behind the report.

About 4^th Gen AMD EPYC 9554 processors

According to AMD, EPYC 9554 processors deliver fast performance “for cloud, enterprise, and HPC workloads—helping accelerate your business.”^[5] EPYC processors include AMD Infinity Guard, which per AMD is “a set of layered, cutting-edge security features that help you protect sensitive data and avoid the costly downtime cause by security breaches.”^[6]

In addition to performance and security features, AMD claims their processors are energy-efficient, which can reduce energy costs and “minimize environmental impacts from data center operations while advancing your company’s sustainability objectives.”^[7]

When comparing SPECCPU Floating Point peak rates and the default thermal design power (TDP) of the AMD EPYC 9554 and the AMD EPYC 7663, the 9554 has 54 percent better performance per watt, which demonstrates the improved power efficiency with the new 4th Gen AMD EPYC process.[8]^,[9]

For more information about 4^th Gen AMD EPYC processors visit: https://www.amd.com/en/processors/epyc-server-cpu-family.

Finding 3: Database performance improved by 44 percent in the new environment

Figure 2 shows the results of our database performance testing using the TPROC-C workload from the HammerDB benchmark suite. The modern OpenShift cluster of Dell PowerEdge R7615 servers outperformed the legacy cluster by 44 percent. This extra capability could benefit companies upgrading to the new environment in several ways. The company could provide a better user experience, perform more work—or support more users—with a given number of servers, or reduce the number of servers necessary to execute a given workload.

Figure 2: Performance in transactions per minute using the TPROC-C workload of the HammerDB benchmark suite. Higher is better. Source: Principled Technologies.

Finding 4: Performance improved in the modern cluster, supporting consolidation, which leads to savings

Based on the results of our performance tests (see Figure 3), a company could consolidate the database workloads of a four-node Dell PowerEdge 7515 cluster with some additional headroom into three modern Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC processors and high-speed 100Gb Broadcom NICs.

The cluster of three modern servers delivered a total of 9,674,180 transactions per minute (3,224,726 TPMs per server). The cluster of three legacy servers delivered a total of 6,714,712 TPM (2,238,237 per server). Based on these results, four legacy servers would achieve a total of 8,952,948 TPM, which would leave 721,231 additional TPM room for growth on the modern three-node cluster.

Reducing the number of servers you need means that operational expenditures such as data center power and cooling and administrator time for maintenance also decrease, leading to ongoing savings.

Figure 3: Performance in transactions per minute that three modern servers and four legacy servers could achieve, based on our hands-on testing. Higher is better. Source: Principled Technologies.

About Dell PowerEdge R7615 servers

The Dell PowerEdge R7615 is a 2U, single-socket rack server. Dell states that it has designed this server to provide “performance and flexible, low-latency storage options in an air or Direct Liquid Cooling (DLC) configuration.”^[10]

According to Dell, this server uses the AMD EPYC 4^th generation processor to deliver up to 50 percent higher core count per single-socket platform in an innovative air-cooled chassis.^[11] It also supports DDR5 at 4800 MT/s memory and PCIe^® Gen5 with double the speed of previous Gen4 for faster access and transport of data, optimizing application output.^[12] It supports up to six single-wide full-length GPUs or three double-wide full-length GPUs, to improve responsiveness or reduce app load time for power users, plus lower-latency, high-performance NVMe SSDs to help maximize compute performance.^[13]

Learn more at https://www.delltechnologies.com/asset/en-us/products/servers/technicalsupport/poweredge-r7615-spec-sheet.pdf.

How high-speed 100Gb Broadcom NICs can help your organization

Even if a 25Gb NIC is sufficient to meet a company’s current networking needs, opting to equip new servers with the high-speed 100Gb Broadcom NIC can be a smart move. Future-proofing your network can allow you to meet the increasing demands of emerging technologies.

Advanced technologies such as artificial intelligence and machine learning, which can require the processing and transmission of large amounts of data, are becoming increasingly prevalent across businesses of all sizes. In a June 2023 survey of small business decision-makers, 74 percent were interested in using AI or automation in their business and 55 percent said their interest in these technologies had grown in the first half of 2023.^[14] Upgrading to a modern environment with a highspeed 100Gb Broadcom NIC positions companies to take advantage of AI applications for social media, content creation, marketing, customer support, and many other use cases.

Another way that investing in the high-speed 100Gb Broadcom NIC can help your company is through improved efficiency. You might be tempted to go with a 25Gb NIC, thinking that as your networking needs increase, you can simply add more NICs of this size. However, consider a 2023 Principled Technologies study that compared the performance of a server solution with a 100Gb Broadcom 57508 NIC and a solution with four 25Gb NICs.^[15] Testing revealed that the 100Gb NIC solution achieved up to 2.3 times the throughput of the solution with 25Gb NICs. It also delivered greater bandwidth consistency, which can translate to providing a better user experience; the report states that applications using the 25Gb NICs network configuration “would experience significant variation in available bandwidth, potentially causing jittery or interrupted service to multiple streams.”^[16]

About the Broadcom BCM57508-P2100G Dual-Port 100GbE PCle 4.0 ethernet controller

A higher performing NIC can reduce latency, increase throughput, and allow the server to transmit and receive a great volume of data. The Dell PowerEdge R7615 we tested features the Broadcom BCM57508-P2100G DualPort 100GbE PCle 4.0 ethernet controller, which supports speeds of up to 200 Gigabits per second. Broadcom designed the BCM57508-P2100G “to build highlyscalable, feature-rich networking solutions in servers for enterprise and cloud-scale networking and storage applications, including high-performance computing, telco, machine learning, storage disaggregation, and data analytics.”^[17]

The BCM57508-P2100G features BroadSAFE^® technology, “to provide unparalleled platform security” and a “unique set of highly-optimized hardware acceleration engines to enhance network performance and improve server efficiency.”^[18]

BCM57508-P2100G Dual-Port 100GbE PCle 4.0 ethernet controller. Image provided by Dell.

Conclusion

If your organization’s transactional databases are running on gear that is several years old, you have much to gain by upgrading to modern servers with new processors and networking components and an OpenShift environment. In our testing, a modern OpenShift environment with a cluster of three Dell PowerEdge R7615 servers with 4^th Generation AMD EPYC processors and high-speed 100Gb Broadcom NICs outperformed a legacy environment with MySQL VMs running on a cluster of three Dell PowerEdge R7515 servers with 3^rd Generation AMD EPYC processors and 25Gb Broadcom NICs. We also easily migrated a VM from the legacy environment to the modern environment, with only a few steps required to set up and less than ten minutes of hands-on time. The performance advantage of the modern servers would allow a company to reduce the number of servers necessary to perform a given amount of database work, thus lowering operational expenditures such as power and cooling and IT staff time for maintenance. The high-speed 100Gb Broadcom NICs in this solution also give companies better network performance and networking capacity to grow as they embrace emerging technologies such as AI that put great demands on networks.

This project was commissioned by Dell Technologies.

May 2024

Principled Technologies is a registered trademark of Principled Technologies, Inc. 

All other product names are the trademarks of their respective owners. 

Read the report on the PT site at https://facts.pt/2V6p3FG and see the science at https://facts.pt/Dj53ZJb.

Author: Principled Technologies

Introduction

Recent advances in AI have significantly accelerated interest in the technology and how it can be applied to revolutionize processes across many different industries. One such industry that is well positioned to benefit from leveraging AI is healthcare. AI is increasingly being used to assist in medical diagnostics, specifically to improve the accuracy and speed of diagnoses made by radiologists and physicians. This paper presents a proof-of-concept (PoC) solution that utilizes an AI image classification model to quickly and precisely detect pneumonia from patient X-ray images.

The PoC shows the practicality of bringing AI into image analysis for healthcare, and sets the stage for healthcare organizations to quickly adopt and deploy AI solutions. Building on the success of the pneumonia detection PoC, the approach can be further extended to modalities such as CT-Scans, MRIs, and others. The PoC overcomes common challenges found both generally in new AI deployments and more specifically in healthcare environments by leveraging and optimizing common CPU-based hardware, customizing a model for a healthcare-specific use case, and deploying a secure, on-premises solution to address healthcare data regulations and privacy concerns.

The PoC was deployed on standard Dell^TM PowerEdge hardware with 64 core 4th Gen AMD^TM EPYC CPUs. The PoC demonstrated impressive performance for both model training and inferencing, without requiring GPUs. Model training was completed in 9 hours with a validation accuracy of 85%. The inferencing process achieved a throughput of 337 images per second with a default configuration. By utilizing additional performance optimizations, the deployment achieved a 4X performance increase, resulting in a throughput of 1,390 images per second.

Importance for the Healthcare Market

Technology advances in healthcare drive greater efficiency and accuracy in medical processes and ultimately improve patient outcomes. Such technology advancements are important to moving the healthcare industry forward.

AI in particular is a technology with great potential to create significant value in the healthcare sector. The possibilities for applying AI technology to healthcare may extend to a wide range of healthcare related areas including pharmaceutical research, hospital workflows, and patient experience. Potential uses for AI in healthcare include AI accelerated drug development, predictive analytics for disease prevention, medical focused chatbots, and intelligent hospital staffing systems.  

Key to adoption of any technology in healthcare is maintaining data privacy and security due to the handling of sensitive patient information and the heavily regulated nature of the industry.  When considering AI, organizations will address this challenge by implementing on-premises solution deployments to maintain control over their private data.

New AI applications must also be capable of integrating with existing technologies, processes, and equipment common to healthcare. Additionally, to adopt AI quickly, healthcare organizations will want to leverage existing hardware or utilize readily available commodity hardware. These challenges may cause uncertainty for organizations who are unfamiliar with AI when planning new deployments, delaying adoption of AI innovation.

The PoC presented in this paper demonstrates an AI implementation that addresses these challenges to assist medical organizations in swift adoption of AI. The PoC serves as a template for using AI to improve diagnoses based on X-ray images. In addition to the benefit of rapid image analysis, the PoC highlights the ability to quickly train an accurate AI model by utilizing a healthcare specific dataset. Notably, the same Dell PowerEdge server was used for both training and inference.

With continued advancements in the field of AI, healthcare executives will recognize the opportunity the technology holds in improving healthcare services and look for ways to leverage it. Developers and IT operations must understand the systems, processes, and effort required for a successful AI deployment, especially when considering vertical specific applications, such as healthcare.

Solution Overview

To demonstrate a practical implementation of a healthcare focused AI solution, Scalers AI^TM, in partnership with Dell, Broadcom^TM, and The Futurum Group, implemented an AI-powered system for detecting pneumonia. The solution utilizes the ResNet50 image classification AI model that was fine tuned to recognize pneumonia in X-ray images.

This PoC showcases the ability for AI to assist doctors in patient diagnoses. Relying solely on human evaluation of X-ray results can lead to delays, due to doctors’ availability and bandwidth, or misdiagnoses due to human error. Delayed diagnosis, in particular, is a significant issue in healthcare, and it has been found to be a leading cause of patient injury claims concerning medical imaging^[1].

These issues are becoming increasingly challenging as the number of patients requiring medical imaging is growing. Meanwhile, hospitals are facing a global shortage of radiologists in the workforce². While AI does not possess the medical expertise required to replace human doctors or radiologists, it provides valuable characteristics that can be leveraged by medical professionals to enhance the efficiency and accuracy of the diagnosis process.

AI image classification models can rapidly detect specific features within images, such as the presence of pneumonia in chest X-rays, and classify them with a high level of accuracy. This approach can be used to quickly identify issues within large amounts of medical images, and in some cases identify issues that may otherwise go undetected. AI provides an efficient and accurate initial classification of images, which will be further analyzed by medical professionals to make a final diagnosis. By leveraging AI to augment the diagnosis processes, medical professionals can provide faster and more accurate diagnoses, ultimately resulting in quicker time to treatment and improved patient results.

Figure 1: AI Medical Image Analysis PoC Solution (Source: Scalers.AI)

To achieve an AI-powered pneumonia detection system, the PoC integrates an image classification AI pipeline with a standard DICOM server for storing and managing medical imaging data. The DICOM server in the PoC stores X-ray images of potential pneumonia cases. The AI pipeline evaluating the X-ray images consists of two additional components – an AI scheduling service and an inferencing server. The AI scheduling service identifies new images, batches them, and sends them to the inferencing server. The inferencing server utilizes a customized version of the ResNet50 AI model, deployed using AMD’s Unified Inferencing Frontend (UIF). X-ray images are inferenced to provide a binary classification regarding the detection of pneumonia. The categorized images are then made available for review with a medical imaging viewer, and returned to the DICOM server.

More information about the specific solution components can be found below:

• DICOM Server and Storage: For the PoC, Orthanc open source DICOM server software was used to manage the medical images.
• AI Service and Scheduler: The AI service and scheduler identify new images, groups them for individual patient evaluation, and send them over Ethernet for analysis. The AI Service then handles the analysis results and sends annotated images back to the DICOM server.
• AI Model Server: The AMD UIF inference engine on the AI model server is an open-source tool that deploys models and makes them available for inferencing. The server is compatible with specific models designed for AMD CPUs.  
• AI Pneumonia Classification Model: The ResNet50 model serves as the foundation for the AI image classification. The model was customized through a transfer learning process using chest X-ray images from the NIH Clinical Center. Transfer learning involves using pre-trained models as a starting point for creating new models, which are further trained for different tasks. In this case, ResNet50 was used to train the model to recognize pneumonia in DICOM images. ResNet50 was chosen for its support of the AMD UIF with ZenDNN Model Zoo. Additionally, ResNet50 is well suited for image detection in critical medical use cases as it is a highly accurate model.

 Figure 2: AI Image Classification Software Overview

The AMD ZenDNN library is used to provide performance optimizations. ZenDNN is a library with APIs designed to accelerate deep learning inference applications on AMD CPUs, aiming to improve performance. ZenDNN performance guide recommendations, along with node pinning and core pinning, were used to optimize the performance of AMD EPYC processors used in the PoC.

• Medical Image Viewer: The PoC used an Open Health Imaging Foundation (OHIF) viewer for viewing the DICOM images. Although radiologists might have distinct preferences for viewers, the PoC specifically used the OHIF viewer as a it is an open source, web-based medical imaging viewer, however, other DICOM viewers could be used.

Additional details about the implementation and performance testing of the PoC have been made available by Dell on GitHub.

Highlights for AI Practitioners

This PoC is notable for AI practitioners as it demonstrates a practical AI application that can be used to enhance healthcare environments. Key to the practicality of the solution is that it utilizes readily available, CPU based hardware, rather than relying on GPUs. A core component of achieving this type of CPU-based approach is utilizing software libraries to simplify and optimize the deployment. The AMD Unified Inference Frontend (UIF) was utilized to easily deploy a model that was optimized to run on AMD EPYC CPUs. While this PoC intentionally utilized a CPU-based deployment to demonstrate running useful AI applications on easily accessible hardware, the use of a model from the UIF model zoo is notable, as the UIF models are transportable across AMD technology stacks. This provides flexibility for organizations who may incorporate GPUs in future deployments as they further expand their use of AI.

AI practitioners should additionally note the performance enhancements that were achieved by utilizing the ZenDNN library, along with core pinning and node pinning configurations. These configurations demonstrated up to a 4X throughput increase, showcasing how the use of software optimization libraries can be leveraged to provide significant inferencing performance without hardware acceleration. Figure 3 shows the ZenDNN parameter configurations utilized.

Figure 3: ZenDNN Configurations

AI practitioners should note that the CPU-based deployment was not only utilized for inferencing. The same Dell PowerEdge server and AMD processor was used for model training. The solution utilizes a pre-trained base model, ResNet50, customized with a transfer learning process. Transfer learning utilizes the foundation of a pre-trained model’s capabilities, and provides further customization to support a new, specific task. In this case, transfer learning was used to teach the ResNet50 image classification model to detect pneumonia in X-ray images. This was done by training the model with a dataset of 29,687 X-ray images. The total training process was completed in around 9 hours, and resulted in 99% training accuracy and 85% validation accuracy. The accuracy of the model is especially critical in this type of medical deployment, as the model is responsible for assisting in the diagnosis of medical patients, and can have a direct impact on patient outcomes. The PoC demonstrates the ability to utilize common CPU-based infrastructure along with pre-trained models for efficient, yet accurate AI model training.

Considerations for IT Operations

This AI implementation is notable for those working in IT operations because it demonstrates an achievable AI deployment that utilizes familiar, readily available hardware for both model training and production. IT operations staff will be very familiar with deploying Dell PowerEdge servers and Broadcom networking, and this PoC provides an example for organizations to understand how these familiar solutions can be leveraged for AI workloads.

The PoC leverages three Dell PowerEdge servers powered by 4th Gen AMD EPYC CPUs to deploy the Orthanc DICOM server, the AI scheduler, and the AI model server. The powerful AMD processors alongside large memory capacity make these servers well suited for AI workloads. This PoC leveraged Dell PowerEdge R7625 servers with AMD EPYC 9554 64-core processors, and 2.95 TB of memory. Additional server specifications can be found in figure 4 below.

Figure 4: Server Details

The Dell PowerEdge R7625 server provides a powerful platform that showcases the ability to run AI on CPUs. For IT operations, this lowers the barrier of entry for supporting AI, allowing them to utilize readily available hardware or leverage their existing infrastructure.

Another notable takeaway of the PoC is its ability to maintain data privacy and security, which are major concerns for IT organizations in the healthcare sector, due to the sensitive nature of medical data and regulations such as HIPAA and HITECH. Dell PowerEdge servers feature a cyber resilient architecture for zero trust IT environments with capabilities such as siliconbased root of trust, multi-factor authentication (MFA), and role-based access controls (RBAC).

The DICOM server, the AI scheduler, and the AI model server are connected with scalable, high bandwidth, Broadcom Ethernet. This high bandwidth connection is crucial to the solution’s ability to support the transmission of medical images, especially as the solution scales. While this PoC demonstrated image classification capabilities using relatively small X-ray images, by implementing a scalable connection, the PoC can be further extended to support larger image files such as MRIs or CT scans.

In addition to providing insight into AI hardware requirements, the PoC provides IT professionals with an understanding of software packages that can be utilized to build a healthcare focused AI solution. The PoC primarily utilized easily accessible, open-source software tools.

Key to deploying the AI model is the AMD Inference Server, which provides an open-source tool to easily deploy AI solutions on AMD hardware. The PoC additionally utilized open-source tools to support the medical imagery workflow, include Orthanc DICOM server and OHIF Viewer. Details of key software utilized, including version and licensing information can be found in figure 5 below.

Figure 5: Software Packages

Healthcare Solution Performance Observations

Key to the performance of this PoC is the throughput of images per second. Quick processing of X-ray images is vital to the solutions overall ability to accelerate patient diagnosis, leading to quicker treatment.

To demonstrate the performance of the PoC, the throughput of images per second were measured with an increasing number of processes, both with default settings as well as with configurations that optimized the performance of the AMD EPYC processor. The optimized variations included the use of the ZenDNN library alongside use of core pinning and node pinning. ZenDNN is a library that optimizes deep learning inferencing for AMD processors. Core pinning and node pinning are configurations that bind processes to specific cores or NUMA nodes.   The performance of each configuration can be seen in Figure 6.

Figure 6: Throughput Performance

The test results demonstrate the ability to significantly improve throughput performance by utilizing ZenDNN with either core pinning or node pinning. When running 64 processes the default configuration achieved a throughput of 337 images per second. Meanwhile, the configuration with ZenDNN and node pinning achieved a 3.9x improvement with 1,338 images per second, and the configuration with ZenDNN and core pinning configuration achieved a 4.1X improvement with 1,390 images per second. Figure 7 includes full testing results of the pneumonia detection PoC throughput performance testing.

Figure 7: Throughput Testing Results

For most hospitals, 1,390 images per second is likely well beyond their typical X-ray image processing requirements. This level of throughput is notable, however, because it provides flexibility for future adaptation of the solution to support more demanding data such as 3D images or other large data formats.

The performance improvements achieved by the ZenDNN configurations are also quite notable because they demonstrate the ability to optimize AI inferencing performance on a CPU. AI performance is often thought of as a hardware problem that requires GPUs or other specialized hardware to solve. This testing showcases the impact that software libraries, such as ZenDNN, can have in dramatically improving performance, even when using off-the-shelf CPU-based hardware. This type of optimization allows organizations to deploy powerful, high performance AI applications with either their existing hardware or readily available hardware, removing the barrier of acquiring GPUs and facilitating quick AI innovation.

Final Thoughts

Strategically applying AI in healthcare has great potential to enhance medical processes and improve patient outcomes, as demonstrated in this successful pneumonia detection PoC.  While the PoC example is focused specifically on pneumonia detection from X-ray images, the solution can be further expanded to analyze patient data from various modalities, allowing trained models to detect a wider range of conditions. The potential for AI to enhance the healthcare industry extends far beyond this type of AI-assisted diagnosis use case.

This PoC demonstrates an AI deployment that utilized off-the-shelf, CPU-based hardware, while providing impressive performance, and meeting the unique requirements of a medical-focused application. The results of this PoC, including the performance details, not only demonstrate a successful example of a healthcare-oriented AI application, but it also emphasizes the broader opportunity for AI to have an immediate impact on improving healthcare processes. AI will prove to be an innovative technology across many areas in healthcare, and healthcare providers should be motivated to adopt the technology quickly, both to maintain competitive advantage in the market and to improve overall patient treatment. By leveraging readily available hardware from Dell and Broadcom, along with the concepts demonstrated in this PoC, healthcare organizations can quickly deploy powerful, innovative new AI solutions.

Resources: 

Introduction

As part of Dell’s ongoing efforts to help make industry-leading AI workflows available to its clients, this paper outlines a sample AI solution for the retail market. The PoC leverages Dell^TM technology to showcase an AI-powered inventory management application for retail organizations.

AI technology has been in development for some time, but recent technological advancements have greatly accelerated AI’s ability to provide value across a wide range of enterprise applications. AI solutions have become a key initiative for many organizations. While the advancement of AI technology provides the basis for a diverse set of AI-powered applications, the specific requirements of different verticals provide distinct hardware and software challenges. IT organizations might be unsure of the technical requirements for deploying such a solution. This uncertainty may be due to unfamiliarity with AI, as well as an expectation that AI applications will require specialized hardware, often with limited availability.

This paper covers a solution specifically designed to capture the requirements of a retail-based AI deployment using a standard AMD^TM CPU for AI training and inference. The solution leverages hardware from Dell, AMD , and Broadcom^TM, to create a solution powerful enough to capture and analyze large-scale video data from cameras in retail environments, as well as flexible enough to scale to the unique needs of individual retail environments. Training of the model was achieved in two days, utilizing the same Dell PowerEdge server that is used for inferencing. The scalability of the solution was tested with up to 20 video streams. The PoC additionally demonstrates AI optimizations for AMD CPUs by utilizing AMD’s ZenDNN library. The utilization of the ZenDNN library, along with node pinning, resulted in an average throughput increase of 1.5x.

While the overall applications of AI in retail environments are much broader than the single inventory management solution outlined in this paper, the PoC demonstrates a framework for how IT organizations can quickly deploy an AI solution that delivers practical value in a retail environment by using readily available hardware.

Importance for the Retail Market

As with many other industries, the retail market has become increasingly data driven. Data can provide greater insight into areas such as customer behavior and product demand, as well as assist in optimizing operational areas such as procurement and inventory management. The emergence of AI technology provides even greater opportunity for valuable data-driven insights and optimizations within the retail industry.

Possibilities for retail-focused AI solutions include both customer experience (CX)-driven solutions and operations-focused applications. CX might be enhanced with personalized recommendation systems based on customer purchase trends, or virtual assistants capable of providing product recommendations for online retail experiences. Retail operations may be optimized through solutions such as AI-enhanced surveillance to detect fraud or theft, inventory management systems, or AI-powered product pricing systems.

These examples, as well as the more in-depth PoC study outlined in this paper, are a small subset of possible AI applications that may be implemented by retail organizations. While the exact solution implementations that are most appropriate may vary between organizations based on several factors such as location, size, type of goods sold, and distribution of online versus in-person sales, it is clear that AI applications can provide immense value in retail environments.

While a proactive approach to AI adoption may be beneficial to retail organizations, unfamiliarity with AI technology and the hardware and software components needed to deploy and optimize such solutions act as a barrier to adoption. The following solution demonstrates a PoC for an AI-powered retail inventory management system that can be quickly deployed and further expanded upon by retail organizations using commonly available hardware.

 Solution Overview

The retail inventory management solution addresses a common challenge in retail environments of inventory distortion. Without accurate and timely inventory management, retail organizations can be challenged with stock levels that are either too low or too high. Both situations can prove to be costly. Too much inventory requires additional storage, commitment of capital, and potential waste of perishable items. Conversely, too low of inventory can lead to customer dissatisfaction and loss of sales. In many cases, low inventory leads to customers purchasing at competitive retailers and may lead to overall loss of brand loyalty. By utilizing computer vision and object detection AI models to monitor and track inventory, retailers can achieve real-time insights into their stock to balance their inventory more appropriately and provide valuable insights back to suppliers.

To demonstrate a real-world example solution of an AI application that could be deployed to address such retail challenges, Scalers AI^TM, in partnership with Dell, Broadcom, and The Futurum Group, implemented a PoC solution for a retail inventory management system. The solution was designed to capture data from store cameras and use an object-detection AI model to monitor and manage product stock levels. The solution was capable of detecting products on store shelves, keeping track of inventory, and raising alerts of low or out of stock items.

All of this was accomplished using standard Dell PowerEdge servers with 32 core 4th Gen AMD EPYC processors and Broadcom networking. No GPUs were required. The CPU-based solution was further optimized with AMD’s Zen Deep Neural Network (ZenDNN) library, which provides optimizations for deep learning inferencing on AMD CPU hardware. AMD’s ZenDNN optimizations delivered an average of 1.5x increased throughput performance to the PoC. By utilizing modest, CPU-based hardware, this PoC solution demonstrates a clear example of a readily deployable and broadly applicable AI retail solution.

To achieve the solution, store shelves were configured in zones with the product names and corresponding x,y coordinate pairs that indicated the shelf location. The products, location, and the maximum capacity for each item were stored as JSON objects.

Highlights for AI Practitioners

It is notable for AI practitioners that the project was not limited to the deployment and inferencing of the AI model. The solution additionally involved customization of the pre-trained base model using a process known as Transfer Learning. The solution began with the SSD_MobileNet_v2 model for object detection, which was an ideal model for this PoC as it provides a one-stage object detection model that does not require exceptional compute power. The model was then customized via Transfer Learning with the SKU110K image data set. The training process involved 23,000 images and resulted in a mean average precision (mAP) of 0.7.  The training process was completed in approximately two days.

Figure 3: Object Detection Software Overview

It should also be noted that both the model training and deployment of the video pipeline solution were accomplished using the same 32 core Dell PowerEdge R7615 server. The PoC demonstrates the ability to achieve useful AI applications on CPU-based hardware that is commonly found in retail environments. The solution is further optimized for inferencing on AMD CPUs by utilizing AMD’s ZenDNN library and node pinning. The ZenDNN library provides performance tuning for deep learning inferencing on AMD CPUs while node pinning can further optimize the application by binding processes to dedicated compute resources.

The below table shows the ZenDNN parameter configurations used.

Figure 4: ZenDNN Configurations

Considerations for IT Operations

The hardware used in this AI application, including Dell PowerEdge R7615 servers with 4th Gen 32 core AMD EPYC 9354P Processors, Dell PowerScale NAS, Dell PowerSwitch Z9664, and Broadcom NetXtreme-E NICs, is familiar and available to IT operations, yet each component provides valuable characteristics needed to support this type of solution.

The Dell PowerEdge servers provide powerful 4th Generation AMD EPYC processors that are capable of supporting both the AI and application workloads, and the Dell PowerScale NAS provides a high-performance, highly scalable NAS storage system capable of handling large-scale video and image data. The solution is then tied together using Broadcom Ethernet capable of supporting the high bandwidth requirements of video streaming. Most notably, these components all provide scalability for IT organizations to further build out this application with more demanding requirements such as additional video streams or additional application logic.

Futurum Group Comment:  The specific use of Dell PowerEdge R7615 servers should be noted, as it demonstrates the ability to run AI workloads on standard hardware, commonly deployed in retail environments. While not considered a high-end compute server, the R7615 servers with mid-range 32 core 9354P Processors proved capable of all processes including model training, inferencing, and the separate visualization engine. This enables retail IT organizations to deploy such solutions without acquiring GPUs or requiring the datacenter level cooling needed for higher end servers. Additionally, by separating the architecture into separate video and visualization pipelines, the solution can be scaled to meet the size and performance requirements of a broad range of retail environments.

The on-premises deployment of this solution additionally enables IT operations to achieve their data security and data privacy requirements. While public cloud has been utilized for many early iterations of AI applications, data privacy becomes a concern for many organizations as they build further AI applications leveraging private data. By deploying this, or similar, retail solutions on-premises, IT operations have greater control over the privacy of their data, which may include sensitive consumer or product information. The on-premises deployment of this solution also offers a potential economic advantage in its ability to avoid cloud storage costs when storing large capacities of video data. It additionally avoids the high networking requirements of uploading many video streams to the cloud.  

Specifications of the Dell PowerEdge servers used in this PoC can be found in Figure 5

Figure 5: Dell PowerEdge Server Details

Retail Solution Performance Observations

A key performance metric for the retail inventory management reference solution is the throughput of images per second as they are streamed by the in-store video cameras, decoded, and inferenced by the video pipeline. Video data is a common source for AI applications in the retail market, due to the prevalence of existing cameras deployed in stores, and the value of information that can be obtained by the video data. Because of this, the throughput performance insights gained from this PoC can translate to additional retail solutions that rely on image processing.

To examine the performance of the 32 core AMD EPYC 9354P processor for data capture and inferencing, the video pipeline was tested both with and without ZenDNN performance tuning, as well as with core pinning and node pinning. ZenDNN is a library that optimizes the performance of AMD processors for deep learning inferencing applications.  The node pinning and core pinning are techniques offer optimization by binding processes to specific NUMA nodes or cores. The tests were run with up to 64 processes running on a 32 core server. The results of this testing can be seen in Figure 6.

Figure 6: Throughput Performance

The performance results demonstrate that the use of ZenDNN with node pinning can provide a dramatic increase in throughput, with mostly lower CPU utilization. On average, ZenDNN with node pinning achieved a throughput increase of approximately 1.5x. Further throughput increases were additionally achieved by utilizing core pinning. Full results can be seen in Figure 7.

Figure 7: Video Pipeline Throughput Test

The performance gains achieved with ZenDNN, core pinning, and node pinning demonstrate the ability to optimize CPUs for AI applications. Commonly, computationally demanding AI processes, such as the computer vision and object detection utilized in this PoC, are expected to require GPUs. Hardware alone, however, is not the only component that affects performance. Software such as ZenDNN plays a key role in optimizing the performance of the chosen hardware, as does configuration details such as utilizing core pinning or node pinning. By utilizing these configurations, organizations can achieve AI applications that meet their performance needs with a CPU-based solution utilizing readily available hardware.

The PoC solution was additionally tested with an increasing number of video streams to assess the bandwidth of the networked video pipeline and visualization service. 1080p video was streamed to the video pipeline where it was decoded and inferenced. It was then transmitted and received by the visualization pipeline to be encoded and shared.  The number of video streams was increased incrementally between 1 and 20 which resulted in an increasing bandwidth utilization. The bandwidth scaled from an average utilization of 1.65 Gbits/s and a max utilization of 3.4 Gbits/s with 1 stream, to an average utilization of 13.9 Gbits/s and a max utilization of 27.4 Gbits/s with 20 streams. An overview of the results can be seen in Figure 8.

Figure 8: Inventory Management System Bandwidth

Notably, the bandwidth does not increase linearly in relation to the number of streams, allowing the solution to scale as additional streams are needed. As the number of streams increases, however, the solution does experience a decrease in frames-per-second. While frames-per-second decreases, the overall utility of the solution is not significantly impacted. Higher frame rates are of greater importance when considering video with large amounts of motion, or when viewing quality is a major priority. In this particular solution, lower frame rates are acceptable as the focus is stationary store shelves, and real time viewing is not the primary use case.  Full results of testing the networked solution, including both bandwidth utilization and frames per second, can be seen in Figure 9.

Figure 9: Inventory Management System Bandwidth Test

The results of this performance testing demonstrate that the bandwidth of the networked servers is capable of scaling alongside more demanding video requirements. The separation of the video pipeline and the visualization service onto distinct servers allows the architecture to independently scale the compute resources for the two services. To capitalize on this architecture however, the networking between the servers must be capable of providing adequate bandwidth between the services. To do so, the PoC solution utilizes Broadcom BCM57508 NetXtreme-E Ethernet controllers capable of supporting up to 200GbE.  By utilizing a modular architecture that’s connected with scalable, high bandwidth networking, the retail inventory management PoC provides a flexible starting point for retail organizations to scale to their individual needs, including the number of video streams, FPS requirements, and additional application logic.

Final Thoughts

With the rapid development of AI technology, the retail market presents many opportunities to deploy valuable new AI-powered applications. With the broad range of value that AI can bring to retail environments, both in improving CX and optimizing store operations, retail organizations should look to be proactive in adopting the emerging technology.

As a new technology, there are many unknowns and misconceptions for those in IT who may be unfamiliar with AI deployments, complicating and delaying new AI applications. A common challenge faced by IT is the expectation that AI applications will require specialized hardware solutions that are inaccessible. The AI-powered retail inventory management solution outlined in this paper serves as a demonstration of a broadly applicable AI solution for retail that can be deployed on off-the-shelf hardware solutions. The Dell hardware solutions used in the PoC deployment were demonstrated to handle the high-bandwidth video requirements as well as the AI modeling and inferencing requirements without the use of purpose-built accelerators, GPUs, or custom hardware.

The PoC solution outlined in this paper additionally serves as a reference for retail organizations to quickly deploy their own inventory management solution. While the solution discussed in this paper is limited to a PoC, it was designed with scalability in mind for organizations to further develop and scale a solution for their needs.

The use of an AI-powered inventory management system can provide real value and cost savings to organizations by avoiding over- or under-stocking products. By using readily available hardware and reference solutions, the barrier of entry for deploying such an AI solution is dramatically lowered, allowing retail organizations to achieve quicker deployments of new AI applications and quicker time to value.

CONTRIBUTORS

Mitch Lewis

Research Analyst  | The Futurum Group

PUBLISHER Daniel Newman

CEO | The Futurum Group

INQUIRIES

Contact us if you would like to discuss this report and The Futurum Group will respond promptly.

CITATIONS

This paper can be cited by accredited press and analysts, but must be cited in-context, displaying author’s name, author’s title, and “The Futurum Group.” Non-press and non-analysts must receive prior written permission by The Futurum Group for any citations.

LICENSING

This document, including any supporting materials, is owned by The Futurum Group. This publication may not be reproduced, distributed, or shared in any form without the prior written permission of The Futurum Group.

DISCLOSURES

The Futurum Group provides research, analysis, advising, and consulting to many high-tech companies, including those mentioned in this paper. No employees at the firm hold any equity positions with any companies cited in this document.

ABOUT THE FUTURUM GROUP

The Futurum Group is an independent research, analysis, and advisory firm, focused on digital innovation and market-disrupting technologies and trends. Every day our analysts, researchers, and advisors help business leaders from around the world anticipate tectonic shifts in their industries and leverage disruptive innovation to either gain or maintain a competitive advantage in their markets. 

Dell PowerEdge R7625 Rack Server & Emulex LPe36002 Host Bus Adapter

Dell R7625 and 64GFC Combine to Accelerate Oracle Analytics Workloads

Tolly Report #224106

Tolly test report demonstrating that Dell PowerEdge R7625 Rack Server outfitted with the Emulex LPe36002 64G Fibre Channel Host Bus Adapter can improve application performance up to 4x vs older generation 32/16G FC technologies.

Executive Summary

New generation technology can be expected to improve performance. There are times, however, when multiple technology advances can combine to provide an outsized advantage. Such is the case when the Dell PowerEdge R7625 Rack Server with AMD EPYC processors is combined with the Broadcom Emulex LPe36002 64G Fibre Channel Host Bus Adapter.

Dell commissioned Tolly to benchmark the analytics workload performance of the Broadcom Emulex LPe36002 64G Fibre Channel dual-port host bus adapter (HBA) running in the Dell PowerEdge R7625 Rack Server. Specifically, this report will focus on illustrating two points: 1) Improved database analytic performance due to the increased input/output (I/O) throughput of 64GFC, 2) Increased application performance when paired with PCIe 4.0/5.0 and the dual-port 64GFC HBA.

Tests showed that the new R7625 AMD EPYC platform's increased CPU power and PCIe 5.0 bus work in conjunction with the Broadcom 64GFC dual-port adapter to deliver line rate, 64G throughput that cannot be matched by earlier generation technology. See Figure 1.

Overview

The goal of these tests, as noted, was to illustrate, simply, that deploying a Dell PowerEdge R7625 Rack Server, powered by AMD EPYC processors, with the Emulex 64G Fibre Channel HBA can improve database analytic performance by providing double and quadruple the I/O throughput of the two prior generation HBAs respectively. Similarly, the tests were used to illustrate the key role of the newer-generation PCIe 5.0 server bus and PCIe 4.0 dual-port 64GFC HBA in increasing server I/O throughput. 

All benchmarking was done using the open source TPROC-H analytics workload of HammerDB. The tests were run using the Oracle 19c database environment but the results are generally applicable to any database or other input/output intensive workload. 

The TPROC-H workload measures how long it takes to run a series of 22 different types of decision support queries.   This type of workload is “read only” with no database updates taking place. The Linux iostat utility was used to measure storage I/O throughput.

Test Background & Results

64GFC vs 16/32GFC

This test was run three times with the only variable being the link speed between the server’s FC HBA and the switch.

Figure 1 (main and inset), on the previous page, summarizes all three tests using two metrics: storage I/O throughput and query execution time as reported by the HammerDB database benchmark. What is important to note are the relative results across the three scenarios. The 16GFC HBA is clearly a bottleneck (blue dots) taking the longest to complete and delivering the lowest throughput. (Note: multiple colors in the inset bar chart represent the different transaction types used in the TPROC-H benchmark.)

Performance is improved, roughly by 2x, when the HBA is configured for 32GFC (gray dots) but, as will be seen, 32GFC still presented a transaction bottleneck.

When run using the 64GFC the database storage IO throughput is the highest and the query execution time is the shortest. Again, performance is improved roughly by a factor of two over the 32GFC results. 

64GFC Dual-Port HBA Performance

The Emulex LPe36002 64GFC HBA is a PCIe 4.0 interface card and is the recommended HBA for the Dell R7625 server with AMD EPYC processors. The card’s total performance capacity is restricted by the bandwidth limitations of older generation servers that utilize PCIe 3.0. 

As in the prior test, the TPROC-H benchmark was run on an Oracle 19c database multiple times using the same card but in servers that implement two different PCIe generation architectures. 

Figure 2, on the previous page, illustrates the how the same dual-port 64GFC HBA delivers dramatically higher throughput and shorter database query times when deployed in a current generation server that implements PCIe 5.0 bus architecture. 

Taking the same dual-port 64GFC HBA and deploying it in a PCIe 5.0 R7625 server improved transaction time by 33% simply by removing the limitations imposed by the maximum bandwidth of the R740 PCIe 3.0 bus. 

Test Setup & Methodology

The HBA under test used current production drivers that are publicly available. Default settings were used. Details of the test environment and systems under test are found in Tables 1-6. Figure 3 shows a composite test environment. 

Server systems were all VMware ESXi 8 hosts running ESXi-8.0U1. Storage volumes mapped to each VM were configured as thick provisioned, eagerly zeroed. PVSCSI controller was used. 

Each VM was assigned 128GB of memory and 24 vCPUs. Each VM was running RHEL 8.9.

Details of the HammerDB tests are found in the “Overview” section above. 

For the 16/32/64GFC comparisons the server’s HBA-to-switch connection was configured to each of the link speeds as required by each test scenario. 

For the PCIe generation comparison test, the R7625 and R740 were not matched with respect to CPU and memory but as the test focused on I/O, the differences were acceptable. 

Table 1. 64G HBA Under Test

 Table 2. R7625 Server Configuration

Table 3. R740 Server Configuration

Table 4. Database Test Tool

 Table 5. Storage Configuration

Table 6. Oracle Database Configuration

About AMD

 For over 50 years, AMD has been at the forefront of driving innovation in high-performance computing, graphics, and visualization technologies. Their products are relied upon by billions of people, leading Fortune 500 businesses, and cutting-edge scientific research institutions worldwide. AMD's mission is to build exceptional products that accelerate next-generation computing experiences and power solutions for the world's most important challenges. 
Visit http://www.amd.com for more information about AMD.

Broadcom Emulex LPe36002

The Broadcom Emulex LPe36000-series Gen 7 Fibre Channel HBAs are designed for demanding mission-critical workloads and emerging applications. The family of adapters features Silicon Root of Trust security, designed to thwart firmware attacks aimed at enterprises and governments.

Gen 7 64G provides seamless backward compatibility to 32G and 16G networks.

Dell sells the LPe36002 64G HBA for the same price as the 32G model. 

About Tolly

 The Tolly Group companies have been delivering world-class IT services for over 30 years. Tolly is a leading global provider of third-party validation services for vendors of IT products, components and services.

You can reach the company by E-mail at sales@tolly.com, or by telephone at
 +1 561.391.5610.

Visit Tolly on the Internet at:
http://www.tolly.com

Tolly Terms Of Usage

 The Tolly Gro This document is provided, free-of-charge, to help you understand whether a given product, technology, or service merits additional investigation for your particular needs. Any decision to purchase a product must be based on your own assessment of suitability based on your needs.  The document should never be used as a substitute for advice from a qualified IT or business professional.  This evaluation was focused on illustrating specific features and/or performance of the product(s) and was conducted under controlled, laboratory conditions. Certain tests January have been tailored to reflect performance under ideal conditions; performance January vary under real-world conditions. Users should run tests based on their own real-world scenarios to validate performance for their own networks.

Reasonable efforts were made to ensure the accuracy of the data contained herein but errors and/or oversights can occur. The test/audit documented herein January also rely on various test tools the accuracy of which is beyond our control. Furthermore, the document relies on certain representations by the sponsor that are beyond our control to verify. Among these is that the software/hardware tested is production or production track and is, or will be, available in equivalent or better form to commercial customers. Accordingly, this document is provided "as is," and Tolly Enterprises, LLC (Tolly) gives no warranty, representation or undertaking, whether express or implied, and accepts no legal responsibility, whether direct or indirect, for the accuracy, completeness, usefulness, or suitability of any information contained herein. By reviewing this document, you agree that your use of any information contained herein is at your own risk, and you accept all risks and responsibility for losses, damages, costs, and other consequences resulting directly or indirectly from any information or material available on it. Tolly is not responsible for, and you agree to hold Tolly and its related affiliates harmless from any loss, harm, injury, or damage resulting from or arising out of your use of or reliance on any of the information provided herein.  

Tolly makes no claim as to whether any product or company described herein is suitable for investment.   You should obtain your own independent professional advice, whether legal, accounting or otherwise, before proceeding with any investment or project related to any information, products or companies described herein. When foreign translations exist, the English document is considered authoritative. To assure accuracy, only use documents downloaded directly from Tolly.com. No part of any document January be reproduced, in whole or in part, without the specific written permission of Tolly.  All trademarks used in the document are owned by their respective owners.  You agree not to use any trademark in or as the whole or part of your own trademarks in connection with any activities, products or services which are not ours, or in a manner which January be confusing, misleading, or deceptive or in a manner that disparages us or our information, projects or developments.

Dell PowerEdge R7625 Rack Server & Emulex LPe36002 Host Bus Adapter

64G Fibre Channel up to 4:1 Server Virtualization Consolidation

Tolly Report #224105

Tolly test report demonstrating that Dell PowerEdge R7625 Rack Server outfitted with the Emulex LPe36002 64G Fibre Channel adapter can improve virtualization server performance up to 4x vs older generation technologies.

Executive Summary

New generation technology can be expected to improve performance. There are times, however, when multiple technology advances can combine to provide an outsized advantage. Such is the case when the Dell PowerEdge R7625 Rack Server with AMD EPYC processors is combined with the Broadcom Emulex LPe36002 64G Fibre Channel Host Bus Adapter.

Dell commissioned Tolly to benchmark the database performance of the Broadcom Emulex LPe36002 64G Fibre Channel dual-port host bus adapter (HBA) running in the Dell PowerEdge R7625 Rack Server and compare that to the same combined workload performance running in four separate, R740-class servers each outfitted with a 16G FC HBA as was standard with that server generation.

Tests showed that the new R7625 AMD EPYC platform's increased CPU power and improved memory performance/capacity provide an environment where the database application can push the Emulex 64G FC HBA to full line rate performance of 64GFC thus matching the combined application throughput of four R740-class Purley platform servers using 16G FC HBAs. See Figure 1.

Overview

The goal of this test was to illustrate, simply, that a single Dell PowerEdge R7625 Rack Server using a single port of a PCIe 4.0-based, dual-port Emulex 64GFC HBA can equal the I/O throughput of four individual, older generation, R740-class servers each using a single port of a 16GFC HBA. 

The Dell PowerEdge R740-class servers use older, less powerful CPUs and use 16GFC HBAs that offer, at best, 25% of the 64GFC HBA’s throughput. The HBAs are constrained by the bandwidth of the PCIe 3.0 bus architecture which would limit the benefits of using the higher FC speed HBAs in the older servers. 

The broader point is that this significant performance improvement means that, for server virtualization applications, a single Dell PowerEdge R7625 Rack Server can be used to replace and consolidate the workloads and operating expenses of up to four older servers.

Test Background & Results

Server virtualization is an important part of IT infrastructure for countless businesses and organizations worldwide. Efficient use of the underlying server hardware components is an important aspect of providing high quality end-user experience while controlling costs. Certain elements of server virtualization can place a tremendous load on I/O resources. In particular, “Boot Storms”  can be impacted severely by lack of sufficient I/O bandwidth. The scenario was run separately on a single Dell PowerEdge R7625 server, powered by AMD EPYC, outfitted with a 64GFC HBA and then, again, simultaneously on four R740-class servers each outfitted with a 16GFC HBA.

“Boot Storm”

This is an informal term applied to situations where multiple VMs are started simultaneously. During the boot process all of the VMs use a workspace profile that will, upon startup, load a standard set of applications and read initial data from the data store simultaneously thus creating the "storm" of I/O requests.

The test was run using two different scenarios. In the first scenario, four of the older servers each booted six VMs simultaneously against the same Dell data store. In the second scenario, the single Dell PowerEdge R7625 booted 24 VMs simultaneously against the same data store.

Figure 1, on the cover page, summarizes the results of the “Boot Storm” tests in terms of storage I/O and startup (boot) time. The I/O throughput difference between the single 64GFC server and the four 16GFC servers is apparent. Where each of the older servers delivers throughput of ~1,600MB/s, the 64GFC server throughput was measured at ~6,400MB/s. 

This increase in throughput on the 64GFC Dell PowerEdge R7625 server results in dramatically faster boot times for each of the 24VMs tested. As shown in the the figure, the average, per-VM boot time for VMs running on the R740-class systems was 31.54. The the average, per-VM boot time for VMs running on the R7625 system was 7.58s. This represents an improvement of 76%.

Test Setup & Methodology

The HBA under test used current production drivers that are publicly available. Default settings were used. Details of the test environment and systems under test are found in Tables 1-9. Figure 2 shows a composite test environment. 

As tests involved basic functions of VMware, no detailed test methodology is required.

     Table 1. 64G HBA Under Test

   Table 2. R7625 Server Configuration

Table 3. VMware Configuration

Table 4. Storage Configuration

 Table 5. 16G HBA Under Test

Table 6. R740 Class Server Configuration Host 1

Table 7. R740 Class Server Configuration Host 2

Table 8. R740 Class Server Configuration Host 3

Table 9. R740 Class Server Configuration Host 4

About AMD

 For over 50 years, AMD has been at the forefront of driving innovation in high-performance computing, graphics, and visualization technologies. Their products are relied upon by billions of people, leading Fortune 500 businesses, and cutting-edge scientific research institutions worldwide. AMD's mission is to build exceptional products that accelerate next-generation computing experiences and power solutions for the world's most important challenges. 
Visit http://www.amd.com for more information about AMD.

Broadcom Emulex LPe36002

The Broadcom Emulex LPe36000-series Gen 7 Fibre Channel HBAs are designed for demanding mission-critical workloads and emerging applications. The family of adapters features Silicon Root of Trust security, designed to thwart firmware attacks aimed at enterprises and governments.

Gen 7 64G provides seamless backward compatibility to 32G and 16G networks.

Dell sells the LPe36002 64G HBA for the same price as the 32G model. 

About Tolly

 The Tolly Group companies have been delivering world-class IT services for over 30 years. Tolly is a leading global provider of third-party validation services for vendors of IT products, components and services.

You can reach the company by E-mail at sales@tolly.com, or by telephone at
 +1 561.391.5610.

Visit Tolly on the Internet at:
http://www.tolly.com 

Tolly Terms Of Usage

 The Tolly Gro This document is provided, free-of-charge, to help you understand whether a given product, technology, or service merits additional investigation for your particular needs. Any decision to purchase a product must be based on your own assessment of suitability based on your needs.  The document should never be used as a substitute for advice from a qualified IT or business professional.  This evaluation was focused on illustrating specific features and/or performance of the product(s) and was conducted under controlled, laboratory conditions. Certain tests January have been tailored to reflect performance under ideal conditions; performance January vary under real-world conditions. Users should run tests based on their own real-world scenarios to validate performance for their own networks.

Reasonable efforts were made to ensure the accuracy of the data contained herein but errors and/or oversights can occur. The test/audit documented herein January also rely on various test tools the accuracy of which is beyond our control. Furthermore, the document relies on certain representations by the sponsor that are beyond our control to verify. Among these is that the software/hardware tested is production or production track and is, or will be, available in equivalent or better form to commercial customers. Accordingly, this document is provided "as is," and Tolly Enterprises, LLC (Tolly) gives no warranty, representation or undertaking, whether express or implied, and accepts no legal responsibility, whether direct or indirect, for the accuracy, completeness, usefulness, or suitability of any information contained herein. By reviewing this document, you agree that your use of any information contained herein is at your own risk, and you accept all risks and responsibility for losses, damages, costs, and other consequences resulting directly or indirectly from any information or material available on it. Tolly is not responsible for, and you agree to hold Tolly and its related affiliates harmless from any loss, harm, injury, or damage resulting from or arising out of your use of or reliance on any of the information provided herein.  

Tolly makes no claim as to whether any product or company described herein is suitable for investment.   You should obtain your own independent professional advice, whether legal, accounting or otherwise, before proceeding with any investment or project related to any information, products or companies described herein. When foreign translations exist, the English document is considered authoritative. To assure accuracy, only use documents downloaded directly from Tolly.com. No part of any document January be reproduced, in whole or in part, without the specific written permission of Tolly.  All trademarks used in the document are owned by their respective owners.  You agree not to use any trademark in or as the whole or part of your own trademarks in connection with any activities, products or services which are not ours, or in a manner which January be confusing, misleading, or deceptive or in a manner that disparages us or our information, projects or developments.

Dell PowerEdge R7625 Rack Server & Emulex LPe36002 Host Bus Adapter

64G Fibre Channel Enables up to 4:1 Application Server Consolidation

Tolly Report #224104

Tolly test report demonstrating that Dell PowerEdge R7625 Rack Server outfitted with the Emulex LPe36002 64G Fibre Channel adapter can improve application performance up to 4x vs older generation technologies.

Executive Summary

New generation technology can be expected to improve performance. There are times, however, when multiple technology advances can combine to provide an outsized advantage. Such is the case when the Dell PowerEdge R7625 Rack Server with AMD EPYC processors is combined with the Broadcom Emulex LPe36002 64G Fibre Channel Host Bus Adapter.

Dell commissioned Tolly to benchmark the database performance of the Broadcom Emulex LPe36002 64G Fibre Channel dual-port host bus adapter (HBA) running in the Dell PowerEdge R7625 Rack Server and compare that to the same combined workload performance running in four separate, R740-class servers each outfitted with a 16G FC HBA as was standard with that server generation.

Tests showed that the new R7625 AMD EPYC platform's increased CPU power and improved memory performance/capacity provide an environment where the database application can push the Emulex 64G FC HBA to full line rate performance of 64GFC thus matching the combined application throughput of four R740-class Purley platform servers using 16G FC HBAs. See Figure 1.

Overview

The goal of this test was to illustrate, simply, that a single Dell PowerEdge R7625 Rack Server using a single port of a PCIe 4.0-based, dual-port Emulex 64G FC can equal the I/O throughput of four individual, older generation, R740-class servers each using a single port of a 16G FC HBA.

The R740-class servers use older, less powerful CPUs and use 16G FC HBAs that offer, at best, 25% of the 64G FC HBA’s throughput. The HBAs are constrained by the bandwidth of the PCIe 3.0 bus architecture which would limit the benefits of using the higher FC speed HBAs in the older servers.

The broader point is that this significant performance improvement means that, for I/O-bound applications, a single Dell PowerEdge R7625 Rack Server can be used to replace and consolidate the workloads and operating expenses of up to four older servers.

Test Background & Results

The same test was run on all of the servers and consisted of running the TPROC-H analytics workload of HammerDB. The tests were run using the Oracle 19c database environment, but the results are generally applicable to any database or other input/output intensive workload.

The TPROC-H workload measures how long it takes to run a series of 22 different types of decision support queries.  This type of workload is “read only” with no database updates taking place.

The test was run using two different scenarios. In the first scenario, four of the older servers ran the HammerDB benchmark simultaneously against the same Dell data store. In the second scenario, the single Dell PowerEdge R7625 ran the benchmark against the same data store.

Figure 1, above the horizontal dividing line, summarizes results of the first scenario. Because those servers were using 16G FC HBAs, 16G was the theoretical maximum for network I/O and, thus a potential bottleneck for each server. As each server finished the test, the reduced load on the target data store allowed subsequent server’s tests to run more quickly. The fastest completion time was 335 seconds and the slowest was 448 seconds with the average being 405.5 seconds.

Figure 1, below the horizontal dividing line, summarizes results of the second scenario. Here, a single Dell PowerEdge R7625 Rack Server outfitted with an Emulex 64G FC HBA was able to complete the same test in 99 seconds. This illustrates that the R7625 could take on the full load of four servers running this type of workload.

Figure 2 shows the results of the same two scenarios overlaid and measured in terms of disk I/O over the course of the tests. The red dots represent the combined disk I/O of all four older generation servers. The blue dots represent the single Dell PowerEdge R7625 Rack Server, powered by AMD EPYC processors. The disk throughput of the single R7625 at 64G matches or exceeds the combined throughput of the four 16G servers.

Figure 3, below, illustrates the networking flow of the four older generation servers, in blue, and the Dell PowerEdge R7625, in red, across the Broadcom Brocade 64G Fibre Channel switch.

Test Setup & Methodology

The HBA under test used current production drivers that are publicly available. Default settings were used. Details of the test environment and systems under test are found in Tables 1-10. Figure 3 shows a composite test environment.

Server systems were all VMware ESXi 8 hosts running ESXi-8.0U1-21495797 (8U2 GA). Storage volumes mapped to each VM were configured as thick provisioned, eagerly zeroed. PVSCSI controller was used.

Each VM was assigned 100GB of memory and 40 vCPUs. Each VM was running RHEL 8.8

Details of the HammerDB tests are found in the “Test Background & Results” section above.

                  Table 1. 64G HBA Under Test

 Table 2. R7625 Server Configuration

Table 3. Database Test Tool

Table 4. Oracle Database Configuration

 Table 5. Storage Configuration

 Table 6. 16G HBA Under Test

Table 7. R740 Class Server Configuration Host 1

Table 8. R740 Class Server Configuration Host 2

Table 9. R740 Class Server Configuration Host 3

Table 10. R740 Class Server Configuration Host 4

About AMD

 For over 50 years, AMD has been at the forefront of driving innovation in high-performance computing, graphics, and visualization technologies. Their products are relied upon by billions of people, leading Fortune 500 businesses, and cutting-edge scientific research institutions worldwide. AMD's mission is to build exceptional products that accelerate next-generation computing experiences and power solutions for the world's most important challenges. 
Visit http://www.amd.com for more information about AMD.

Broadcom Emulex LPe36002

The Broadcom Emulex LPe36000-series Gen 7 Fibre Channel HBAs are designed for demanding mission-critical workloads and emerging applications. The family of adapters features Silicon Root of Trust security, designed to thwart firmware attacks aimed at enterprises and governments.

Gen 7 64G provides seamless backward compatibility to 32G and 16G networks.

Dell sells the LPe36002 64G HBA for the same price as the 32G model. 

About Tolly

 The Tolly Group companies have been delivering world-class IT services for over 30 years. Tolly is a leading global provider of third-party validation services for vendors of IT products, components and services.

You can reach the company by E-mail at sales@tolly.com, or by telephone at
 +1 561.391.5610.

Visit Tolly on the Internet at:
http://www.tolly.com

Tolly Terms Of Usage

 The Tolly Gro This document is provided, free-of-charge, to help you understand whether a given product, technology, or service merits additional investigation for your particular needs. Any decision to purchase a product must be based on your own assessment of suitability based on your needs.  The document should never be used as a substitute for advice from a qualified IT or business professional.  This evaluation was focused on illustrating specific features and/or performance of the product(s) and was conducted under controlled, laboratory conditions. Certain tests January have been tailored to reflect performance under ideal conditions; performance January vary under real-world conditions. Users should run tests based on their own real-world scenarios to validate performance for their own networks.

Reasonable efforts were made to ensure the accuracy of the data contained herein but errors and/or oversights can occur. The test/audit documented herein January also rely on various test tools the accuracy of which is beyond our control. Furthermore, the document relies on certain representations by the sponsor that are beyond our control to verify. Among these is that the software/hardware tested is production or production track and is, or will be, available in equivalent or better form to commercial customers. Accordingly, this document is provided "as is," and Tolly Enterprises, LLC (Tolly) gives no warranty, representation or undertaking, whether express or implied, and accepts no legal responsibility, whether direct or indirect, for the accuracy, completeness, usefulness, or suitability of any information contained herein. By reviewing this document, you agree that your use of any information contained herein is at your own risk, and you accept all risks and responsibility for losses, damages, costs, and other consequences resulting directly or indirectly from any information or material available on it. Tolly is not responsible for, and you agree to hold Tolly and its related affiliates harmless from any loss, harm, injury, or damage resulting from or arising out of your use of or reliance on any of the information provided herein.  

Tolly makes no claim as to whether any product or company described herein is suitable for investment.   You should obtain your own independent professional advice, whether legal, accounting or otherwise, before proceeding with any investment or project related to any information, products or companies described herein. When foreign translations exist, the English document is considered authoritative. To assure accuracy, only use documents downloaded directly from Tolly.com. No part of any document January be reproduced, in whole or in part, without the specific written permission of Tolly.  All trademarks used in the document are owned by their respective owners.  You agree not to use any trademark in or as the whole or part of your own trademarks in connection with any activities, products or services which are not ours, or in a manner which January be confusing, misleading, or deceptive or in a manner that disparages us or our information, projects or developments.

Dell PowerEdge R7625 Rack Server & Emulex LPe36002 Host Bus Adapter

64G Fibre Channel Microsoft SQL Server Performance – NVMe/FC vs. SCSI/FC

Tolly Report #224107

Tolly test report demonstrating that Dell PowerEdge R7625 Rack Server outfitted with the Emulex LPe36002 Host Bus Adapter using NVMe/FC can improve application performance vs older generation SCSI/FC.

Executive Summary

New generation servers can bring higher performance across a range of areas. This is certainly the case with Dell’s 16th-generation server line. Similarly, newer protocols like NVM Express (NVMe) over Fibre Channel (FC) can provide greater throughput and efficiency than older SCSI over FC. Dell is unique in offering an end-to-end NVMe/FC connectivity solution in the mid-range storage marketplace with the PowerStore line.

Dell commissioned Tolly to benchmark the performance of the Broadcom Emulex LPe36002 64G Fibre Channel dual-port host bus adapter (HBA) running in the Dell PowerEdge R7625 Rack Server with AMD EPYC processors by testing using actual database applications rather than simulated I/O microbenchmarks. Testing focused on evaluating the database throughput, latency, and CPU efficiency of accessing Microsoft SQL Server 2019 for Linux systems over older SCSI/FC and newer NVMe/FC. Databases were stored on a Dell PowerStore 9200T storage appliance.

Tests showed significant improvements in transaction throughput, latency reduction, and CPU efficiency. See Figure 1 for a summary of relative improvements.

Overview

The goal of this test was to illustrate the performance benefits of using the newer, more-efficient NVMe/FC protocol in lieu of the older, less-efficient SCSI/FC protocol in conjunction with Emulex 64G FC HBAs running under Linux in a Dell PowerEdge R7625 Rack Server.  (Dell sells the Emulex 64G FC HBA for the same price as the Emulex 32G FC HBA.)

The test was run using Microsoft SQL Server 2019 for Linux accessing the database via SCSI and then via NVMe. 

While low-level component benchmarks are instructive,  ultimately system architects are rightly most interested in how network-level improvements can translate into application performance improvements. This benchmarking was done with HammerDB which generates actual user transactions against an actual database. The test was focused on TPROC-C which is the HammerDB, database-oriented implementation of the de facto standard TPC-C online transaction processing benchmark.

Tests showed significant improvements in key benchmarks. 

Test Results

Microsoft SQL Server 2019 for Linux

Transaction Processing. The NVMe/FC results were significantly better than the SCSI/FC results. When run over NVMe/FC, 38% more transactions per minute were processed. 

CPU Efficiency. The NVMe/FC results were significantly better than the SCSI/FC results. When run over NVMe/FC, the CPU efficiency was improved by 50%.

P95 Stored Procedure Latency.  Similarly, the NVMe/FC results were significantly better than the SCSI/FC results. When run over NVMe/FC, the latency was reduced by 35%.

Test Setup & Methodology

The HBA under test used current production drivers that are publicly available. Default settings were used. Details of the test environment and systems under test are found in Tables 1-5. Figure 2 shows a composite test environment. 

Database Test

The goal of this test was to benchmark the database transaction performance of each HBA running the HammerDB “TPROC-C” workload which, as noted earlier, is the HammerDB, database version of the Transaction Processing Council’s TPC-C OLTP benchmarked

A Dell PowerEdge R7625 server, powered by AMD EPYC processors, was configured with the HBA under test. The Broadcom Emulex LPe36002 64G HBA connected to a Dell PowerStore 9200T via a Dell Connectrix 64G Fibre Channel switch. The test utilized a single 64G FC port of the Emulex HBA.   

The server ran RHEL 8.9. SCSI Device Mapper and NVMe native multipath were enabled for the respective devices. NUMA was set to off and “transparent huge pages” was disabled.

For storage, path selection policy for NVMe native multipath was set to “round-robin". For SCSI Device mapper multipath was set to "queue-length 0”. 

This test was run using Microsoft SQL Server 2019 for Linux,

The open source HammerDB test tool was used to populate the database schema and run the workload.

Table 1. HBA Under Test

Table 2. Server Configuration

Table 3. Microsoft Database Configuration

Table 4. Database Test Tool

Table 5. Storage Configuration

About AMD

 For over 50 years, AMD has been at the forefront of driving innovation in high-performance computing, graphics, and visualization technologies. Their products are relied upon by billions of people, leading Fortune 500 businesses, and cutting-edge scientific research institutions worldwide. AMD's mission is to build exceptional products that accelerate next-generation computing experiences and power solutions for the world's most important challenges. 
Visit http://www.amd.com for more information about AMD.

Broadcom Emulex LPe36002

The Broadcom Emulex LPe36000-series Gen 7 Fibre Channel HBAs are designed for demanding mission-critical workloads and emerging applications. The family of adapters features Silicon Root of Trust security, designed to thwart firmware attacks aimed at enterprises and governments.

Gen 7 64G provides seamless backward compatibility to 32G and 16G networks.

Dell sells the LPe36002 64G HBA for the same price as the 32G model. 

About Tolly

 The Tolly Group companies have been delivering world-class IT services for over 30 years. Tolly is a leading global provider of third-party validation services for vendors of IT products, components and services.

You can reach the company by E-mail at sales@tolly.com, or by telephone at
 +1 561.391.5610.

Visit Tolly on the Internet at:
http://www.tolly.com

Tolly Terms Of Usage

 The Tolly Gro This document is provided, free-of-charge, to help you understand whether a given product, technology, or service merits additional investigation for your particular needs. Any decision to purchase a product must be based on your own assessment of suitability based on your needs.  The document should never be used as a substitute for advice from a qualified IT or business professional.  This evaluation was focused on illustrating specific features and/or performance of the product(s) and was conducted under controlled, laboratory conditions. Certain tests January have been tailored to reflect performance under ideal conditions; performance January vary under real-world conditions. Users should run tests based on their own real-world scenarios to validate performance for their own networks.

Reasonable efforts were made to ensure the accuracy of the data contained herein but errors and/or oversights can occur. The test/audit documented herein January also rely on various test tools the accuracy of which is beyond our control. Furthermore, the document relies on certain representations by the sponsor that are beyond our control to verify. Among these is that the software/hardware tested is production or production track and is, or will be, available in equivalent or better form to commercial customers. Accordingly, this document is provided "as is," and Tolly Enterprises, LLC (Tolly) gives no warranty, representation or undertaking, whether express or implied, and accepts no legal responsibility, whether direct or indirect, for the accuracy, completeness, usefulness, or suitability of any information contained herein. By reviewing this document, you agree that your use of any information contained herein is at your own risk, and you accept all risks and responsibility for losses, damages, costs, and other consequences resulting directly or indirectly from any information or material available on it. Tolly is not responsible for, and you agree to hold Tolly and its related affiliates harmless from any loss, harm, injury, or damage resulting from or arising out of your use of or reliance on any of the information provided herein.  

Tolly makes no claim as to whether any product or company described herein is suitable for investment.   You should obtain your own independent professional advice, whether legal, accounting or otherwise, before proceeding with any investment or project related to any information, products or companies described herein. When foreign translations exist, the English document is considered authoritative. To assure accuracy, only use documents downloaded directly from Tolly.com. No part of any document January be reproduced, in whole or in part, without the specific written permission of Tolly.  All trademarks used in the document are owned by their respective owners.  You agree not to use any trademark in or as the whole or part of your own trademarks in connection with any activities, products or services which are not ours, or in a manner which January be confusing, misleading, or deceptive or in a manner that disparages us or our information, projects or developments.

Dell PowerEdge R7625 servers with AMD EPYC processors & Emulex 64G Fibre Channel LPe36002 Host Bus Adapters demonstrate Application Advantages

Executive Summary

New generation technology can be expected to improve performance. There are times, however, when multiple technology advances can combine to provide an outsized advantage. Such is the case when the Dell PowerEdge R7625 Rack Server is combined with the Broadcom Emulex LPe36002 64G Fibre Channel Host Bus Adapter. 

Dell commissioned Tolly to benchmark the database performance of the Broadcom Emulex LPe36002 64G Fibre Channel dual-port host bus adapter (HBA) running in the Dell PowerEdge R7625 Rack Server and compare that to the same combined workload performance running in four separate, R740- class servers each outfitted with a 16G FC HBA as was standard with that server generation.

Following is a summary of the 4 tests conducted:

1. The first test measured HammerDB “TPROC-C” Online Transaction Processing (OLTP) workload performance with Microsoft SQL Server Database to compare the NVMe/FC vs SCSI/FC performance on a Dell PowerEdge R7625 server with Broadcom Emulex LPe36002 64G Fibre Channel HBA.

Key Findings:

• Improved database transactions by up to 38%
• Reduced database stored procedure latency by up to 35%
• Improved server CPU efficiency by up to 50%

READ THE FULL STUDY HERE:

• 64G Fibre Channel Microsoft SQL Performance – NVMe/FC vs. SCSI/FC

2. The second test measured the HammerDB “TPROC-H” Decision Support System (DSS) analytics workload queries on a single Dell R7625 AMD EPYC-based platform and found that it pushed Emulex 64G Fibre Channel HBA to full line rate performance of 64G Fibre Channel, thus matching the combined application throughput of four previous generation R740-class Purley platform servers using 16G Fibre Channel HBAs.

Key Findings:

• Impressive database analytics throughput consolidation- from four R740 servers with 16G Fibre Channel HBAs to a single R7625 with 64G Fibre Channel HBA
• Consolidating analytics workload can significantly reduce I/O bound query times

READ THE FULL STUDY HERE:

• 64G Fibre Channel Enables up to 4:1 Application Server Consolidation

3. The third test revealed a 4:1 server consolidation benefit for Virtualization workloads where a single Dell R7625 AMD EPYC-based platform with 64G Fibre channel HBA matched the combined application throughput of four Dell R740-class Purley platform servers using 16G Fibre Channel HBAs.

Key Findings:

• Consolidation of virtual machine (VM) “Boot Storm” - Virtualization workloads throughput from four Dell R740 servers with 16GFC HBA to a singleDell R7625 with Emulex 64G Fibre Channel

• A VDI boot storm is the consumption of compute and disk I/O resources during the initial startup of end-user desktop virtual images that results in poor performance for all users. VDI environments need read I/O at boot (Bootstorm).

READ THE FULL STUDY HERE:

• 64G Fibre Channel up to 4:1 Server Virtualization Consolidation

4. The final test determined that the Dell R7625 with PCIe Gen5 and Emulex 64G Fibre Channel HBA combined to overcome bottlenecks for Oracle database HammerDB “TPROC-H” DSS analytics workload queries, achieving maximum throughput 

Key Findings:

• R7625 with 64GFC HBA can achieve 4x the database analytics throughput of the16GFC HBA and 2x the throughput of the 32GFC HBA 
• 42% improvement in complex database ad hoc query processing time when running the dual-port 64GFC HBA on the PCIe 5.0-based R7625 server compared to the older generation R740 server 

READ THE FULL STUDY HERE:

• Dell R7625 and 64GFC Combine to Accelerate Oracle Analytics Workloads

Key performance indicators (KPIs) show that a hardware refresh with the latest-generation Dell PowerEdge servers and AMD EPYC^™ processors can help enterprises improve the performance, efficiency, and security of their server infrastructures.

 Executive Summary

Forrester Consulting reports that data centers that refresh their servers at least every three years can gain technological and business benefits compared to data centers that do not.[1] These benefits manifest themselves through higher performance, increased efficiency, and better security. Prowess Consulting investigated these benefits further by examining results from industry-standard benchmarks and environmental ratings. Based on our research, we concur with the Forrester Consulting opinion that the benefits of a server refresh can easily outweigh the costs.

If you are still wondering whether it’s time to refresh your servers, you can use this study to help you decide. We examined the effects of upgrading legacy servers running on x86-based processors that are more than three years old to Dell PowerEdge servers powered by 4th Generation AMD EPYC processors. Examples of the kinds of benefits we uncovered in the course of our investigation include:

• Up to 232% higher performance per watt[2]
• Up to 48% lower processor cost[3]
• Up to 40% lower software licensing costs through 5:1 server consolidation[4]

Exploring the Value of a Server Refresh

A 2019 report by Forrester Consulting determined that in order to be more agile and productive, data centers should be refreshing their servers at least every three years.¹ The online survey showed numerous technical benefits to be gained from a server refresh, and it concluded that organizations that keep their servers modernized and updated tend to earn greater benefits from their infrastructure investments.¹ Security is also a critical concern for businesses with aging server platforms. Older-generation processors might not have the latest security features necessary to protect against modern security threats.

These findings suggest that if you are running legacy servers powered by processors more than three years old, you simply cannot afford not to consider a server refresh. With the innovative hardware technologies being released in 2023, Prowess Consulting believes that now is an excellent time to look at the latest server and processor offerings. In this article, we examine the performance, efficiency, and security benefits of upgrading your legacy server platforms to the latest PowerEdge servers built on 4th Gen AMD EPYC processors.

With the goal of identifying the potential benefits you could enjoy by refreshing to latest-generation server hardware, we looked at the popular combination of Dell servers and AMD processors. Our analysis indicates that upgrading to PowerEdge servers with 4th Gen AMD EPYC processors can help improve performance, efficiency, and security. To quantify these improvements, we used a variety of industry-standard benchmarks, published results, and environmental ratings. We also evaluated qualitative benefits of refreshing servers, such as the security benefits provided by current-generation servers.

Much of this study refers to a hypothetical update scenario that involves moving from a two-node cluster of 2S 2U Fujitsu  PRIMERGY^® RX2540 M5 servers with two Intel^® Xeon^® Platinum 8280 processors each to a two-node cluster of 2S 2U Dell PowerEdge R7615 servers with a single AMD EPYC 9654P processor each. This tangible comparison helps illustrate how a server refresh can help with performance, efficiency, and security.

Total Cost of Ownership (TCO)

The total cost of owning and running a server—and its reciprocal, the value of upgrading legacy servers to the latest generation—is complex. Specific benefits from a server refresh will vary from organization to organization and from use case to use case. This study does not attempt to generate a single number that quantifies the TCO benefits of a server refresh, but we found that an upgrade from three-to-five-year-old x86 processors to 4th Gen AMD EPYC processors can provide several indicative benefits:

• Up to 40% lower software licensing costs through 5:1 server consolidation⁴
• Up to 38% lower software licensing costs per unit of performance[5]
• Up to 31% reduction in average energy cost[6]

These figures offer a sense of the cost benefits that can come with a server refresh. And while this analysis lays out specific benefits from refreshing legacy servers in the context of performance, efficiency, and security, all of these kinds of benefits have a direct bearing on the cost of ownership for servers—and the gains from refreshing them.

Boost Performance

A server refresh can help you lower TCO while delivering the insights you need when you need them. Newer processors can deliver higher performance per core, meaning you can run the most demanding AI and high-performance computing (HPC) workloads while still lowering your power consumption and physical footprint.

Get Higher Performance Per Core and Per Watt

Based on SPEC^® benchmarking results comparing high-performance processors from several generations, we found that refreshing the two-socket Fujitsu PRIMERGY RX2540 M5 server with two Intel Xeon Platinum 8280 processors (28 cores) to a PowerEdge R7615 server with a single AMD EPYC 9654P processor (96 cores) could deliver up to twice the performance (102% higher) per core.[7]

Raw performance is an important pillar in understanding the full story of a server’s capabilities and cost of ownership. For example, virtualization continues to be a vital workload for many businesses, and while mere computational horsepower alone cannot capture how good a server might be for hosting virtual machines (VMs), it is still an important factor. With that fact in mind, we used VMmark^® 3.x benchmarking results to analyze this same refresh scenario looking specifically at performance/watt for virtualization workloads. A refresh from servers powered by three-to-five-year-old x86 processors to 4th Gen AMD EPYC processors can provide up to 232% higher performance per watt for virtualization workloads.²

A single AMD EPYC 9654P processor has more cores than two Intel Xeon Platinum 8280 processors combined. However, even accounting for this difference in core count, the refreshed servers powered by a 4th Gen AMD EPYC processor can provide up to 93% higher performance/watt/core than the legacy servers powered by three-to-five-year-old x86 processors.² Higher performance per watt and per core mean that you can either shrink your energy costs or server footprint for the same performance, or increase performance while holding power consumption and server footprint the same.

Increase Efficiency

IT budgets are being cut everywhere, and IT organizations are being told to do more with less. In short, improving the efficiency of hardware is critical to companies of all sizes.

Reducing capital expenditures (CapEx) is often the first consideration for organizations seeking to increase efficiency with a server refresh. Reduced costs upfront get reflected in lower amortized costs over the life of a server. The good news from our investigation is that upgrading to servers powered by current-generation processors can actually cost less than the legacy systems originally did.

Consider again the example of the legacy Fujitsu PRIMERGY RX2540 M5 servers running 2nd Gen Intel Xeon Platinum 8280 processors being refreshed to PowerEdge R7615 servers powered by 4th Gen AMD EPYC 9654P processors. Pricing servers is complex and multidimensional, but the majority of the price comes from the processors and the memory. If we hold memory roughly even between these two systems, processor price can give a rough idea of the relative prices of the two servers.

The two 2nd Gen Intel Xeon Scalable processors in each legacy server have a total MSRP of $22,920, compared to an MSRP of $11,805 for the single 4th Gen AMD EPYC processor in each new server.³ The representative 48% lower price can translate directly into lower system cost for the newer server—or, more likely, it can help absorb some of the cost of putting more memory into the new server to increase system efficiency, such as by hosting more VMs.

Improve License Efficiency

Using fewer servers to do the same amount of computing offers a number of savings opportunities, notably by reducing costs for software licensed by the server core. Licensing costs can end up forming a sizeable plurality if not an outright majority of the TCO of a server. Reducing the number of cores that you need to license can be a powerful way to reduce licensing costs.

To cite just one example, a study conducted by Dell Technologies showed that the latest-generation PowerEdge R7625 server with 4th Gen AMD EPYC processors offers 5:1 server consolidation compared to legacy servers using 1st Gen Intel Xeon Scalable processors. Specifically, 380 VMs running on five 2S legacy servers using 10 Intel Xeon Platinum 8180 processors (28 cores, 205 W) could be successfully migrated to one 2S 2U PowerEdge R7625 server powered by two AMD EPYC 9654 processors (96 cores, 360 W).⁴

Figure 1. Dell PowerEdge servers and 4th Gen AMD EPYC processors can help consolidate your data center footprint⁴

The refreshed server uses 31% fewer cores, which can help reduce virtualization licensing costs. For example, you could reduce the number of VMware^® licenses from 10 licenses for the five legacy 2S servers to six licenses for the new 2S server, a 40% cost savings on VMware licensing.⁴

In another example, the newer-generation processors were more performant than the three-to-five-year-old processors they replaced and so could provide the same level of performance using fewer cores. In this case, the lower core count due to the refresh lowered VMware licensing costs per unit of performance by up to 38%.⁵

Streamline Infrastructure Costs

Beyond savings on software costs, consolidating your servers with a refresh can save money on your physical infrastructure too. For example, fewer servers consume fewer networking resources, which can help reduce the cost of your networking infrastructure. A smaller number of servers also takes up less rack space, which can help reduce the footprint in your own data center—or it can translate directly into lower monthly costs if you use a co-location facility to host your data center (such as with a 5:1 server consolidation).⁴

Manage Power Consumption

Consolidating workloads from legacy servers to the newest-generation hardware can also lower power consumption. In our example, the 10 legacy processors in the consolidation scenario illustrated in Figure 1 are rated to have a combined maximum power draw of 2,050 W, compared to the total 720 W maximally drawn by the newest-generation processors, which represents a 64% reduction in power consumption by the processors.

Even if your server refresh plans call for keeping the same number of servers from generation to generation, you have options. If you anticipate needing additional performance, you could replace a legacy two-socket server with a newer two-socket model and gain the benefits of the higher core count in newest-generation processors. Alternatively, you could replace a two-socket legacy server with a single-socket server that provides similar performance but consumes less power. For example, VMmark benchmarking for the server-upgrade path discussed earlier recorded average usage for the Fujitsu PRIMERGY RX2540 M5 server running 2nd Gen Intel Xeon Platinum 8280 processors at 1,425.14 W and an average power draw for the PowerEdge R7615 server powered by a 4th Gen AMD EPYC 9654P processor of 982.42 W, demonstrating a drop of 31% in average power consumption.[8]

A server refresh allows you to take advantage of the latest advancements in management features, which you can use to improve performance, efficiency, and sustainability across your data center. For example, Dell OpenManage Enterprise Power Manager can help optimize the energy usage and power consumption of PowerEdge servers and servers from other top server vendors. You can use its real-time monitoring to identify power-hungry applications and devices or “zombie servers” that are running but not in use. Hardware and software telemetry helps you configure policies that will automatically take steps to reduce energy consumption or set power caps at the rack or group level. Predictive analytics can help identify power-usage trends so that you can proactively make changes to lower power consumption. For example, you can schedule low-demand workloads outside of regular business hours and take advantage of off-peak electricity rates.

Figure 2. Dell OpenManage Enterprise Power Manager (www.dell.com/en-us/dt/solutions/openmanage/power-management.htm) lets you set up alerts for excessive power usage and temperature

Cost-Effective Ways of Keeping Your Servers Cool

The latest-generation Dell PowerEdge servers include high-efficiency cooling technologies designed to reduce the amount of power needed to cool your servers. PowerEdge servers are designed with Dell Smart Cooling (www.dell.com/en-us/dt/servers/power-and-cooling.htm), which uses state-of-the-art thermal and mechanical simulation tools to ensure optimal cooling and sustained system performance.

• Multi Vector Cooling (MVC) (https://infohub.delltechnologies.com/p/multi-vector-cooling-2-0-for-next-generation-poweredge-servers/) intelligently adapts to changing environments and configurations by regulating the speed of PowerEdge servers’ case fans so that they use less energy, which also helps increase the lifespan of the servers.
• Dell Smart Flow (https://infohub.delltechnologies.com/p/improved-poweredge-server-thermal-capability-with-smart-flow/) is a PowerEdge server chassis design that pairs a larger air intake with less obstructed internal airflow, which helps to move more air more efficiently.
• Dell Direct Liquid Cooling (https://infohub.delltechnologies.com/p/dell-technologies-direct-liquid-cooling-support-for-new-poweredge-servers/) technology is available for the latest PowerEdge servers. Liquid cooling used alone or combined with air cooling provides highly efficient temperature management for processors with high thermal design power (TDP) limits.

Improve Sustainability

Dell PowerEdge servers can help “green up” your data center. As of July 2023, PowerEdge servers are the only Silver-rated data center servers listed in the Global Electronics Council’s Electronic Product Environmental Assessment Tool (EPEAT^™) (www.epa.gov/greenerproducts/electronic-product-environmental-assessment-tool-epeat).[9] EPEAT ranks qualifying products as Gold, Silver, or Bronze according to a set of required and optional criteria for environmental and social responsibility (https://globalelectronicscouncil.org/wp-content/uploads/NSF-426-2019.pdf); in achieving Silver ranking, PowerEdge servers meet all the required criteria and at least half of the optional criteria set out by EPEAT.[10]

Harden Security

With the increasing frequency and severity of cyberattacks, organizations must be proactive in ensuring that their security measures align with the latest cybersecurity standards. An upgraded server platform allows you to implement the latest multi-layered security, deploy advanced platform monitoring and management capabilities, and enable hardware security features.

Holistically Address Security with PowerEdge Servers

We found that PowerEdge servers are designed from the ground up with security in mind, and they thus provide holistic security. Holistic security for servers refers both to the defenses that OEMs such as Dell Technologies provide to protect servers from attack and to the design ideals that help support actions in response to attacks that succeed. PowerEdge servers are designed to conform to the US National Institute of Standards and Technology (NIST) Cybersecurity Framework. The NIST Cybersecurity Framework (www.nist.gov/cyberframework) consists of standards, guidelines, and best practices for organizations through five phases of cyberattacks: identification, protection, detection, response, and recovery.

A subset of this framework is the zero-trust paradigm for cybersecurity. Zero-trust is a cyber-protection paradigm that assumes all users and devices are untrusted until proven otherwise. For Dell hardware, this paradigm starts with its immutable hardware root of trust, hardware-based encryption that is used to verify subsequent operations within the server, such as booting. This verification establishes a chain of trust that extends throughout the server lifecycle, from deployment through maintenance to decommissioning. If a step in the boot process fails verification, the server shuts down so that automatic BIOS recovery can begin.

Similarly, PowerEdge servers use digital signatures on firmware updates to attest to the authenticity of the firmware running on a given server. Organizations can also use Dell management tools to maintain server firmware to a specified baseline. OpenManage Enterprise (www.dell.com/en-us/dt/solutions/openmanage/enterprise.htm) is a platform-management solution that can detect deviations from the baseline. Organizations can then use the Integrated Dell Remote Access Controller (iDRAC) (www.dell.com/en-us/dt/solutions/openmanage/idrac.htm) management controller to schedule repairs for the next time servers are rebooted for maintenance.

OpenManage Enterprise also helps deploy end-to-end security across all servers in an organization in other ways. Centralized management provided by the software uses real-time monitoring to detect potential threats, examine server activity, track user access, and analyze security logs. This makes it easier to identify and respond to potential threats before they can cause significant damage.

OpenManage Enterprise can help you quickly recover from a security breach with data backup and restoration capabilities. We highly recommend scheduling regular backups and restoration checks, which can help minimize the impact of an attack and ensure your data is protected.

Harness Hardware-Based Security with AMD EPYC Processors

4th Gen AMD EPYC processors offer a suite of hardened security technologies called AMD Infinity Guard (www.amd.com/en/processors/epyc-5-reasons-security), designed to complement your existing software- and hardware-based security. These built-into-the-silicon features can help you extend protections holistically across your x86 server platforms, regardless of what workloads they are running, who is accessing them, or where they are physically located.

AMD Infinity Guard consists of five CPU-enforced security technologies:

1. AMD Secure Processor works with the immutable Dell hardware root of trust to secure BIOS boot, ensuring that only validated and verified components are allowed to boot up and run.
2. Secure Memory Encryption (SME) helps protect against threats that target system memory, such as memory-scraping attacks. Even if an attacker gains access to system memory, they are unable to read or modify the encrypted data.
3. AMD Shadow Stack protects in-memory data against return-oriented programming (ROP) attacks. This feature supports Microsoft hardware-enforced stack protection.
4. Secure Encrypted Virtualization (SEV) blocks attacks against VMs by keeping guest operating systems and the hypervisor environment isolated from each other. The SEV Encrypted State (SEV-ES) extension adds another layer of protection for data in use.
5. SEV-Secure Nested Paging (SEV-SNP) helps protect the integrity of the hypervisor, ensuring that a corrupted VM cannot access the hypervisor's memory.

Insights and Support for Complex Infrastructures

Management decisions that optimize your IT environment can help you gain even more benefits from a server refresh. For example, Dell Live Optics (www.dell.com/en-us/dt/live-optics/index.htm) is a tool that lets you see into file systems, storage and database servers, on-premises and cloud environments, workloads, and data-protection operations. You can use these insights to get your server platforms running as performantly and efficiently as possible.

The last thing you want to happen after upgrading your servers is a disruption to resource availability and user productivity. However, achieving a seamless transition to the latest and emerging technologies might require a higher level of expertise than you have available in-house. In that case, you might choose to engage additional IT support, such as Dell ProSupport for Enterprise (www.dell.com/en-us/dt/services/support-services/prosupport-infrastructure-suite.htm).

Conclusion

Organizations that adopt a modernized server strategy, which includes a three-year hardware refresh cycle, can lower the TCO of their server estates. This lower cost of ownership can manifest itself both through aggregated costs and benefits for their overall server performance, efficiency, and security.

Research conducted by Prowess Consulting found that refreshing your servers to the latest-generation Dell PowerEdge servers and AMD EPYC processors can:

• Improve performance/watt by up to 232% after upgrading from 2nd Gen AMD EPYC processors²
• More than double performance/core after upgrading from 2nd Gen Intel Xeon Scalable processors⁷

Refreshing your servers can also improve efficiency in a number of ways, with:

• Up to 5:1 server consolidation after upgrading from 1st Gen Intel Xeon Scalable processors, helping with server-license efficiency⁴
• Up to 38% lower VMware vSphere^® licensing costs per unit of performance⁵
• Up to 31% lower average power consumption after upgrading from 2nd Gen Intel Xeon Scalable processors⁶

Moreover, newer environmentally and socially responsible server infrastructures can help reduce power and cooling costs for your data center.⁹

Finally, refreshing to newer servers can help holistically improve security for your server estate. Crucially, new servers with the latest-generation processors can help you adopt a zero-trust paradigm through features such as the Dell hardware root of trust and AMD Secure Processor, which require cryptographic authentication for each step of the server-boot process in order to head off attacks through compromised firmware. And features like AMD SME, SEV, and SEV-ES can help protect server operating systems and the VMs that depend upon them from low-level attacks.

Learn More

Learn more about Dell PowerEdge servers with 4th Gen AMD EPYC processors: www.dell.com/en-us/dt/servers/amd.htm

Discover other research reports by Prowess Consulting: https://prowessconsulting.com/labs/

Appendix

Table A1. Benchmarks and registry used for this study

The analysis in this document was done by Prowess Consulting and commissioned by Dell Technologies.

Prowess and the Prowess logo are trademarks of Prowess Consulting, LLC.

Copyright © 2023 Prowess Consulting, LLC. All rights reserved.

Other trademarks are the property of their respective owners.

Author: Prowess Consulting, LLC

Principled Technologies examined the performance improvements and cost savings associated with upgrading to the 16th Generation Dell PowerEdge R7625 for machine learning algorithms

Overview

Recent years have seen a dramatic increase in the amount of data organizations store and analyze. Between 2010 and 2020, the amount of data people and organizations created, copied, consumed, and stored increased from 2 zettabytes to 64 zettabytes.[i] Machine learning (ML) tools can help companies put this data to work by analyzing it and extracting key insights, enabling more informed, data-driven business decisions. To meet this need, ML tools have become more powerful—but these workloads also put more demand on data centers.

We used the HiBench benchmark to understand the benefits of upgrading from the 15G Dell^™ PowerEdge^™ R7525 server to the 16G Dell PowerEdge R7625 server powered by Broadcom^® network interface cards (NICs) and PERC 11 storage controllers. Both servers feature two AMD EPYC^™ 64-core processors for a direct core-to-core generational comparison. We measured the throughput and time to complete k-means clustering and Bayesian classification workloads using both servers. We found the latest-generation PowerEdge R7625 offered better performance for the same amount of cores running both workloads. This means that organizations that upgrade to the latest-generation PowerEdge R7625 servers could process ML workloads faster, allowing them to update their models with new data more frequently for more timely insights. Plus, organizations that choose PowerEdge R7625 servers could save money by reducing the number of servers required to do the same amount of work as PowerEdge R7525 servers, which could reduce energy/cooling costs as well as licensing costs—up to $10,178.99 per year per consolidated server on Red Hat OpenShift licensing.

The challenges of data proliferation and compute‑intensive workloads

The rise of the Internet of Things (IoT), cloud computing, and smartphones have made it possible for businesses to harvest data from a wide range of sources and utilize it to improve their operations. Retailers can use data to track customer behavior and make their marketing more effective; manufacturers can use data to make their processes more efficient; and financial institutions can use data to detect fraud or predict market changes. As businesses gain access to new sources of data and use new technologies to analyze that data, the demand for more powerful servers will continue to grow.

Machine learning and artificial intelligence (AI) workloads have enormous potential to improve business operations, but as they gain popularity, they consume increasing amounts of processing power.[ii] According to OpenAI, developers of ChatGPT, the computing power of their AI system doubles every 3.4 month.[iii] As the ML applications organizations use become more demanding, they will need more powerful servers in their data centers as well as efficient data analysis tools in the ML pipeline. Among those data analysis tools is Apache Spark^™.

Apache Spark is an open-source computing framework that converts very large data sets into smaller blocks of data for the purpose of applying machine learning algorithms and analyzing the data quickly using a distributed network of devices. For algorithms that operate on chunks of data, Spark is effective because it farms the data out to servers in the cluster, the servers process the chunks of data, and then Spark combines them for the final result. One of the main advantages of using Spark is that it can split data sets into chunks that fit in memory (when the entire data set might not) and operate with data that is entirely in memory—it doesn’t need to write to disk, which saves time. Spark is scalable: users can expand the size of their data set by adding more nodes. According to Databricks^®, Spark can process “multiple petabytes of data on clusters of over 8,000 nodes,” and Spark supports a variety of data sources, including Hadoop HDFS. [iv]

We focused on two Apache Spark capabilities—k-means clustering and Bayesian classification—in our examination of the value of upgrading to the 16G Dell EMC PowerEdge R7625 server powered by 4^th Gen AMD EPYC processors along with Broadcom NICs and PERC 11 storage controllers. Using these workloads, we measured the throughput and speed of the servers. A server with better throughput and speed can process more data, handle more concurrent users, handle heavier workloads, and improve response times.

About Dell EMC PowerEdge R7625 servers

The Dell EMC PowerEdge R7625 server we tested features two AMD EPYC^™ 9554 processors that each contain 64 cores and a Broadcom BCM5720 NIC. According to Dell, “the PowerEdge R7625 is a highly scalable two-socket, 2U rack server packed with 50 percent more cores and up to 6 GPUs in a package that combines powerful performance and flexible configuration.”[v] According to Dell, the R7625 features:

• “Up to two 4th Gen AMD EPYC processors with up to 96 cores
• Available with either liquid or air-cooled configurations
• Low-latency storage options”[vi]

How we tested

We tested the following configurations:

• One 16G Dell PowerEdge R7625 server powered by 4^th Gen AMD EPYC 64-core processors along with Broadcom NICs and PERC 11 storage controllers
• One 15G Dell PowerEdge R7525 server powered by 3^rd Gen AMD EPYC 64-core processors along with Broadcom NICs and PERC 10 storage controllers

We configured both systems at maximum RDIMM capacity. The R7625 has a higher maximum capacity at 3TB and higher speed RAM at 4800 MT/s than the R7525 at 2TB and 3200MT/s, which is a useful upgrade for processing memory-intensive Spark workloads. We used Red Hat^® OpenShift^® virtualization. OpenShift is an open-source, Kubernetes-based container platform that offers a set of tools to manage, scale, and deploy containerized applications. For our deployment of OpenShift, we used a single-node deployment mode which is a new feature that is meant for proof of concept type environments. A typical OpenShift deployment uses three or more servers in a clustered configuration.

On each system, we created 10 OpenShift VMs with 24 cores, 96GB RAM, and one OpenShift VM with 12 cores, 32GB RAM, and one 30GB storage volume. We used this network for Spark cluster communications and Spark testing. We used Red Hat Enterprise Linux^® 8 for the OS and installed Java^™ 1.8.0, Python2^®, and Apache Maven^® 3.5.4; Apache Spark 3.0.3 with the Apache Hadoop 3.2 libraries; Apache Hadoop 3.2.4 for its HDFS capabilities; and the HiBench testing framework, version 7.1.1 with updates up to June 12, 2023 from its GitHub repository. We configured the 12-core VM as the Spark primary, and as the Hadoop manager for HDFS. We configured the remaining 10 VMs as Spark workers and Hadoop data nodes for HDFS. We used the storage volume for both the OS and for HDFS. We ran HiBench Bayes and k-means workloads from the Spark primary VM. Below is a table showing a summary of the system configurations we used in testing. For more details about our testing and configurations, read the science behind the report.

Table 1: System configurations we used in testing. Source: Principled Technologies.

About 4^th Gen AMD EPYC 9554 processors

According to AMD, EPYC 9554 processors deliver fast performance “for cloud, enterprise, and HPC workloads- helping accelerate your business.”[vii] EPYC processors include AMD Infinity Guard, which per AMD is “a set of layered, cutting-edge security features that help you protect sensitive data and avoid the costly downtime cause by security breaches.”[viii]

The EPYC 9554 has support for AVX512 processor extensions that speed up AI inference, including the use of the BFloat 16 data type (AVX512_BF16), and Vector Neural Network Instructions (AVX512_VNNI). In contrast, the EPYC 7763 processor has no support for AVX512 instructions.

In addition to performance and security features, AMD claims their processors are energy-efficient, which can reduce energy costs and “minimize environmental impacts from data center operations while advancing your company’s sustainability objectives.”[ix]

For more information about 4th Gen AMD EPYC processors visit: https://www.amd.com/en/processors/epyc-server-cpu-family.

About the HiBench benchmark suite

According to its GitHub repository, the HiBench benchmark suite “is a big data benchmark suite that helps evaluate different big data frameworks in terms of speed, throughput and system resource utilizations.”[x] The HiBench benchmark suite offers performance testing for 29 different types of workloads, including the machine learning algorithms associated with Bayesian Classification (Bayes) and k-means clustering.

Our results

K-means clustering

For large data sets, it isn’t possible for a human to analyze the data as efficiently or effectively as a machine learning algorithm can. K-means clustering is a machine learning algorithm that aims to group similar or dissimilar data points together in clusters. By finding similarities between data points that wouldn’t be obvious with other means of analysis, k-means clustering can unlock valuable insights into individual data points, whether they are about the customers of a business, the manufacturing processes of a factory, or some other aspect of a business. These insights could help an e-commerce company offer promotions to similar types of customers or help an insurance company detect anomalies or fraud. Using the latest generation of server technology has the potential to help businesses unlock these actionable data insights faster. Tools like RapidMiner^®, ELKI, Orange, Weka^®, and MATLAB^™ rely on k-means clustering for some of types of calculations.

To better understand how upgrading server technology might benefit organizations that use k-means clustering to analyze their data, we used the HiBench benchmark suite to compare the k-means performance in terms of throughput (megabytes per second) and speed (seconds). As Figures 1 and 2 show, the new Dell PowerEdge R7625 server outperformed the previous-generation server in both measurements. The latest-generation server had 70.0 percent higher throughput and completed the k-means workload 41.2 percent faster than the previous-generation device.

These results suggest that organizations that frequently use k-means clustering to gain insights might benefit from upgrading their older servers. For an e-commerce company that provides personalized product recommendations to millions of users based on data, better throughput and faster k-means speed could allow them to tailor their recommendations more quickly. Faster throughput and speed could allow the e-commerce company to update their clustering model more frequently so that it adapts to changing customer behavior in real time. These improvements could lead to more customer engagement and higher sales.

Figure 1: A comparison of the k-means throughput of the two servers in megabytes per second. Higher is better. Source: Principled Technologies.

Figure 2: A comparison of the times, in seconds, that the two servers took to complete the test k-means workload. Lower is better. Source: Principled Technologies.

Bayesian classification

Bayesian classification (or Bayesian inference) is a method of estimating the probability of an outcome and calculating the uncertainty around this probability using historical data. By analyzing prior outcomes, Bayesian machine learning can give organizations a statistical probability for a future outcome. A retailer may want to know the probability of a customer making a purchase after receiving a coupon code, for example. More advanced applications of Bayesian inference have helped scientists develop new drugs and assign probability to the accuracy of diagnostic tests.[xi]^,[xii] Being able to quickly analyze data sets for predictions about the future can be a powerful tool for businesses and organizations.

To evaluate the Bayesian analysis performance of the servers, we used the HiBench benchmark suite to compare the total throughput, measured in megabytes per second, and the speed of analysis, in seconds. As Figure 3 shows, the 16G Dell PowerEdge R7625 achieved 19.5 percent more throughput than the previous-generation server. As Figure 4 shows, the new server was 16.3 percent faster at completing the Bayesian classification workload than the previous-generation server we compared it to.

These results indicate just how much organizations that use Bayesian machine learning to make probabilistic calculations might benefit from upgrading their aging servers. For a financial services company that uses Bayesian analysis to make investment decisions and assess risk, higher throughput and speed could allow them to handle larger data sets and run more complex models to make more accurate, real-time decisions. Alternatively, a healthcare system that uses Bayesian models for diagnosis and treatment could update patient models faster and more frequently, leading to more accurate diagnosis and better health outcomes for patients.

Figure 3: A comparison of the Bayes throughput of the two servers in megabytes per second. Higher is better. Source: Principled Technologies.

Figure 4: A comparison of the times, in seconds, that the two servers took to complete the test Bayes workload. Lower is better. Source: Principled Technologies.

Performance and value – How these results can impact the bottom line

With any decision to upgrade a server environment, companies want to know that their upfront investment in new technology provides opportunities to save money further down the road. New technologies come at a price, but improvements in performance and efficiency can pay off in the long run.

Organizations can potentially save money by consolidating older servers with higher-performing, newer servers that each do more work. In our testing, a single Dell PowerEdge R7625 outperformed the Dell PowerEdge R7525 by up to 70 percent, completing 1.7 times as much k-means work as a single PowerEdge R7525. This means that two PowerEdge R7625 servers could process 3.4 times as much k-means work as one PowerEdge R7525 server. In other words, two PowerEdge R7625 servers can process the same amount of work as three PowerEdge R7525 servers with an additional 40 percent headroom. Thus, an organization that upgrades the servers in their data centers could likely reduce the total number of servers and still process the same workloads.

For each server a company can consolidate onto new gear, they can reduce their licensing cost for Red Hat OpenShift Platform Plus licensing costs for a standard 1-year subscription by $10,178.99 or by $27,820.99 for a standard 3-year subscription.[xiii]^,[xiv] These savings don’t even take into account premium subscriptions or additional support add-ons, which would further reduce annual licensing and support costs. By reducing server counts, companies could also find savings in the reduction of cooling costs, power costs, and data center footprints. As the number of servers in a data center scales, so too do the savings associated with upgrading to the latest-generation PowerEdge R7625 servers.

About Broadcom Gigabit Ethernet BCM5720 Controller

The Dell PowerEdge servers we tested feature Broadcom Gigabit Ethernet BCM5720 controllers. According to Broadcom, its 1G Ethernet Controllers are “the ideal solution for multicore servers, delivering full line-rate throughput across all ports.”[xv]

The BCM5720 Dual-Port 1GBASE-T PCle 2.1 Ethernet Controller is a 13th generation 10/100/1000BASE-T and 10/100/1000BASE-X Ethernet LAN controller solution. The host interface supports a separate PCle function for each LAN interface and the controller includes I/O Virtualization (IOV) features such as 17 receive and 16 transmit queues, and 17 MSI-X vectors with flexible vector-to-queue association. These IOV features enable the BCM5720 to support the VMware^® NetQueue and Microsoft VMQ technologies.[xvi]

Broadcom also states that this controller has “a comprehensive set of hardware features that the system may use to implement IEEE 1588 or IEEE 802.1AS-based time synchronization. These hardware features include a high-precision clock, timestamp registers for receive/transmit packets, and programmable trigger inputs and watchdog outputs.”[xvii]

Learn more at https://www.broadcom.com/products/ethernet-connectivity/network-adapters/bcm5720-1gbase-t-ic.

About Broadcom PERC 11 PERC H755N controllers

The PERC11 series of adapters presents a diverse range of notable features. It ensures dependable, high-performance, and fault-tolerant management of the disk subsystem. These adapters possess extensive RAID control capabilities, offering support for multiple RAID levels, such as 0, 1, 5, 6, 10, 50, and 60.[xviii] This facilitates efficient data safegaurding and redundancy mechanisms within the system.

Regarding compatibility, the PERC11 adapters conform to the Serial Attached SCSI (SAS) 3.0 standard, which facilitates a maximum data throughput of 12 Gb/s. This adherence ensures streamlined data transfer and seamless operations within the storage environment. Furthermore, the adapters boast extensive compatibility with a wide array of storage devices. They seamlessly integrate with Dell-qualified Serial Attached SCSI (SAS) and SATA hard drives, solid-state drives (SSDs), and PCle SSDs (NVMe). This versatility empowers users to leverage diverse storage options that align with their specific requirements and preferences.

Conclusion

As data proliferates and the sizes of databases grow, the potential to unlock valuable insights from them becomes increasingly dependent on fast architectures that can handle compute-intensive machine learning workloads such as k-means clustering and Bayesian inference. By upgrading to the latest servers, organizations can scale their processing power to meet the growing demands of their databases.

Larger databases and more powerful algorithms have the potential to give organizations a competitive edge. Faster servers can improve the accuracy of data-driven decisions by allowing organizations to use more complex algorithms and update ML models more frequently. To consider just two examples, improved performance could allow an e-commerce company to make better recommendations to customers and a financial services company to assess risks more accurately.

When we compared the machine learning performance of a 16G Dell PowerEdge R7625 server powered by 4^th Gen AMD EPYC 64-core processors with Broadcom NICs and PERC 11 storage controllers to a previous-generation PowerEdge server, we found performance enhancements in terms of throughput and speed, whether running k-means clustering or Bayesian workloads. These findings suggest that organizations that rely on machine learning algorithms might gain performance advantages by upgrading to the latest generation of these Dell servers.

This project was commissioned by Dell Technologies.

September 2023

Principled Technologies is a registered trademark of Principled Technologies, Inc.

All other product names are the trademarks of their respective owners.

A cluster of these servers ran a mix of applications with up to 27 percent better application performance than a previous-generation cluster, which could allow companies to do a given amount of work with fewer servers

Introduction

COVID-19 forced many small or medium-sized businesses (SMBs) to make changes, such as shifting to new markets or moving portions of their business online. Given the overall mood of uncertainty during the pandemic, some companies chose to delay technology purchases. Supply chain issues also affected the availability of some hardware. As conditions have stabilized, however, decision makers may be looking at the legacy gear in their data centers and questioning its ability to meet current requirements.

When upgrading, purchasers have a choice between investing in the latest-generation hardware or trying to reduce their capital expenditure (CAPEX) by going with previous-generation gear. To help those in this position understand the implications of both options, Principled Technologies conducted a series of tests on two three-node Microsoft Windows Server 2022 clusters with Hyper-V and Storage Spaces Direct. One cluster used previous-generation single-socket 15G Dell^™ PowerEdge^™ R7515 servers powered by 3^rd Gen AMD EPYC 7543P processors; the other used latest-generation single-socket 16G Dell PowerEdge R7615 servers powered by 4^th Gen AMD EPYC^™ 9354P processors along with Broadcom^® network interface cards (NICs) and PERC 11 storage controllers. We measured each cluster’s capabilities by making it simultaneously handle a database workload, a container-based application, and a web app—a mix of workloads similar to the ones that many SMBs run.

On all three workloads, the new cluster demonstrated significant performance advantages over the previous-generation cluster, to the point where you would need fewer new servers to do a given amount of work. With software licensing being such a large expense, the savings you would reap from being able to eliminate one server could more than offset the purchase price of the new servers. This would help your company deliver a better experience to end users while also lowering other costs, such as power and cooling and IT staff time for maintenance.

Considerations SMBs face when upgrading IT infrastructure

When preparing to replace their outdated servers with modern ones, small and medium-sized businesses face a wide range of challenges, but three common ones are cost, staffing, and equipment longevity.

IT budgets are limited, and it can be easy to underestimate the true cost of new gear if decision makers account for only the CAPEX of the hardware purchase. Companies should also consider the ongoing operating expenditures (OPEX) involved with servers, such as rack space and power for servers, IT staffing resources, and the most expensive item: software licensing.

Researching technology solutions, deploying servers, and providing support once the new equipment is up and running can all be extremely time-consuming tasks. By choosing a solution that minimizes these IT burdens, companies can free their in-house admin teams to take care of other needs or limit costs for third-party IT.

Choosing a server solution that is a good fit for the unique needs of your business can feel like walking a tightrope. On the one hand, you want to avoid overinvesting in technologies with capabilities that exceed the requirements of your workloads. On the other hand, underinvesting can also be a mistake, leaving you with servers that lack the power and reliability necessary for mission-critical workloads for the lifespan of the new equipment, cannot handle future growth well, and risk delivering an unsatisfactory experience to both customers and employees. An underpowered solution could have a shorter lifecycle, which would put you back at square one of the decision-making process sooner. Perhaps the greatest downside to choosing a previous-generation solution is that doing so can require you to purchase operating system and application software licenses for an additional server.

All these considerations make it very important to take time to assess your current and future needs, such as the types of workloads you run, the number of customers and employees you support, and the growth you anticipate. By doing so, you greatly improve the likelihood of selecting a cost-effective hardware solution that will suit your needs for the life span you hope the solution to have.

The limitations of cloud

Before we dive into data center upgrades, we must consider the cloud. While many companies have shifted business applications to the cloud, there are potential disadvantages and limitations, which you should weigh against the convenience of this approach. These include security concerns, dependence on the internet, lack of control of resources, occasional downtime, vendor compatibility, and cost.

Security concerns

While cloud service providers (CSPs) typically apply multiple security measures to keep their cloud infrastructure safe from attack, data breaches do occur. For instance, a 2021 flaw in the Microsoft Azure Cosmos DB database resulted in customer information being exposed to hackers.[i] While threats such as this one do not make cloud computing entirely insecure, they demonstrate “a higher chance of successful attacks or data breaches when there is human error in cloud setup and issues with endpoint configurations.”[ii]

Limited flexibility and control over resources

Cloud providers typically do not allow business owners to manage and monitor the hardware in their cloud environment. This limits the visibility into potential future problems or hardware failures, leaving the business completely reliant on the cloud provider’s planning and reliability. CSPs can also place limits on the tools, applications, and data that customers can deploy on cloud servers.[iii]

Occasional downtime

When cloud servers go down, forcing users to wait until a connection is restored, businesses can lose customers and revenue.[iv] One example of downtime affecting cloud-based businesses was the hour-long 2020 blackout of all Google services.[v] This type of downtime may be rare, but it can have an enormous negative impact.

Vendor compatibility issues

Transitioning from one CSP to another is not necessarily a seamless experience. Applications working properly in one cloud platform will not always be compatible with another provider’s platform, a risk that can make decision makers feel “locked in” with a single provider.[vi]

Cost

A company’s monthly CSP bill increases along with usage, making cloud potentially very expensive. As Wang and Casado outline in the Andreesen Horowitz paper “The Cost of Cloud, a Trillion Dollar Paradox, ”paying a “flexibility tax” for the public cloud often makes good business sense early in a company’s journey, but can lead to large OPEX outlays that can offset the flexibility benefits.[vii]

One company that left the cloud for economic reasons was project management platform Basecamp. In October 2022, Basecamp CTO David Heinemeier Hansson wrote, “Renting computers is (mostly) a bad deal for medium-sized companies like ours with stable growth. The savings promised in reduced complexity never materialized.”[viii]

A look at the applications SMBs are running in their data centers

These downsides of the cloud are some of the reasons decision makers run certain applications in their data centers. Another reason is the nature of the applications themselves. For example, companies may choose to keep internal applications such as company portals and human resources applications on servers that are on site. We tested with three types of applications companies might place on on-site servers.

Kubernetes containers for containerized apps

A container is a unit of software packaged with everything required to run that software in a standalone state, including binaries, libraries, dependencies, and of course, the application itself. Kubernetes^® is an open-source platform for deploying and managing applications that run in containerized environments.

Organizations deploy applications in Kubernetes containers for scalability and flexibility; containers also give them the ability to burst to cloud when necessary. Thanks to software improvements, Kubernetes technology has become more accessible in recent years. Running your containerized applications on high-performing servers is a win because the smaller footprint of containerized applications lets you take better advantage of the increased resources of those servers.

Kubernetes containerized environments can allow organizations to offer a high-quality user experience for multi-tiered web applications, such as those for online auctions and ecommerce.

WordPress for websites

Websites are a critical resource for many small and medium-sized businesses, and WordPress is an extremely popular web platform for businesses of all shapes and sizes. According to WordPress, “More bloggers, small businesses, and Fortune 500 companies use WordPress than all other options combined.”[ix] Almost as important as having a website is having it perform well in terms of speed and responsiveness. For example, if your site takes more than 3 seconds to load, 40 percent of potential customers will abandon it.[x]

The WordPress platform provides a way for companies to have a web presence, which is obviously a vital component of success because web searches may well be the way most customers will find businesses. Because users expect web pages to load quickly and have little patience when they fail to do so, strong WordPress performance can translate to attracting and keeping customers, while poor WordPress performance can cause you to lose customers before they even see your site.

Online transaction processing (OLTP) databases that underlie many critical business applications

For SMBs, OLTP databases are essential tools for organizing and tracking customers, inventory, employees, and finances. Examples of OLTP databases include:

• Customer relationship management (CRM) databases maintain all the information about accounts, contacts, leads, and opportunities. The record for an individual customer might track not only contact information and order history, but also details of service calls and more. Companies can also use CRM databases to manage marketing and promotions, export email addresses, and generate shipping labels.[xi]
• Inventory tracking databases help companies keep tabs on how much inventory they have and where it is located. The database can include integrated bar codes and scanners that employees can use to track and monitor items as they travel from one location to another, and can send alerts when supplies of critical items are running low.[xii]
• Payroll and scheduling databases keep track of employee data, such as wages, accrued vacation time, and benefits.[xiii]

One such OLTP database application is Microsoft SQL Server, a widely recognized relational database management system (RDBMS) that utilizes the SQL programming language. At the center of its architecture is the Database Engine, a relational engine for query processing and a storage engine for database file and index management. It also includes other data-related services such as SQL Server Integration Services (SSIS), SQL Server Analysis Services (SSAS), and SQL Server Reporting Services (SSRS).

For those running these and other SMB database applications, performance is an important element of effectiveness. Servers that deliver database results quickly mean less waiting and frustration for your employees as they perform their jobs, and putting important information into the hands of decision makers sooner.

Testing this mixed, multi-application workload

To help small and medium-sized businesses considering upgrading their legacy servers, we conducted testing using two different three-node Microsoft Windows Server 2022 clusters with Hyper-V and Storage Spaces Direct:

• One using previous-generation single-socket 15G Dell PowerEdge R7515 servers with 3^rd Gen AMD EPYC processors
• One using current-generation 2U, single-socket 16G Dell PowerEdge R7615 servers powered by 4^th Gen AMD EPYC processors along with Broadcom NICs and PERC11 storage controllers

We chose the PowerEdge R7615 for a number of reasons. As a 2U rack server, it offers better storage options than a 1U server. The fact that it uses single-socket processors provides a financial advantage over multi-socket servers in terms of both its purchase price and its licensing requirements. Any software that uses a per-socket licensing structure will be less expensive to license. We configured the PowerEdge R7615 servers with PERC11 storage controllers because of their effect on both redundancy and performance. We selected 16G servers with AMD EPYC 9354P processors because they strike a balance between strong performance and optimized cost and because 32 cores is a sweet spot for licensing. The 9354P is also less expensive than the two-socket 9354 version of the processor.[xiv]

Our mixed workload included a Microsoft SQL Server database component, a multi-tier web app (Weathervane) on Kubernetes, and a WordPress component. All applications ran simultaneously to simulate an organization using a single cluster of three servers to run multiple concurrent applications.

Table 1 shows the server hardware we used, Table 2 shows the software we used, and Figure 1 shows a diagram of our test bed. Note that given the differences in memory channel architecture between the two server generations, we could not match the RAM capacities while also configuring the systems in a balanced, optimized configuration. We chose to ensure a balanced configuration to optimize for performance. As a result, the 16G servers had a greater memory capacity than the 15G servers.

Table 1: Server configuration information.

Table 2: Software we used.

About Dell PowerEdge R7615 servers

The Dell PowerEdge R7615 is a 2U, single-socket rack server. Dell states that it has designed this server to provide “performance and flexible, low-latency storage options in an air or Direct Liquid Cooling (DLC) configuration.”[xv]

According to Dell, this server uses the AMD EPYC 4^th generation processor to deliver up to 50 percent higher core count per single-socket platform in an innovative air-cooled chassis and supports DDR5 at 4800 MT/s memory and PCIe^® Gen5 with double the speed of previous Gen4 for faster access and transport of data, optimizing application output.[xvi] It supports up to six single-wide full-length GPUs or three double-wide full-length GPUs to improve responsiveness or reduce app load time for power users and supports lower-latency, high-performance NVMe SSDs in a hardware RAID solution to help maximize compute performance.[xvii]

Learn more at https://www.delltechnologies.com/asset/en-us/products/servers/technicalsupport/poweredge-r7615-spec-sheet.pdf.

Figure 1: Test bed diagram. Source: Principled Technologies.

SQL Server with DVD Store workload

We conducted a series of tests on an OLTP workload that we set up using the DVD Store 3 benchmarking tool.[xviii] DVD Store, an open-source test and benchmark tool, emulates an online store specializing in DVD sales. The test utility simulates customers logging in, browsing products by title or author, accessing reviews, submitting new reviews, rating existing reviews, signing up for premium membership, and making purchases.

To gauge performance, the benchmarking tool generates a metric of orders placed per minute. For our testing, we generated a pre-sized database elsewhere, then restored that database backup in our environment for testing.

About AMD EPYC 9354P processors

Part of the Platform Server Product Family and the AMD EPYC 9004 Series, these 32-core, 64-thread processors have a maximum boost clock of 3.8GHz, an all-core boost speed of 3.75GHz, a base clock of 3.25GHz, and a 256MB L3 cache.[xix] These are the single-socket versions of the 64-core processors that are cost optimized for single-socket servers.

Learn more at https://www.amd.com/en/products/cpu/amd-epyc-9354p.

Kubernetes with Weathervane 2.1 benchmark

According to VMware, “Weathervane 2.1 is an application-level performance benchmark which lets users investigate the performance characteristics of on-premises and cloud-based Kubernetes clusters… by deploying one or more applications on the cluster and then driving a load against those applications.”[xx] Weathervane uses a multi-tier web application that includes both stateless and stateful services. The Weathervane benchmark provides a variety of pre-tuned configurations (i.e., deployment sizes) for the app, allowing users to select a configuration appropriate for their cluster sizing. The Weathervane workload driver generates the load and runs on a Kubernetes cluster. Users can configure Weathervane “to generate a steady load using a fixed number of simulated users, or to automatically vary the number of users to find the maximum number that can be supported on the cluster without violating quality-of-service (QoS) requirements.”[xxi]

In the fixed-load scenario, Weathervane gives the test a passing score only if the run completes without violating the QoS requirements. In the maximum-user scenario, Weathervane reports the highest number of simulated users that completed the test without violating the QoS requirements. Weathervane refers to this number as the peak WvUsers. In our testing, we used the fixed-load scenario to allow us more control over system resource utilization while running our three different workloads. We ran one Kubernetes cluster using Docker on one VM per node for a total of three Kubernetes clusters per physical cluster. We then deployed an instance of the Weathervane workload to each Kubernetes cluster.

WordPress with Siege benchmark

Siege is an open-source HTTP load testing benchmark utility designed to measure a website or multiple websites performance under stress. It can test a single URL with a set number of simulated users, or can read multiple URLs into memory and stress them simultaneously.[xxii] According to the Siege GitHub page, “Siege supports HTTP/1.0 and 1.1 protocols, the GET and POST directives, cookies, transaction logging, and basic authentication. Its features are configurable on a per user basis.”[xxiii] In our testing, we used Siege to target a default WordPress install on Ubuntu 20.04. We ran the test for 30 minutes and report the average transactions per second.

Our test configuration

We configured each cluster with the same number of VMs:

• Six OLTP (SQL Server) VMs, with two on each physical node
• Six WordPress VMs, with two on each physical node
• Three Weathervane VMs, with one on each physical node. Each of the three VMs ran containers in a single Kubernetes cluster, with a single Weathervane instance targeting that Kubernetes cluster.

 We sized the VM memory to mostly fill the host capacity (192 GB and 128 GB for the 16G server and the 15G server, respectively). Table 3 provides details of our test configuration.

Table 3: Details of our test configuration.

We ran the following parameters:

Database (DVD Store 3)

• 16 threads
• 5ms think time
• 30s run time
• 24s warmup time
• 10 users per second ramp rate 

WordPress

• 25 users per VM
• 0ms think time

Weathervane

• 2,500 users per VM

About Broadcom PERC11 PERC H755N controllers

The PERC11 series of adapters presents a diverse range of notable features. To begin with, it ensures dependable, high-performance, and fault-tolerant management of the disk subsystem. These adapters possess extensive RAID control capabilities, offering support for multiple RAID levels, such as 0, 1, 5, 6, 10, 50, and 60.[xxiv] This facilitates efficient data safeguarding and redundancy mechanisms within the system.

Regarding compatibility, the PERC11 adapters conform to the Serial Attached SCSI (SAS) 3.0 standard, which facilitates a maximum data throughput of 12 Gb/sec. This adherence ensures streamlined data transfer and seamless operations within the storage environment. Furthermore, the adapters boast extensive compatibility with a wide array of storage devices. They seamlessly integrate with Dell-qualified Serial Attached SCSI (SAS) and SATA hard drives, solid-state drives (SSDs), and PCIe SSDs (NVMe).[xxv] This versatility empowers users to leverage diverse storage options that align with their specific requirements and preferences.

What testing revealed

In the sections below, we present the findings of the three workloads we ran simultaneously on our two clusters, each of which comprised three servers. We identified the highest-performing server in each cluster and present the results that server achieved on each of our three workloads.

Kubernetes/Weathervane

As we noted earlier, our Weathervane testing consisted of a fixed-user scenario with the same number of WvUsers on both clusters. With an almost identical throughput rate, response time on the highest-performing new Dell PowerEdge R7615 server was half that of the highest-performing previous-generation Dell PowerEdge R7515 server (see Figure 2). This performance advantage could translate to higher numbers of supported users, or lower latencies for a fixed set of users, improving user experience due to reduced response time while interacting with the site.

We identified the response time from the single Weathervane application on the best-performing server and present that time here.

Figure 2: Weathervane response time on the highest-performing server in each cluster. Lower is better. Source: Principled Technologies.

WordPress/Siege

After we ran the Siege benchmark to measure WordPress performance, we added the transactions per second from the two VMs on the best-performing server in each cluster and present those sums here.

As Figure 3 shows, the highest-performing server in the cluster of new Dell PowerEdge R7615 servers achieved a rate of WordPress requests per second that was 27.4 percent higher than that of the highest-performing server in the previous-generation cluster. This performance advantage could translate to speedier load times, which would position your business much better in the competitive landscape where “88% of online users won’t return to a site after a bad experience.”[xxvi]

Figure 3: Total WordPress transactions per second on the highest-performing server in each cluster. Higher is better. Source: Principled Technologies.

SQL Server/DVD Store 3

After we ran the DVD Store 3 benchmark to measure SQL Server database performance, we added the orders per minute from the two VMs on the best-performing server in each cluster and present those sums here.

As Figure 4 shows, the highest-performing server in the cluster of new Dell PowerEdge R7615 servers achieved a rate of OPM that was 24.7 percent higher than that of the highest-performing server in the previous-generation cluster. This performance advantage could translate to speedier and more responsive behavior on the part of many business database applications, such as those we noted earlier—customer relationship management, inventory, and business data analysis.

Figure 4: Total DVD Store 3 transactions per second on the highest-performing server in each cluster. Higher is better. Source: Principled Technologies.

Why was performance on the new Dell PowerEdge R7615 cluster better?

Any time you undertake a system upgrade such as the one in our test scenario, multiple factors work together to improve performance. In our testing, we saw clear advantages of the Dell PowerEdge R7615 with Dell PowerEdge RAID Controller 11 cluster on the mixed workload we tested. We can attribute a portion of this improvement to this solution’s use of latest 4^th Gen AMD EPYC processors, which have a base CPU frequency of 3.25 GHz and support up to 4800 MT/s DDR5 RAM, a considerable improvement over the 2.80 GHz base CPU frequency and 3200 MT/s DDR4 RAM of the older AMD EPYC processors in the previous-generation servers. If we compare the SPEC^®2017 test results for the Dell PowerEdge R7515 and Dell PowerEdge R7615 with the same processors our test servers used, we see increases ranging from 33 percent on Integer Base to 66 percent on Floating Point Base.[xxvii]

In addition to its more powerful processor, the Dell PowerEdge R7615 also has faster and more RAM with DDR5 and supports 24Gbps SAS storage. (Note that both solutions used the same SAS storage drives, which are rated for 24Gbps SAS data transfer speeds. However, the previous-generation PowerEdge R7515 supported only up 12Gbps SAS, while the PowerEdge R7615 could run at the full 24Gbps rate.)

Spotlight on licensing costs

While increased performance is a major decision in any server purchase, SMBs must also consider cost. The CAPEX of purchasing gear is unavoidable, but how does the choice of server model affect software licensing?

To answer this, we use pricing as of June 30, 2023. Let’s first look at the operating system software licensing. For Windows Server 2022 Datacenter edition, customers can purchase core-based licensing in 16-core packs for $6,155.28.[xxviii] In our testing, each previous- and current-generation server contained one 32-core processor. Therefore, if a customer were purchasing new OS licenses for either environment, they would need two of these license packs, for a total of $12,310 per server ($36,930 per cluster).

Next, let’s look at SQL Server 2022 Enterprise licensing costs. In a virtualized environment, customers have two choices: They can license all cores on a server or, if they are enrolled in the Software Assurance program, they can license by the number of vCPUs per SQL Server VM. Enrolling in the Software Assurance program offers several advantages, including software upgrades at no additional cost. Because our performance testing used only a fraction of the CPU threads for SQL Server, we are assuming enrollment in Software Assurance and using the vCPU-based pricing. Each test server had two SQL Server VMs with 10 vCPUs each, for a total of 20 vCPUs needing licenses, or 60 vCPUs per three-node cluster. SQL Server Enterprise comes in a two-core pack for $15,123.[xxix] Each cluster would need 30 of these licenses, for a total of $453,690.

As Table 4 shows, the total cost to license three servers is $613,275. Dividing this figure by three gives us $204,425, the total per-server licensing cost. After the first year, annual Software Assurance costs for a single server would be $40,885.

Table 4: Licensing and software assurance costs as of July 14, 2023.

The remaining workloads used open-source software such as Ubuntu, WordPress, and Tanzu Community Edition, which are all free. While numerous support and security packages are available for these open-source solutions, we are excluding them from this analysis.

The costs above assume that customers are purchasing the license as a part of their CAPEX investment. However, customers can also choose to transfer licenses from existing servers and continue paying annual OPEX fees related to the software. As we mentioned earlier, we assume customers are enrolled in the Microsoft Software Assurance program, which provides the added benefit of fine-tuning the licensing costs related to SQL Server by licensing vCPUs instead of whole CPUs, as well as the benefit to upgrade to major software versions at no additional cost. A ComputerWorld article discusses the many additional benefits to the program.

In our cost analysis, we include Software Assurance for both Windows Server and SQL Server. The annual cost of Software Assurance for enterprise software is 25 percent of licensing costs.[xxxiii] In our comparison, the total licensing costs for each of our test clusters is $490,620, which would incur an annual fee of $122,655 if the company chose to maintain the licenses with Software Assurance. This annual fee, like all the licensing fees we have discussed, is identical for the previous-generation cluster and the current-generation cluster.

About the Broadcom Ethernet controllers in our test servers

Our testing used Broadcom Gigabit Ethernet BCM5720 and Broadcom 57414 Dual Port 10/25GbE SFP28 NICs.

The BCM5720 Dual-Port 1GBASE-T PCIe 2.1 Ethernet Controller is a 13th generation 10/100/1000BASE-T Ethernet LAN controller solution. According to Broadcom, the BCM5720 “provides a PCI Express^® v2.0-compliant host interface, which can operate at 5 GT/s or at 2.5 GT/s at x2 link width.” It also has “I/O Virtualization (IOV) support for VMWare^® NetQueue and Microsoft^® VMQ” and also supports Energy Efficient Ethernet.[xxxiv]

Learn more at https://www.broadcom.com/ products/ethernet-connectivity/network-adapters/bcm5720-1gbase-t-ic.

The BCM 57414 Dual Port 10/25GbE SFP28 controller features two network interface ports that support both SFP28 for 25Gb/s speeds and SFP+ for 10Gb/s modules. According to Broadcom, these NICs are ideal for supporting both on-premises data centers and cloud computing backends. The BM57414 also supports advanced networking features such as SR-IOV, vSwitch acceleration, TruFlow^™ flow processing, and RDMA over converged Ethernet (RoCE), the last of which we used for our Storage Spaces Direct backend.[xxxv]

For more information see : https://docs.broadcom.com/doc/957414A4142CC-DS.

How improved performance can lead to needing fewer servers, which in turn reduces licensing costs

Earlier, we discussed how the superior performance of the Dell PowerEdge R7615 on our mixed workload could improve business outcomes by delivering a speedier experience for end users, whether they are employees or current or potential customers. We then looked at licensing costs and saw that an equal number of previous-generation Dell PowerEdge R7515 servers and current-generation Dell PowerEdge R7615 servers would have the same per-server cost for Windows Server, SQL Server, and Software Assurance.

Another enormous potential benefit of choosing the current-generation Dell PowerEdge R7615 is the savings that result from a lower server count. Being able to perform a given amount of work with fewer servers can not only lead to savings on OPEX such as power and cooling and IT staffing resources, but it can reduce licensing costs as well.

Let’s take the performance results we saw on the SQL Server workload and use them as a rough proxy for the different performance levels of the two server models we tested and server counts a hypothetical company might require depending on which generation it chose. Based on the number of database orders per minute the highest-performing servers in each cluster achieved, we can set a performance level, such as approximately 90,000 OPM, that a company needs to achieve to meet service-level agreements or other criteria. Given this hypothetical requirement, a company could purchase only three 16G Dell PowerEdge R7615 servers rather than the four 15G Dell PowerEdge R7515 servers that would be necessary to perform the same level of work. Having one fewer server would save the company over $200,000 on the first year of licensing and Software Assurance costs and an additional $40,000 every subsequent year. This savings would be more than enough to offset the higher purchase price of the 16G Dell PowerEdge R7615 server. Additionally, the company would spend less on power and cooling and IT management time.

Table 5: Licensing and software assurance costs as of July 14, 2023.

Conclusion

As you do your best to balance timing, budget, IT resources, and your current and anticipated server needs, consider how opting for newer servers could help your business. As our testing showed, there are clear benefits to choosing servers that support such workload requirements as keeping databases running at a quick pace and delivering speedy hosting for your business’s website. Plus, a solution that offers the capacity and software features to perform well while natively supporting Kubernetes containers could add value in terms of setup, flexibility, scalability, and cost-effectiveness. And you can achieve all of this and possibly reduce OPEX in the process.

In our testing with a mixed workload that reflects some of the needs common to small and medium businesses, a cluster of 16G Dell PowerEdge R7615 single-socket servers powered by 4^th Gen AMD EPYC processors outperformed a cluster of previous-generation 15G Dell PowerEdge R7515 servers, with improvements of up to 27 percent and latency reduction of up to 50 percent. These results show that upgrading to the new Dell solution can be a smart step toward meeting the needs of your users now and in the years to come.

This project was commissioned by Dell Technologies.

August 2023

Principled Technologies is a registered trademark of Principled Technologies, Inc.

All other product names are the trademarks of their respective owners.

Many factors can impact Oracle Database performance, and optimizing performance often requires a combination of strategies. Here are a few things that can help improve Oracle Database performance, but none more than the hardware:

• Hardware:  Ensure your hardware meets the recommended requirements and expectations for running the database. This includes CPU, memory, disk space, and raid controllers. Other factors can also include:
• Indexing: Create indexes on frequently accessed columns to improve query performance.
• Tuning: Regularly tune the database parameters to ensure optimal performance. This includes adjusting the buffer cache, sorting area, and network packet size.
• Partitioning: Consider partitioning large tables to improve query performance.
• Compression: Consider using compression to reduce the size of the database and improve I/O performance.
• Clustering: Implement clustering to distribute the load across multiple servers.
• Query optimization: Optimize SQL queries to reduce the number of I/O operations required.
• Data archiving: Archive old or infrequently accessed data to improve query performance.

Note that the best approach for improving Oracle Database performance will depend on your specific environment and hardware to enhance the workload.

This study analyzes the benefit of migrating from legacy Dell™ PowerEdge™ R7525 servers running Oracle Database to the Dell PowerEdge R7625, including AMD Epyc 4th generation processors equipped with a PERC 12.  We also analyzed the average CPU IOWait times.

We found that the Dell PowerEdge R7625 will let you support more customers and realize better system efficiency, which could lead to savings related to server consolidation.

• CPU-targeted: Retailers with a small list of tiered options with high throughput (1.46 x the New Orders Per Minute)
• IO-targeted (storage): Online retailers with medium-to-large product selections (2.39 X the new orders per minute)
• Balanced CPU/IO: Retailers with smaller product inventory and medium throughput (1.71 X the new orders per minute)