Date:	Fri, 23 Jan 2009 04:55:03 +0100
From:	Nick Piggin <npiggin@...e.de>
To:	Hugh Dickins <hugh@...itas.com>
Cc:	Pekka Enberg <penberg@...helsinki.fi>,
	Linux Memory Management List <linux-mm@...ck.org>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Lin Ming <ming.m.lin@...el.com>,
	"Zhang, Yanmin" <yanmin_zhang@...ux.intel.com>,
	Christoph Lameter <cl@...ux-foundation.org>
Subject: Re: [patch] SLQB slab allocator

On Wed, Jan 21, 2009 at 06:10:12PM +0000, Hugh Dickins wrote:
> On Wed, 21 Jan 2009, Nick Piggin wrote:
> > 
> > Since last posted, I've cleaned up a few bits and pieces, (hopefully)
> > fixed a known bug where it wouldn't boot on memoryless nodes (I don't
> > have a system to test with), and improved performance and reduced
> > locking somewhat for node-specific and interleaved allocations.
> 
> I haven't reviewed your postings, but I did give the previous version
> of your patch a try on all my machines.  Some observations and one patch.

Great, thanks!

 
> I was initially _very_ impressed by how well it did on my venerable
> tmpfs loop swapping loads, where I'd expected next to no effect; but
> that turned out to be because on three machines I'd been using SLUB,
> without remembering how default slub_max_order got raised from 1 to 3
> in 2.6.26 (hmm, and Documentation/vm/slub.txt not updated).
> 
> That's been making SLUB behave pretty badly (e.g. elapsed time 30%
> more than SLAB) with swapping loads on most of my machines.  Though
> oddly one seems immune, and another takes four times as long: guess
> it depends on how close to thrashing, but probably more to investigate
> there.  I think my original SLUB versus SLAB comparisons were done on
> the immune one: as I remember, SLUB and SLAB were equivalent on those
> loads when SLUB came in, but even with boot option slub_max_order=1,
> SLUB is still slower than SLAB on such tests (e.g. 2% slower).
> FWIW - swapping loads are not what anybody should tune for.

Yeah, that's to be expected with higher order allocations I think. Does
your immune machine simply have fewer CPUs and thus doesn't use such
high order allocations?

 
> So in fact SLQB comes in very much like SLAB, as I think you'd expect:
> slightly ahead of it on most of the machines, but probably in the noise.
> (SLOB behaves decently: not a winner, but no catastrophic behaviour.)
> 
> What I love most about SLUB is the way you can reasonably build with
> CONFIG_SLUB_DEBUG=y, very little impact, then switch on the specific
> debugging you want with a boot option when you want it.  That was a
> great stride forward, which you've followed in SLQB: so I'd have to
> prefer SLQB to SLAB (on debuggability) and to SLUB (on high orders).

It is nice. All credit to Christoph for that (and the fine grained
sysfs code). 

 
> I do hate the name SLQB.  Despite having no experience of databases,
> I find it almost impossible to type, coming out as SQLB most times.
> Wish you'd invented a plausible vowel instead of the Q; but probably
> too late for that.

Yeah, apologies for the name :P

 
> init/Kconfig describes it as "Qeued allocator": should say "Queued".

Thanks.


> Documentation/vm/slqbinfo.c gives several compilation warnings:
> I'd rather leave it to you to fix them, maybe the unused variables
> are about to be used, or maybe there's much worse wrong with it
> than a few compilation warnings, I didn't investigate.

OK.
 

> The only bug I found (but you'll probably want to change the patch
> - which I've rediffed to today's slqb.c, but not retested).
> 
> On fake NUMA I hit kernel BUG at mm/slqb.c:1107!  claim_remote_free_list()
> is doing several things without remote_free.lock: that VM_BUG_ON is unsafe
> for one, and even if others are somehow safe today, it will be more robust
> to take the lock sooner.
 
Good catch, thanks. The BUG should be OK where it is if we only
claim the remote free list when remote_free_check is is set, but
some of the periodic reaping and teardown code calls it unconditionally.
But it's not critical so it should definitely go inside the lock.


> I moved the prefetchw(head) down to where we know it's going to be the head,
> and replaced the offending VM_BUG_ON by a later WARN_ON which you'd probably
> better remove altogether: once we got the lock, it's hardly interesting.

Right, I'll probably do that. Thanks!

> Signed-off-by: Hugh Dickins <hugh@...itas.com>
> ---
> 
>  mm/slqb.c |   17 +++++++++--------
>  1 file changed, 9 insertions(+), 8 deletions(-)
> 
> --- slqb/mm/slqb.c.orig	2009-01-21 15:23:54.000000000 +0000
> +++ slqb/mm/slqb.c	2009-01-21 15:32:44.000000000 +0000
> @@ -1115,17 +1115,12 @@ static void claim_remote_free_list(struc
>  	void **head, **tail;
>  	int nr;
>  
> -	VM_BUG_ON(!l->remote_free.list.head != !l->remote_free.list.tail);
> -
>  	if (!l->remote_free.list.nr)
>  		return;
>  
> +	spin_lock(&l->remote_free.lock);
>  	l->remote_free_check = 0;
>  	head = l->remote_free.list.head;
> -	/* Get the head hot for the likely subsequent allocation or flush */
> -	prefetchw(head);
> -
> -	spin_lock(&l->remote_free.lock);
>  	l->remote_free.list.head = NULL;
>  	tail = l->remote_free.list.tail;
>  	l->remote_free.list.tail = NULL;
> @@ -1133,9 +1128,15 @@ static void claim_remote_free_list(struc
>  	l->remote_free.list.nr = 0;
>  	spin_unlock(&l->remote_free.lock);
>  
> -	if (!l->freelist.nr)
> +	WARN_ON(!head + !tail != !nr + !nr);
> +	if (!nr)
> +		return;
> +
> +	if (!l->freelist.nr) {
> +		/* Get head hot for likely subsequent allocation or flush */
> +		prefetchw(head);
>  		l->freelist.head = head;
> -	else
> +	} else
>  		set_freepointer(s, l->freelist.tail, head);
>  	l->freelist.tail = tail;
>  
--
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