At STH, we have been covering DPUs for many years. Enough time to see some solutions flourish and others fail. While DPUs have been deployed in the cloud for many years, bringing that cloud-like deployment model to the Enterprise has been much slower, if not impossible. Now with VMware vSphere 8, organizations with VMware deployments can start taking advantage of DPUs. That project has gone from theory to deployment and we were able to check it out.
As a quick note, we normally would not be prioritized to get access for this NVIDIA LaunchPad demo. We are saying that NVIDIA is sponsoring this piece. Like all pieces on STH, NVIDIA, and VMware did not get to review this piece before it went live to maintain editorial independence. NVIDIA did help show us around the demo environment before we started our walk-through and testing. That was necessary in this case to ensure our first time setting this up was not going to be done incorrectly.
What is the NVIDIA BlueField-2 DPU?
We have covered the NVIDIA BlueField-2 DPU on STH many times. One way to think about BlueField-2 is that it is a NVIDIA ConnectX-6 network interface with an Arm core complex, its own memory, storage, out-of-band networking, a PCIe switch, and accelerators.
The BlueField-2 cards come in various form factors. Here is one from our recent AMD Ryzen Server the ASRock Rack 1U4LW-B650/2L2T Review.
Here is the topology when logged directly into the card’s Arm complex.
We showed in ZFS without a Server Using the NVIDIA BlueField-2 DPU a fun demo with these cards running storage stacks using the Arm cores.
While the demo we built last year was fun, it was not meant for production. The VMware-NVIDIA networking integration is. That is the solution we are going to review today.
Next, we are going to show how this works and how we set up the demo environment. If you just want to see the “so what” performance and impact, that will be on the last page.