[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20080110212519.GA29847@lazybastard.org>
Date: Thu, 10 Jan 2008 22:25:20 +0100
From: Jörn Engel <joern@...fs.org>
To: Matt Mackall <mpm@...enic.com>
Cc: Linus Torvalds <torvalds@...ux-foundation.org>,
Pekka J Enberg <penberg@...helsinki.fi>,
Christoph Lameter <clameter@....com>,
Ingo Molnar <mingo@...e.hu>, Hugh Dickins <hugh@...itas.com>,
Andi Kleen <andi@...stfloor.org>,
Peter Zijlstra <a.p.zijlstra@...llo.nl>,
Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [RFC PATCH] greatly reduce SLOB external fragmentation
On Thu, 10 January 2008 11:49:25 -0600, Matt Mackall wrote:
>
> b) grouping objects of the same -type- (not size) together should mean
> they have similar lifetimes and thereby keep fragmentation low
>
> (b) is known to be false, you just have to look at our dcache and icache
> pinning.
(b) is half-true, actually. The grouping by lifetime makes a lot of
sense. LogFS has a similar problem to slabs (only full segments are
useful, a single object can pin the segment). And when I grouped my
objects very roughly by their life expectency, the impact was *HUGE*!
In both cases, you want slabs/segments that are either close to 100%
full or close to 0% full. It matters a lot when you have to move
objects around and I would bet it matters even more when you cannot move
objects and the slab just remains pinned.
So just because the type alone is a relatively bad heuristic for life
expectency does not make the concept false. Bonwick was onto something.
He just failed in picking a good heuristic. Quite likely spreading by
type was even a bonus when slab was developed, because even such a crude
heuristic is slightly better than completely randomized lifetimes.
I've been meaning to split the dentry cache into 2-3 seperate ones for a
while and kept spending my time elsewhere. But I remain convinced that
this will make a measurable difference.
Jörn
--
Never argue with idiots - first they drag you down to their level,
then they beat you with experience.
-- unknown
--
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