Messages in this thread | | | Date | Fri, 11 Apr 2008 10:06:17 +0200 | From | Andi Kleen <> | Subject | Re: [PATCH] x86: set_cyc2ns_scale() remove tsc_now and ns_now |
| |
On Fri, Apr 11, 2008 at 09:55:54AM +0200, Ingo Molnar wrote: > > * Andi Kleen <andi@firstfloor.org> wrote: > > > (actually it is still the wrong time -- really needs a grace period > > during which the TSC is not used > > ftp://firstfloor.org/pub/ak/quilt/patches/sched-clock implemented some > > of these ideas against an older kernel) > > recent CPUs have constant-freq TSCs so it's mostly a legacy issue, but
Actually there millions of non constant freq TSC CPUs shipped each quarter ...
> we dont really have to worry about complications like grace periods - > higher layers in the scheduler protect against temporary sched_clock() > outliers.
But you still get scheduling hickups even with the sanity check. If the scheduler depends on a smooth time that is not good and my (admittedly much less than yours) understanding of CFS is that it relies on that. Especially ondemand can cause quite a lot of cpufreq changes on some workloads.
> So i think this can all be done much simpler. Just get rid of > the global cpu_khz notion, sched_clock() should simply follow the ->freq > value - and that's it.
At some point you have to generate an offset to something and that offset must be different for different frequencies, otherwise you get large systematic errors
(<imagine complicated mathematical proof why this is so, but it should be fairly obvious>)
-Andi
| |