Step #4: Analyze the Results
Our results.txt file contains the results of both our baseline test and our UPT test. Let’s take a look.
Standard vSwitch (no UPT) |
NSX vSwitch on DPU (UPT enabled) |
---|---|
Average latency = .245 ms |
Average latency = .179 ms |
Operations/sec = 635k |
Operations/sec = 847k |
Throughput = 74 Mbps |
Throughput = 99 Mbps |
Your numbers will not exactly match these results, but should reflect similar performance improvements. You can execute these tests multiple times.
As you can see from our test results, UPT performs much better than our baseline test in latency, operations/sec, and throughput. This illustrates that UPT-enabled VMs, backed by Nvidia Bluefield DPUs and VMware NSX, not only perform better, but also get the benefits of added functionality. In this case, we’re adding Geneve overlay overhead to the UPT test case, and even with this added functionality it still performs better than our baseline test.
This testing was done with a single redis server and benchmark client. Even at this small scale, we begin to see the benefits of UPT. In a real operational environment, many more redis server instances and threads would be servicing read/write requests. The performance benefits of DPU offload and acceleration will scale up with the size of customer deployments.
As users adopt vSphere 8 and NSX 4 with Nvidia Bluefield DPUs, they will benefit from even more offloaded features than we’ve experimented with in this lab. Many of NSX’s security services can also be offloaded and accelerated on Nvidia DPUs.