You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Add AMD CI runners with their own performance series #410
CI's microVM and platform runners are Azure VMs with Intel Xeon CPUs (Ice Lake-SP or Emerald Rapids). So no CI job exercises an AMD CPU profile or the AMD paths of the backends, and the performance gate covers only Intel hosts. #396 deferred AMD runners (plan item 7), and doc/benchmarks.md says how one would get its own series.
The built-in AMD profiles now serve the CPUs of Azure's AMD VMs: Milan, Genoa, and Turin. #412 qualified all three on KVM in GitHub-hosted Actions runners, which are such VMs, with the full default test-microvm suite. GitHub-hosted runners can't hold a performance series, though: each job lands on any of those CPUs, or on an Intel one.
Plan
Provision self-hosted AMD runners for the backends that CI covers, one AMD VM type each, so that each series measures one CPU:
KVM and MSHV on Azure Linux VMs with nested virtualization;
WHP on Windows VMs.
Route the microVM test and platform jobs to them, and give them their own performance series, with an -amd suffix such as linux-kvm-virtual-machine-amd, as doc/benchmarks.md describes.
Record their baselines from dev runs before the performance gate compares them.
Activity
openvmm: serve AMD EPYC Genoa and Turin hosts
openvmm: serve AMD EPYC Genoa and Turin hosts