AMD EPYC Zen3 Delivers 20% More SQL Server Performance
Mon, 03 May 2021 14:04:47 -0000
|Read Time: 0 minutes
It’s a common question: “How much database performance will we gain when upgrading to the newest server technology?” The person asking the question wants to justify the investment based on measured benefit. An engineering team here at Dell Technologies ran a load test comparing the prior generation of AMD EPYC processors to the new EPYC Zen3 processors. These test findings show a double-digit gain in performance for a typical write-heavy transactional workload using the new AMD EPYC Zen3 processors.
We used two Dell EMC PowerEdge R7525 servers with different generations of AMD EPYC processors. One server was configured with two 32-core AMD EPYC 75F3 processors running with a base clock speed of 2.95 GHz that can boost up to 4 GHz.
The table below compares the AMD EPYC Zen3 processors to the Zen2 processors used in the other server. We see that the new Zen3 processors have a higher boost clock speed and double the L3 cache. These new processors should accelerate database workloads by providing greater performance.
Component | PowerEdge R7525 with AMD EPYC Zen3 | PowerEdge R7525 with AMD EPYC Zen2 |
AMD EPYC CPU | 75F3 | 7542 |
Base clock speed | 2.95 GHz | 2.9 GHz |
Boost clock speed | 4.0 GHz | 3.4 GHz |
L1 cache | 96K per core | 96K per core |
L2 cache | 512K per core | 512K per core |
L3 cache | 256 MB shared | 128 MB shared |
Both generations of processors support 8 memory channels, with each memory channel supporting up to 2 DIMMS. We believe the faster boost clock speed combined with doubling the L3 cache for the new generation AMD EPYC Zen3 processors will drive greater database performance; however, there are many more features that we haven’t covered. This AMD webpage covers the EPYC 75F3 processors. For a deep technical dive into performance tuning database workloads we recommend this RDBMS tuning guide.
The table below shows a comparison summary of two PowerEdge servers used in the testing.
Component | PowerEdge Milan (Zen3) | PowerEdge ROME (Zen2) |
Processor | 2 x AMD EPYC 75F3 32 core processor | 2 x AMD EPYC 7542 32 core processor |
Memory | 2,048 GB 3.2 GHz | 2,048 GB 3.2 GHz |
Disk Storage | 8 x Dell Express Flash NVMe P4610 1.6 TB | 8 x Dell Express Flash NVMe P4610 1.6 TB |
Embedded NIC | 1 x Broadcom Gigabit Ethernet BCM5720 | 1 x Broadcom Gigabit Ethernet BCM5720 |
Integrated NIC | 1 x Broadcom Adv. Dual port 25 GB Ethernet | 1 x Broadcom Adv. Dual port 25 GB Ethernet |
Microsoft SQL Server Enterprise Edition 2019 was virtualized to reflect the most common configuration we see at customer sites. The performance differences between bare-metal and VMware virtualization are rarely a consideration for customers since AMD and VMware vSphere 7.0 CU2 are continuously optimizing performance. The paper, “Performance Optimizations in VMware vSphere 7.0 U2 CPU Scheduler for AMD EPYC Processors” shows that the VMware CPU scheduler archives up to 50% better performance than 7.0 U1. Virtualized data management systems have also enabled Database-as-a-Service offerings on-premise for many enterprises. The capability to quickly provision database copies can significantly benefit many IT priorities and programs.
The Dell Engineering team used VMware virtualization to create a SQL Server virtual machine template for this testing. That template allowed the team to quickly provision four copies of the exact same virtualized database across the two PowerEdge servers. We hosted two virtualized SQL Server databases on each PowerEdge server. The figure below shows the infrastructure configuration for this performance test.
The virtual machine configuration for the four SQL Servers databases and the memory allocations for SQL Server are detailed below.
Component | Virtual Machine Configuration and Memory of SQL Server |
vCPU | 28 |
Memory | 886 GB |
Disk Storage | 2.7 TB |
Memory for SQL Server | 758 GB |
Both databases used VMware vSphere’s Virtual Machine File System (VMFS) on Direct Attached Storage (DAS). In the table below shows the sizes of each storage volume used for the virtual machine. There are two data volumes (data 1 and data 2) to increase I/O bandwidth to the disk storage. By having two volumes, reads and writes are split between the volumes increasing storage performance.
Storage Group | Size (GB) |
Operating System | 100 |
Data 1 | 600 |
Data 2 | 600 |
TempDB and TempDB Log | 200 |
Lob | 200 |
Backup | 1,000 |
In summary, the entire SQL Server software stack consisted of:
- Microsoft SQL Server Enterprise Edition 2019 CU9
- Red Hat Enterprise Linux 8.3
- VMware vSphere ESXi 7.0 Update 2
To create an Online Transaction Processing (OLTP) workload on the two SQL Server databases, the team used HammerDB. HammerDB is a leading benchmark tool used with databases like Microsoft SQL Server, Oracle, and others. We used HammerDB to generate a TPC-C workload that simulates terminal operators executing transactions, which is characteristic of an OLTP workload. A typical OLTP workload sends thousands of small read and write requests per minute to that database that must get committed to storage. New Orders Per Minute (NOPM) indicates the number of orders that were fully processed in one minute and is a metric we can use to compare two database systems. Below are the HammerDB settings used for the TPC-C workload test.
Setting | Value |
Time Driver Script | Yes |
Total Transactions per user | 1,000,000 |
Minutes of Ramp Up Time | 10 |
Minutes of Test Duration | 20 |
Use All Warehouses | Yes |
Number of Virtual Users | 100 |
The new AMD EPYC Zen3 processors delivered a 20% increase in New Orders Per Minute performance over the prior generation processor! That is a substantial improvement across the two virtualized SQL Server databases running on the PowerEdge R7525 server. See the comparison chart below.
It’s very likely that the boost memory speed of 4.0 GHz and the larger L3 cache in the AMD EPYC Zen3 processors contributed significantly to the 20% performance gain. Every database is different, and results will vary. However, this performance test provides value in terms of understanding the potential gains in moving to AMDs new Zen3 processors. For enterprises considering migrating their databases to a new server platform, serious consideration should be given to using the PowerEdge R7525 servers with the new generation of AMD processors.