* Rusty Russell <ru...@rustcorp.com.au> wrote:
> Ingo Molnar <mi...@kernel.org> writes:
> > * Madhavan Srinivasan <ma...@linux.vnet.ibm.com> wrote:
> >
> >> Performance data for different FAULT_AROUND_ORDER values from 4 socket
> >> Power7 system (128 Threads and 128GB memory). perf stat with repeat of 5
> >> is used to get the stddev values. Test ran in v3.14 kernel (Baseline) and
> >> v3.15-rc1 for different fault around order values.
> >>
> >> FAULT_AROUND_ORDER Baseline 1 3 4
> >> 5 8
> >>
> >> Linux build (make -j64)
> >> minor-faults 47,437,359 35,279,286 25,425,347
> >> 23,461,275 22,002,189 21,435,836
> >> times in seconds 347.302528420 344.061588460 340.974022391
> >> 348.193508116 348.673900158 350.986543618
> >> stddev for time ( +- 1.50% ) ( +- 0.73% ) ( +- 1.13% ) (
> >> +- 1.01% ) ( +- 1.89% ) ( +- 1.55% )
> >> %chg time to baseline -0.9% -1.8%
> >> 0.2% 0.39% 1.06%
> >
> > Probably too noisy.
>
> A little, but 3 still looks like the winner.
>
> >> Linux rebuild (make -j64)
> >> minor-faults 941,552 718,319 486,625
> >> 440,124 410,510 397,416
> >> times in seconds 30.569834718 31.219637539 31.319370649
> >> 31.434285472 31.972367174 31.443043580
> >> stddev for time ( +- 1.07% ) ( +- 0.13% ) ( +- 0.43% ) (
> >> +- 0.18% ) ( +- 0.95% ) ( +- 0.58% )
> >> %chg time to baseline 2.1% 2.4%
> >> 2.8% 4.58% 2.85%
> >
> > Here it looks like a speedup. Optimal value: 5+.
>
> No, lower time is better. Baseline (no faultaround) wins.
>
>
> etc.
ah, yeah, you are right. Brainfart of the week...
> It's not a huge surprise that a 64k page arch wants a smaller value
> than a 4k system. But I agree: I don't see much upside for FAO > 0,
> but I do see downside.
>
> Most extreme results:
> Order 1: 2% loss on recompile. 10% win 4% loss on seq. 9% loss random.
> Order 3: 2% loss on recompile. 6% win 5% loss on seq. 14% loss on random.
> Order 4: 2.8% loss on recompile. 10% win 7% loss on seq. 9% loss on random.
>
> > I'm starting to suspect that maybe workloads ought to be given a
> > choice in this matter, via madvise() or such.
>
> I really don't think they'll be able to use it; it'll change far too
> much with machine and kernel updates. [...]
Do we know that?
> [...] I think we should apply patch
> #1 (with fixes) to make it a variable, then set it to 0 for PPC.
Ok, agreed - at least until contrary data comes around.
Thanks,
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
