Public bug reported: This PR backports the steal-governor[0] patches from mainline and enables the driver as a module for amd64 and arm64. The governor uses guest steal time to adjust the preferred CPU set without offlining CPUs or overriding task affinity.
This addresses performance degradation under vCPU overcommitment, where lock-holder preemption can contribute to the overhead: https://nvbugspro.nvidia.com/bug/6720994 The backport adapts the helper to the existing three-argument kcpustat_field() API and leaves fs/proc/uptime.c unchanged to preserve tickless idle accounting. These changes are explained in the affected commit. Testing: Earlier GB200 and Vera performance results: https://nvbugspro.nvidia.com/bug/6720994?commentNumber=5 [0] https://lore.kernel.org/all/[email protected]/ ** Affects: linux-nvidia (Ubuntu) Importance: Undecided Status: New ** Affects: linux-nvidia-bos (Ubuntu) Importance: Undecided Status: New ** Also affects: linux-nvidia (Ubuntu) Importance: Undecided Status: New -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2172891 Title: NVIDIA: Backport steal governor To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux-nvidia/+bug/2172891/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
