On 27/01/19 11:17, Nir Weiner wrote: > The optimal value changes between different workloads. It depends on > outliers rate and polling sessions length. > As this value has significant effect on the dynamic halt-polling > algorithm, it should be configurable and exposed. What are the cases where you found the effect to be strongly negative? In practice we observed only three workloads: mostly idle (no polling), mostly busy (no HLT), repeatedly switching between idle and polling (always polling). The start value basically gives you a measure of how likely KVM is to "think" it's in the third workload. I'm not against the series, but perhaps it's possible to make the auto tuning better too. Paolo
