Home > Workload Solutions > High Performance Computing > White Papers > Enhance Availability of Storage Services with an NFS Storage System > Performance vs. previous version
The following table summarizes the performance differences observed between the Validated Design for NFS Storage and the last version, NSS7.4-HA (metadata only compared for empty files, since there is no NSS7.4 data for 4KiB files).
Peak Performance Deltas |
DTVD NFS Storage |
NSS7.4 – HA Storage |
Seq. 1MB Writes Peak: 3.12% increase |
4,985 MB/s |
4,834 MB/s |
Seq. 1MB Reads Peak: 3.22% increase |
7,250 MB/s |
7,024 MB/s |
Random 4KB Writes Peak: 0.12% decrease |
7,281 IOps |
7,290 IOps |
Random 4KB Reads Peak: 31.32% decrease |
11,405 IOps |
16,607 IOps |
Create operations/second Peak: 63.44% increase |
88,579 Op/s |
54,197 Op/s |
Stat operations/second Peak: 11.39% increase |
581,734 Op/s |
522,231 Op/s |
Remove operations/second Peak: 84.75% increase |
87,469 Op/s |
47,345 Op/s |
To conclude, the Dell Technologies Validated Design for HPC NFS Storage will benefit from the new operating system version, since RHEL 8.3 XFS supported capacity limitation allows you to use any supported HDD drive on the PowerVault ME4084, up to the current 18TB HDDs.
The system has significant performance benefits from improvements to operating system components, faster components on the server Dell EMC PowerEdge R750, NVIDIA Mellanox ConnectX-6 adapters with higher data and messages rates. For example, when running benchmarks with a few threads (1-16) and for metadata operations like create and remove, regardless of the number of threads. However, more work is needed to understand some of the performance drops and try to mitigate or prevent them.
For future work, a better tool than IOzone is needed to characterize 4 KiB blocks random reads, to allow more control of the IO load, and the execution time, while reducing or eliminating caching effects everywhere (clients and server). A suitable candidate for local and remote execution is FIO.
In future MDtest benchmarking efforts, the goal will be modified to use 2,097,152 files total per metadata operation = 2^20 = 2Mi (two million in powers of two). Maintaining the number of files per directory fixed to 2048, allowing to increase the number of users from 1 to 1024 in powers of two, but use the number of directories per thread to keep constant the total number of files, fixed at 2Mi.