Hi Nick,
On Tue, Nov 22, 2005 at 03:48:18PM +0800, Wu Fengguang wrote:
> Hi, your vm-kswapd-incmin.patch looks nice, and I'd like to base my age
> balancing patch upon it. But while trying it, I ran into a problem.
>
> $ cp bigfile /dev/null
> $ free -m
> total used free shared buffers cached
> Mem: 501 495 6 0 2 321
> -/+ buffers/cache: 171 330
> Swap: 127 2 125
> $ sleep 10
> $ free -m
> total used free shared buffers cached
> Mem: 501 244 257 0 4 66
> -/+ buffers/cache: 173 328
> Swap: 127 2 125
>
> In a short time, the bigfile was totally evicted from page cache?
>
> Before/after the huge free pages:
>
> $ cat /proc/sys/fs/dentry-state
> 3393 28 45 0 0 0
> $ cat /proc/sys/fs/dentry-state
> 3626 260 45 0 0 0
>
>
> linux-2.6.15-rc1-mm2 without the patch seems ok.
> Any suggestions? Thanks.
Maybe I found the answer to it :)
I listed the dentry-state in several other normal machines, they all have
more than 10k unused dentries:
54215 42155 45 0 0 0
91704 81376 45 0 0 0
96304 88832 45 0 0 0
30057 26999 45 0 0 0
Then I disabled the shrinker by:
echo 0 > /proc/sys/vm/vfs_cache_pressure
That increased the number from
3393 28 45 0 0 0
to
6247 2672 45 0 0 0
And there is no sudden huge increases of free pages any more.
Maybe your patch is shrinking the slabs much more, though I cannot confirm this
from the source code. But one thing I'm sure is that there should be a lower
bound for the unused dentries, either absolutely or relatively, something like
this:
--- linux-2.6.15-rc1-mm2.orig/fs/dcache.c
+++ linux-2.6.15-rc1-mm2/fs/dcache.c
@@ -860,7 +860,7 @@ static int shrink_dcache_memory(int nr,
return -1;
prune_dcache(nr);
}
- return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
+ return (dentry_stat.nr_unused / 1000) * 10 * sysctl_vfs_cache_pressure;
}
/**
The original 100 is way too small. It should be much more than the SHRINK_BATCH(128).
Thanks,
Wu
Wu Fengguang wrote:
> Hi Nick,
>
Hi Wu,
> Then I disabled the shrinker by:
> echo 0 > /proc/sys/vm/vfs_cache_pressure
>
> That increased the number from
> 3393 28 45 0 0 0
> to
> 6247 2672 45 0 0 0
> And there is no sudden huge increases of free pages any more.
>
Yes, the inode shrinker can discard all pagecache from a file if
it is under a lot of pressure to free inodes. This is what you
are seeing I guess.
> Maybe your patch is shrinking the slabs much more, though I cannot confirm this
> from the source code. But one thing I'm sure is that there should be a lower
> bound for the unused dentries, either absolutely or relatively, something like
> this:
>
> --- linux-2.6.15-rc1-mm2.orig/fs/dcache.c
> +++ linux-2.6.15-rc1-mm2/fs/dcache.c
> @@ -860,7 +860,7 @@ static int shrink_dcache_memory(int nr,
> return -1;
> prune_dcache(nr);
> }
> - return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
> + return (dentry_stat.nr_unused / 1000) * 10 * sysctl_vfs_cache_pressure;
> }
>
I don't think that kswapd-incmin puts much more pressure on the slab
(unless there is a bug), but I'll take a look. It could just be a
"weird" interaction in the reclaim code, or possibly a rounding issue.
Changing the pressure calculation here is probably not the right way
to do this, but rather in vmscan.c. I know Andrea has recently been
looking at a problem with slab shrinking too.
Thanks,
Nick
--
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com