lkml.org 
[lkml]   [2002]   [Sep]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC] [PATCH] Include LRU in page count
Date
On Sunday 01 September 2002 00:30, William Lee Irwin III wrote:
> On Sat, Aug 31, 2002 at 11:05:02PM +0200, Daniel Phillips wrote:
> > The current patch seems satisfactory performance-wise and if it's
> > also raceless as it's supposed to be, it gives us something that works,
> > and we can evaluate alternatives at our leisure. Right now I'm afraid
> > we have something that just works most of the time.
> > I think we're getting to the point where this needs to get some heavy
> > beating up, to see what happens.
>
> It's not going to get much heavier than how I'm beating on it.
> Although it seems my box is mighty bored during 64 simultaneous
> tiobench 256's (i.e. 16384 tasks). I got bored & compiled a kernel:
>
> make -j64 bzImage 304.60s user 848.70s system 694% cpu 2:46.05 total
>
> The cpus are 95+% idle except for when I touch /proc/, where the task
> fishing around /proc/ gets stuck spinning hard in the kernel for
> anywhere from 30 minutes to several hours before killing it succeeds.
> It didn't quite finish the run, as the tty deadlock happened again. The
> VM doesn't appear to be oopsing, though I should slap on the OOM fixes.

Is that good or bad? It sounds like the VM cycle is doing what it's
supposed to. I've been running lots_of_forks for about 5 hours and
things have reached a steady state with free memory slowly racheting
down 1% then, apparently in a sudden fit of scanning, jumping to 2.5%.
Pretty normal looking behaviour, though I don't much like the way the
free list refill is so sporadic. I can't think of any reason why that is
good. It doesn't cost a lot of cpu though. There's been a little bit of
(useless) swapping, so I guess I can consider that path at least somewhat
tested.

Every now and then, instead of free jumping up to 2.5% it jumps up to
over 10%. I doubt that's my fault. Even though I'm missing an if
(--nr_pages) at the orphan collection point, there aren't enough orphans
to cause that effect - about 3 million out of 12 billion lru_cache_dels.

The orphans appear during or shortly after the episodes of free list
filling, as you would expect, since page_cache_release fails the trylock
whenever shrink_cache is active. There are lots of parallel
page_cache_releases going on, but they don't seem to create any
noticable number of orphans.

I think what's happening normally is this:

if (page_count(page) == 2 /* no, it's 3 */
put_page(page);
if (page_count(page) == 2 /* now it is */
if (PageLRU(page) && page_count(page) == 2)
__lru_cache_del(page);
as opposed to this:

if (page_count(page) == 2 /* no, it's 3 */
if (page_count(page) == 2 /* no, still 3 */
put_page(page);
put_page(page);

which requires hitting a 3 instruction window or on the same page.
It's not going to happen often.

Statm_pte_range looks like it's going to get even more confused about
what constitutes a shared page than it already is. The solution to
this is to examine the rmap.

--
Daniel
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:28    [W:0.170 / U:2.480 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site