Re: [Bug #15713] hackbench regression due to commit 9dfc6e68bfe6e

From: Pekka Enberg
Date: Mon Apr 26 2010 - 03:22:24 EST


Hi Yanmin,

On Mon, Apr 26, 2010 at 9:59 AM, Zhang, Yanmin
<yanmin_zhang@xxxxxxxxxxxxxxx> wrote:
>> I haven't been able to reproduce this either on my Core 2 machine.
> Mostly, the regression exists on Nehalem machines. I suspect it's related to
> hyper-threading machine.

OK, so does anyone know why hyper-threading would change things for
the per-CPU allocator?

>> Yanmin, does something like this help on your machines?
> A quick testing doesn't show any help.

So it's unlikely to be false sharing, I suppose.

> I did a new testing. After the machine boots, I hot remove 8 hyper-threading cpu
> which means last 8 are just cores. The regression between 2.6.33 and 2.6.34-rc becomes
> small.
>
> My opinion is we needn't revert the patch, but still keep an eye on it when testing other
> new RC kernel releases. One reason is volanoMark and netperf have no such regression.
> Is it ok?

We need to get this fixed. In my experience, it's pretty common that
slab regressions pop up only in one or few benchmarks. The problem is
likely to pop up in some real-world workload where it's even more
difficult to track down because basic CPU profiles don't pin-point the
problem.

Do we have some Intel CPU expert hanging around here that could
enlighten me of the effects of hyper-threading on CPU caching? I also
wonder why it's showing up with the new per-CPU allocator and not with
the homebrewn one we had in SLUB previously.

Pekka
--
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/