Re: [PATCH net-next] net: allocate skbs on local node
From: Pekka Enberg
Date: Tue Oct 12 2010 - 07:09:16 EST
On 10/12/10 10:58 AM, Andrew Morton wrote:
On Tue, 12 Oct 2010 09:49:53 +0200 Eric Dumazet<eric.dumazet@xxxxxxxxx> wrote:
Le mardi 12 octobre 2010 à 00:24 -0700, Andrew Morton a écrit :
I'd love to forget it, but it's faster for some things (I forget
which). Which is why it's still around.
Yes, two years ago it was true on pathological/obscure cases.
Every time I did the comparison, SLUB won.
You asked me, I did yet another test this morning, and 40% is pretty
serious, I believe.
And the ghastly thing about this is that you're forced to care about it
too because some people are, apparently, still using it.
Yes, some people (in my company) still use linux 2.6.9 32bit on HP G6/G7
machines, I know...
I am not saying we should not care, but for any serious network workload
on NUMA arches, SLUB is the best, and seeing Christoph recent work, it
might even get better.
BTW, I believe all modern distros ship SLUB, dont they ?
Dunno.
Pekka, why haven't we deleted slab yet??
To make a long story short, we still have relevant performance
regressions that need to be taken care of. The most interesting one is a
regression in netperf TCP_RR that's been reported by David Rientjes a
while back. There's bunch of SLUB cleanups queued for 2.6.37 that pave
the way for Christoph's SLUB queueing work that should hopefully fix
that particular issue for 2.6.38.
There's little point in discussing the removal of SLAB as long as there
are performance regressions for real workloads from people who are
willing to share results and test patches. I'm optimistic that we'll be
able to try removing SLAB some time next year unless something
interesting pops up...
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/