On Thu, Sep 21, 2023 at 1:23 PM George Dunlap <george.dun...@cloud.com> wrote: > > On large systems with many vcpus yielding due to spinlock priority > inversion, it's not uncommon for a vcpu to yield its timeslice, only > to be immediately stolen by another pcpu looking for higher-priority > work. > > To prevent this: > > * Keep the YIELD flag until a vcpu is removed from a runqueue > > * When looking for work to steal, skip vcpus which have yielded > > NB that this does mean that sometimes a VM is inserted into an empty > runqueue; handle that case. > > Signed-off-by: George Dunlap <george.dun...@cloud.com>
Marcus, Just wanted to verify my interpretation of the testing you did of this patch several months ago: 1. On the problematic workload, mean execution time for the task under heavy load was around 12 seconds 2. With only patch 2 of this series (0004 in your tests), mean execution time under heavy load was around 5 seconds 3. With only patch 1 of this series (0003 in your tests), mean execution time under heavy load was around 3 seconds 4. With both patch 1 and patch 2 of this series (0003+0004 in your tests), mean execution time under heavy load was also around 3 seconds So both patches independently exhibit an improvement; but the combined effect is about the same as the first patch. Assuming those results are accurate, I would argue that we should take both patches. Does anyone want to argue we should only take the first one? -George
