Re: -mm seems significanty slower than mainline on kernbench

From: Peter Williams
Date: Sat Jan 28 2006 - 18:18:38 EST


Martin J. Bligh wrote:

Thanks, but I have that. What do you think those vertical bars on the graph are for? ;-) They're deviation of 5 runs. I throw away the best and worst first.



Not very good scientific practice :-)



Bollocks to scientific practice ;-) It works, it produces very stable results, has done for a few years now. Things like cron or sunspots can kick in. Yes, yes, I did stats at University ... but the real world doesn't work like that. The visuals in the graph speak for it.


Looking at the other 6 kernbench graphs, I see that it also occurs for elm3b70 but no others (including elm3b6 and elm3b67). Are there any differences between the various elm3b systems that could explain this?

Yes. They're all completely different architectures - there's a brief description at the top of the main page. elm3b67 should be ignored, nay thrown out of the window. It's an unstable POS that randomly loses processors. I've removed it from the pages.

elm3b70 is PPC64 (8 cpu)
elm3b6 is x86_64.
elm3b132 is a 4x SMP ia32 Pentium 3.
moe is 16x NUMA-Q (ia32).
gekko-lp1 is a 2x PPC blade.

Use the visuals in the graph .. it's very telling. -mm is *broken*.
It may well not be the same issue as last time though, I shouldn't
have jumped to that conclusion.



It's very hard to understand how it could be an issue on a system that doesn't have a lot of abnormally niced (i.e. non zero) tasks that are fairly active as it's now mathematically equivalent to the original in the absence of such tasks. Do these two systems have many such tasks running?

Would it be possible to get a run with the following patches backed out:

+sched-modified-nice-support-for-smp-load-balancing-fix.patch
+sched-modified-nice-support-for-smp-load-balancing-fix-fix.patch



Yup, that should prove or disprove it. It's probably something completely un-scheduler-related ;-)

M.

Looking at the latest results for 2.6.16-rc1-mm3, it appears to me that this is no longer an issue. Do you agree?

Peter
--
Peter Williams pwil3058@xxxxxxxxxxxxxx

"Learning, n. The kind of ignorance distinguishing the studious."
-- Ambrose Bierce
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/