On 12/04/08 19:21, Li, Aubrey wrote: > Bill Holler wrote: > > >> Deep cstate avoidance based on the number of non-idle cpus in the >> system does not play well with powertop! Odd. >> May be the dtrace probes with powertop causes all cpus to be busy at >> the same times? >> >> >> When powertop is not running, non-idle cpu count throttling of >> deep c-states does not effect idle systems. >> >> > > How is a cpu considered to be non-idle? > The algorithm that is working well counts the number of times the cpu has gone idle over a time span of 10 milliseconds.
> we need to figure out the workload is light or heavy on this cpu. > The algorithm that did not work well computed the number of bits set in the cpu partition's mc_haltset cpuset_t. That is the number of idle cpus in the partition. There is usually one cpu partition for the whole system. This metric was surprisingly useless. It might help with cpu-bound workloads. > Consider my patch for 5444, it makes powertop happy, :-) > Can you run your patch through SPECPower? I will have a patch tonight. Hopefully you can also run SPECpower on the new patch. Bill > -Aubrey >
