Epic Goal

• To use ML benchmarks to assess the performance of vGPUs in VMs on RHEL, as well as in openshift virtualization, in comparison to direct GPU use.

Why is this important?

• Currently unknown/undocumented information, potentially desirable to customer(s)

Scenarios

1. Running mlperf SSD + SSDv2 training benchmarks and nvidiadl BERT benchmark on baremetal RHEL8, single GPU
2. Starting a single VM on RHEL8 w/ a single vGPU (full capacity), running same benchmarks
3. Running multiple workloads on baremetal w/ single GPU, running same benchmarks
4. Starting multiple VMs (same as step 3 amount of workloads) on RHEL8, each with a vGPU, running same benchmarks
5. Adding SNO to baremetal, running benchmarks in openshift w/ single GPU
6. Adding SNO to the single/multi VM environments, running same benchmarks in openshift
7. Using openshift virtualization to test VM/vGPU performance within openshift using SNO on baremetal

Acceptance Criteria

• Document for performance report covering all scenarios
• Potential presentation on results found
• Automation code (if needed) cleaned, documented, and checked into repo

Dependencies (internal and external)

1. ...

Previous Work (Optional):

1. …

Open questions::

1. …