ZFS not caching on i386 with kmem_size >1GB

Hi,

About 2 months ago I reported that I found ZFS extremely
slow for
some tasks (specifically upgrading ports).  This was because
ZFS
was only using the absolute minimum cache size at all
times.

The problem is here in
/sys/contrib/opensolaris/uts/common/fs/zfs/arc.c:

static int
arc_reclaim_needed(void)
{
...
        if (kmem_used() > (kmem_size() * 4) / 5)
                return (1);
}

I'm running on i386 with kmem_size set to 1GB.  As a result,
the
multiplication overflows and the test becomes (kmem_used()
> 0).  ZFS then
always tries to shrink the cache, and never grows it above
the absolute
minimum size (about 30MB for each of c and p)

The patch I have attached fixes the problem for me, although
there is probably
a better way to avoid the overflow (without calling
kmem_size() twice).
Best of all, portupgrade is now an order of magnitude
faster!

Of course, I'm now worried that my previously rock-solid
settings will actually
trigger the kmem_map too small panics when the cache
actually fills up.

-- 
David Taylor
_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"

David Taylor wrote:
> Hi,
> 
> About 2 months ago I reported that I found ZFS
extremely slow for
> some tasks (specifically upgrading ports).  This was
because ZFS
> was only using the absolute minimum cache size at all
times.
> 
> The problem is here in
/sys/contrib/opensolaris/uts/common/fs/zfs/arc.c:
> 
> static int
> arc_reclaim_needed(void)
> {
> ...
>         if (kmem_used() > (kmem_size() * 4) / 5)
>                 return (1);
> }
> 
> I'm running on i386 with kmem_size set to 1GB.  As a
result, the
> multiplication overflows and the test becomes
(kmem_used() > 0).  ZFS then
> always tries to shrink the cache, and never grows it
above the absolute
> minimum size (about 30MB for each of c and p)
> 
> The patch I have attached fixes the problem for me,
although there is probably
> a better way to avoid the overflow (without calling
kmem_size() twice).
> Best of all, portupgrade is now an order of magnitude
faster!
> 
> Of course, I'm now worried that my previously
rock-solid settings will actually
> trigger the kmem_map too small panics when the cache
actually fills up.
> 

FYI, kmem_size > 1GB makes no sense unless you also
increase KVA_PAGES 
since the entire kernel only has 1GB of address space on
i386.

Kris
_______________________________________________
freebsd-fsfreebsd.org mailing list

http://lists.freebsd.org/mailman/listinfo/freebsd-fs
To unsubscribe, send any mail to
"freebsd-fs-unsubscribefreebsd.org"