[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <1190060393.5460.143.camel@localhost>
Date: Mon, 17 Sep 2007 16:19:53 -0400
From: Lee Schermerhorn <Lee.Schermerhorn@...com>
To: Christoph Lameter <clameter@....com>
Cc: linux-kernel <linux-kernel@...r.kernel.org>,
akpm@...ux-foundation.org, ak@...e.de, eric.whitney@...com,
Mel Gorman <mel@....ul.ie>
Subject: Re: [PATCH] 2.6.23-rc6: Fix NUMA Memory Policy Reference Counting
On Mon, 2007-09-17 at 12:37 -0700, Christoph Lameter wrote:
> On Mon, 17 Sep 2007, Lee Schermerhorn wrote:
>
> > Here is the 23-rc6 verison of the patch. Andi considers it a high
> > priority bug fix for .23. I'm a bit uncomfortable with this, this late
> > in the 23 cycle. I've not heard of problems w/o this patch, but then,
> > maybe no one notices if they leak a memory policy struct now and then,
> > or occasionally allocate memory on the wrong node because they used a
> > prematurely freed memory policy.
>
> The patch does require concurrent increments and decrements in the main
> fault patch. The potential is to create another bouncing cacheline for
> concurrent faults. This looks like it would cause a performance issue.
Only for vma policy, right? show_numa_maps() isn't a performance path,
and shared policies are already reference counted--just not unref'd!
>
> > Kernel Build [16cpu, 32GB, ia64] - average of 10 runs:
> >
> > w/o patch w/ refcount patch
> > Avg Std Devn Avg Std Devn
> > Real: 100.59 0.38 100.63 0.43
> > User: 1209.60 0.37 1209.91 0.31
> > System: 81.52 0.42 81.64 0.34
>
> Single threaded build? I would suggest to try concurrently faulting memory
> from multiple processors. You may not see this on a kernel build even if
> this is run with -j16 because concurrent faults are rare.
Well, it was a 32-way parallel build [-j32] on a 16-cpu system--my usual
build method. But, I'm guessing that all of the build tools are single
threaded and all using default policy, so no reference counting is
needed.
I'm taking a look at your 'pft' program, and I'll try that.
I do have some ideas for enhancements to memtoy to test vma policies in
a multi-threaded task. I have the basic multi-threading infrastructure
that binds threads to cpus, allocates node local stacks, thread state
structs, ... in my mmtrace tool that I can probably hack for use in
memtoy to provoke cacheline bouncing of the mem policy. But, if pft
does the trick, I won't rush the memtoy enhancments...
Meanwhile, we do have a mem policy ref counting bug in the mainline.
Later,
Lee
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists