On 2009-06-07 at 14:23:53 [+0200], Michael Lotz <mmlr@xxxxxxxx> wrote: > Hi Cyan > > > I can reproduce the poor "system load" results on this R5 system > > by enabling the HLT instruction, and leaving the processors idle. > > In other words, calling HLT seems to require some time to let the > > processor recover (maybe the first interrupt is dropped/delayed?) > > Spinning all four cores in a loop, even at minimum priority, > > defeats the HLT calls and eliminates the jitter (as does disabling > > the HLT instruction in the kernel config file). > > > > I don't suppose this problem should crop up under Haiku, being so > > different at the kernel level? > > Actually, see these test results, both are from an idle Haiku after > boot on a 2.4GHz Quad Core: [...] > The twist is that the first one was all normal and in the second the > HLT instruction was simply disabled. I could imagine that our current > scheduling policy of scheduling threads on idle CPUs is responsible. > Obviously it's a good idea to schedule on a core that's currently idle, > but on the downside you have to "wake" that core from a halt state > which seems to have a certain overhead attached to it. Further > investigation would be nice. A run with SMP disabled would be interesting. If the "core wake-up" theory is right, the results should improve. CU, Ingo