Dell Technologies has completed the successful submission of MLPerf Training, which marks the seventh round of submission to MLCommons™. This blog provides an overview and highlights the performance of the Dell PowerEdge R750xa, XE8545, and DSS8440 servers that were used for the submission.

What’s new in MLPerf Training v2.1?

This round of submission does not include new benchmarks or changes in the existing benchmarks. A change is introduced in the submission compliance checker. 

This round adds one-sided normalization to the checker to reduce variance in the number of steps to converge. This change means that if a result converges faster than the RCP mean within a certain range, the checker normalizes the results to the RCP mean. This normalization was not available in earlier rounds of submission.

What’s new in MLPerf Training v2.1 with Dell submissions?

For Dell submission for MLPerf Training v2.1, we included:

• Improved performance with BERT and Mask R-CNN models
• Minigo submission results on Dell PowerEdge R750xa server with A100 PCIe GPUs

Overall Dell Submissions

Figure 1.     Overall submissions for all Dell PowerEdge servers in MLPerf Training v2.1

Figure 1 shows our submission in which the workloads span across image classification, lightweight and heavy object detection, speech recognition, natural language processing, recommender systems, medical image segmentation, and reinforcement learning. There were different NVIDIA GPUs including the A100, with PCIe and SXM4 form factors having 40 GB and 80 GB VRAM and A30.

The Minigo on the PowerEdge R750xa server is a first-time submission, and it takes around 516 minutes to run to target quality. That submission has 4x A100 PCIe 80 GB GPUs.

Our results have increased in count from 41 to 45. This increased number of submissions helps customers see the performance of the systems using different PowerEdge servers, GPUs, and CPUs. With more results, customers can expect to see the influence of using different hardware settings that can play a vital role in time to convergence.

We have several procured winning titles that demonstrate the higher performance of our systems in relation to other submitters, starting with the highest number of results across all the submitters. Some other titles include the top position in the time to converge for BERT, ResNet, and Mask R-CNN with our PowerEdge XE8545 server powered by NVIDIA A100-40GB GPUs.

Improvement in Performance for BERT and Mask R-CNN

Figure 2.     Performance gains from MLPerf v2.0 to MLPerf v2.1 running BERT

Figure 2 shows the improvements seen with the PowerEdge R750xa and PowerEdge XE8545 servers with A100 GPUs from MLPerf training v2.0 to MLPerf training v2.1 running BERT language model workload. The PowerEdge XE8545 server with A100-80GB has the fastest time to convergence and the highest improvement at 13.1 percent, whereas the PowerEdge XE8545 server with A100-40GB has 7.74 percent followed by the PowerEdge R750xa server with A100-PCIe at 5.35 percent.

Figure 3.     Performance gains from MLPerf v2.0 to MLPerf v2.1 running Mask R-CNN

Figure 3 shows the improvements seen with the PowerEdge XE8545 server with A100 GPUs. There is a 3.31 percent improvement in time to convergence with MLPerf v2.1.

For both BERT and Mask R-CNN, the improvements are software-based. These results show that software-only improvements can reduce convergence time. Customers can benefit from similar improvements without any changes in their hardware environment.

The following sections compare the performance differences between SXM and PCIe form factor GPUs.

Performance Difference Between PCIe and SXM4 Form Factor with A100 GPUs

Figure 4.     SXM4 form factor compared to PCIe for the BERT

Figure 5.     SXM4 form factor compared to PCIe for Resnet50 v1.5

Figure 6.     SXM4 form factor compared to PCIe for the RNN-T

Table 1:

Figures 4, 5, and 6 and Table 1 show that SXM form factor is faster than the PCIe form factor for BERT, Resnet50 v1.5, and RNN-T workloads.

The SXM form factor typically consumes more power and is faster than PCIe. For the above workloads, the minimum percentage improvement in convergence that customers can expect is in double digits, ranging from approximately 12 percent to 40 percent, depending on the workload.

Multinode Results Comparison

Multinode performance assessment is more important than ever. With the advent of large models and different parallelism techniques, customers have an ever-increasing need to find results faster. Therefore, we have submitted several multinode results to assess scaling performance.

Figure 7.     BERT multinode results with PowerEdge R750xa and XE8545 servers

Figure 7 indicates multinode results from three different systems with the following configurations:

1. R750xa with 4 A100-PCIe-80GB GPUs
2. XE8545 with 4 A100-SXM-40GB GPUs
3. XE8545 with 4 A100-SXM-80GB GPUs

Every node of the above system has four GPUs each. When the graph shows eight GPUs, it means that the performance results are derived from two nodes. Similarly, for 16 GPUs the results are derived from four nodes, and so on.

Figure 8.     Resnet50 multinode results with R750xa and XE8545 servers

Figure 9.     Mask R-CNN multinode results with R750xa and XE8545 servers

As shown in Figures 7, 8, and 9, the multinode scaling results of the BERT, Resnet50, and Mask R-CNN are linear or nearly linear scaled. This shows that Dell servers offer outstanding performance with single-node and multinode scaling.

Conclusion

The findings described in this blog show that:

• Dell servers can run all types of workloads in the MLPerf Training submission.
• Software-only enhancements reduce time to solution for our customers, as shown in our MLPerf Training v2.1 submission, and customers can expect to see improvements in their environments.
• Dell PowerEdge XE8545 and PowerEdge R750xa servers with NVIDIA A100 with PCIe and SXM4 form factors are both great selections for all deep learning models.
• PCIe-based PowerEdge R750xa servers can deliver reinforcement learning workloads in addition to other classes of workloads, such as image classification, lightweight and heavy object detection, speech recognition, natural language processing, and medical image segmentation.
• The single-node results of our submission indicate that Dell servers deliver outstanding performance and that multinode run scales well and helps to reduce time to solution across a distinct set of workload types, making Dell servers apt for single-node and multinode deep learning training workloads.
• The single-node results of our submission indicate that Dell servers deliver outstanding performance and that multinode results show a well-scaled performance that helps to reduce time to solution across a distinct set of workload types. This makes Dell servers apt for both small training workloads on single nodes and large deep learning training workloads on multinodes.

Appendix

System Under Test

MLPerf system configurations for PowerEdge XE8545 systems

MLPerf system configurations for Dell PowerEdge R750xa servers

Introduction 

Graphics Processing Units (GPUs) provide exceptional acceleration to power modern Artificial Intelligence (AI) and Deep Learning (DL) workloads. GPU resource allocation and isolation are some of the key components that data scientists working in a shared environment use to run their DL experiments effectively. The need for this allocation and isolation becomes apparent when a single user uses only a small percentage of the GPU, resulting in underutilized resources. Due to the complexity of the design and architecture, maximizing the use of GPU resources in shared environments has been a challenge. The introduction of Multi-Instance GPU (MIG) capabilities in the NVIDIA Ampere GPU architecture provides a way to partition NVIDIA A100 GPUs and allow complete isolation between GPU instances. The Dell Validated Design showcases the benefits of virtualization for AI workloads and MIG performance analysis. This design uses the most recent version of VMware vSphere along with the NVIDIA AI Enterprise suite on Dell PowerEdge servers and VxRail Hyperconverged Infrastructure (HCI). Also, the architecture incorporates Dell PowerScale storage that supplies the required analytical performance and parallelism at scale to feed the most data-hungry AI algorithms reliably.

In this blog, we examine some key concepts, setup, and MLPerf Inference v1.1 performance characterization for VMs hosted on Dell PowerEdge R750xa servers configured with MIG profiles on NVIDIA A100 80 GB GPUs. We compare the inference results for the ResNet50 and Bidirectional Encoder Representations from Transformers (BERT) models.

Key Concepts

Key concepts include:

• Multi-Instance GPU (MIG)—MIG capability is an innovative technology released with the NVIDIA A100 GPU that enables partitioning of the A100 GPU up to seven instances or independent MIG devices. Each MIG device operates in parallel and is equipped with its own memory, cache, and streaming multiprocessors.

In the following figure, each block shows a possible MIG device configuration in a single A100 80 GB GPU:

  Figure 1- MIG device configuration - A100 80 GB GPU

The figure illustrates the physical location of GPU instances after they have been instantiated on the GPU. Because GPU instances are generated and destroyed at various locations, fragmentation might occur. The physical location of one GPU instance influences whether more GPU instances can be formed next to it.

Supported profiles for the A100 80GB GPU include:

• 1g.10gb 
• 2g.20gb 
• 3g.40gb
• 4g.40gb
• 7g.80gb 

In Figure 1, a valid combination is constructed by beginning with an instance profile on the left and progressing to the right, ensuring that no two profiles overlap vertically. For detailed information about NVIDIA MIG profiles, see the NVIDIA Multi-Instance GPU User Guide.

• MLPERF—MLCommons™ is a consortium of leading researchers in AI from academia, research labs, and industry. Its mission is to "develop fair and useful benchmarks" that provide unbiased evaluations of training and inference performance for hardware, software, and services—all under controlled conditions. The foundation for  MLCommons began with the MLPerf benchmark in 2018, which rapidly scaled as a set of industry metrics to measure machine learning performance and promote transparency of machine learning techniques. To stay current with industry trends, MLPerf is always evolving, conducting new tests, and adding new workloads that represent the state of the art in AI. 

Setup for MLPerf Inference

A system under test consists of an ESXi host that can be operated from vSphere.

System details

The following table provides the system details.

Table 1: System details

System configuration for MLPerf Inference

The configuration for MLPerf Inference on a virtualized environment requires the following steps:

1. Boot the host with ESXi (see Installing ESXi on the management hosts), install the NVIDIA bootbank driver, enable MIG, and restart the host.
2. Create a virtual machine (VM) on the ESXi host with EFI boot mode (see Using GPUs with Virtual Machines on vSphere – Part 2: VMDirectPath I/O) and add the following advanced configuration settings:
   
   

   pciPassthru.use64bitMMIO: TRUE
   pciPassthru.allowP2P: TRUE
   pciPassthru.64bitMMIOSizeGB: 64

3. Change the VM settings and add a new PCIe device with a MIG profile (see Using GPUs with Virtual Machines on vSphere – Part 3: Installing the NVIDIA Virtual GPU Technology).
4. Boot the Linux-based operating system and run the following steps in the VM.
5. Install Docker, CMake (see Installing CMake), the build-essentials package, and CURL
6. Download and install the NVIDIA MIG driver (grid driver).
7. Install the nvidia-docker repository (see NVIDIA Container Toolkit Installation Guide) for running nvidia-containers.
8. Configure the nvidia-grid service to use the vGPU setting on the VM (see Using GPUs with Virtual Machines on vSphere – Part 3: Installing the NVIDIA Virtual GPU Technology) and update the licenses.
9. Run the following command to verify that the setup is successful:

nvidia-smi

Note: Each VM consists of 32 vCPUs and 64 GB memory.

MLPerf Inference configuration for MIG

When the system has been configured, configure MLPerf v1.1 on the MIG VMs. To run the MLPerf Inference benchmarks on a MIG-enabled system under test, do the following:

1. Add MIG details in the inference configuration file:
   Figure 2- Example configuration for running inferences using MIG enabled VMs
   
   
2. Add valid MIG specifications to the system variable in the system_list.py file.

Figure 3- Example system entry with MIG profiles

These steps complete the system setup, which is followed by building the image, generating engines, and running the benchmark. For detailed instructions, see our previous blog about running MLPerf v1.1 on bare metal systems.

MLPerf v1.1 Benchmarking 

Benchmarking scenarios

We assessed inference latency and throughput for ResNet50 and BERT models using MLPerf Inference v1.1. The scenarios in the following table identify the number of VMs and corresponding MIG profiles used in performance tests. The total number of tests for each scenario is 57. The results are averaged based on three runs.
Note: We used MLPerf Inference v1.1 for benchmarking but the results shown in this blog are not part of the official MLPerf submissions.

Table 2: Scenarios configuration

ResNet50

ResNet50 (see Deep Residual Learning for Image Recognition) is a widely used deep convolutional neural network for various computer vision applications. This neural network can address the disappearing gradients problem by allowing gradients to traverse the network's layers using the concept of skip connections. The following figure shows an example configuration for ResNet50 inference:

Figure 4- Configuration for running inference using Resnet50 model

The following figure shows ResNet50 inference performance based on the scenarios in Table 2:

Figure 5- ResNet50 Performance throughput of MLPerf Inference v1.1 across various VMs with MIG profiles

Multiple data scientists can use all the available GPU resources while running their individual workloads on separate instances, improving overall system throughput. This result is clearly seen on Scenarios 6 through 8, which contain multiple instances, compared to Scenario 1 which consists of a single instance with the largest MIG profile for A100 80 GB. Scenario 6 achieves the highest overall system throughput (5.77 percent improvement) compared to Scenario 1. Also, Scenario 8 shows seven VMs equipped with individual GPU instances that can be built for up to seven data scientists who can fine-tune their ResNet50 base models.

BERT

BERT (see BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding) is a state-of-the-art language representational model. BERT is essentially a stack of Transformer encoders. It is suitable for neural machine translation, question answering, sentiment analysis, and text summarization, all of which require a working knowledge of the target language.

BERT is trained in two stages:

• Pretrain—During which the model acquires language and context understanding
• Fine-tuning—During which the model acquires task-specific knowledge such as querying and response.

The following figure shows an example configuration for BERT inference:

Figure 6- Configuration for running inference using BERT model

The following figure shows BERT inference performance based on scenarios in Table 2:

Figure 7- BERT Performance throughput of MLPerf Inference v1.1 across various VMs with MIG profiles

Like Resnet50 Inference performance, we clearly see that Scenarios 6 through 8, which contain multiple instances, perform better compared to Scenario 1. Particularly, Scenario 7 achieves the highest overall system throughput (21 percent improvement) compared to Scenario 1 while achieving 99.9 percent accuracy target. Also, Scenario 8 shows seven VMs equipped with individual GPU instances that can be built for up to seven data scientists who want to fine-tune their BERT base models.

Conclusion

In this blog, we describe how to install and configure MLPerf Inference v1.1 on Dell PowerEdge 750xa servers using a VMware-virtualized infrastructure and NVIDIA A100 GPUs. Furthermore, we examine the performance of single- and multi-MIG profiles running on the A100 GPU. If your ML workload is primarily inferencing-focused and response time is not an issue, enabling MIG on the A100 GPU can ensure complete GPU use with maximum throughput. Developers can use VMs with an independent GPU compute allocated to them. Also, in cases where the largest MIG profiles are used, performance is comparable to bare metal systems. Inference results from ResNet50 and BERT models demonstrate that overall system performance using either the whole GPU or multiple VMs with MIG instances hosted on an R750xa system with VMware ESXi and NVIDIA A100 GPUs performed well and produced valid results for MLPerf Inference v1.1. In both the cases, the average throughput and latency are equal. This result confirms that MIG provides predictable latency and throughput independent of other processes operating on the MIG instances on the GPU.

 There is a MIG limitation for GPU profiling on the VMs. Due to the shared nature of the hardware performance across all MIG devices, only one GPU profiling session can run on a VM; parallel GPU profiling sessions on a single VM are not possible.