Still getting kmem exhausted panic

[Ben Kelly]

On Sep 28, 2010, at 5:30 PM, Andriy Gapon wrote:

<< snipped lots of good info here... probably won't have time to look at it in detail until the weekend >>

>> there seems to be a layering violation in that the buffer cache signals
>> directly to the upper page daemon layer to trigger page reclamation.)
> 
> Umm, not sure if that is a fact.

I was referring to the code in vfs_bio.c that used to twiddle vm_pageout_deficit directly.  That seems to have been replaced with a call to vm_page_grab().

>> The old (ancient) patch I tried previously to help reduce the arc working set
>> and allow it to shrink is here:
>> 
>> http://www.wanderview.com/svn/public/misc/zfs/zfs_kmem_limit.diff
>> 
>> Unfortunately, there are a couple ideas on fighting fragmentation mixed into
>> that patch.  See the part about arc_reclaim_pages().  This patch did seem to
>> allow my arc to stay under the target maximum even when under load that
>> previously caused the system to exceed the maximum.  When I update this
>> weekend I'll try a stripped down version of the patch to see if it helps or
>> not with the latest zfs.
>> 
>> Thanks for your help in understanding this stuff!
> 
> The patch seems good, especially the part about taking into account the kmem
> fragmentation.  But it also seems to be heavily tuned towards "tiny ARC" systems
> like yours, so I am not sure yet how suitable it is for "mainstream" systems.

Thanks.  Yea, there is a lot of aggressive tuning there.  In particular, the slow growth algorithm is somewhat dubious.  What I found, though, was that the fragmentation jumped whenever the arc was reduced in size, so it was an attempt to make the size slowly approach peak load without overshooting.

A better long term solution would probably be to enhance UMA to support custom slab sizes on a zone-by-zone basis.  That way all zfs/arc allocations can use slabs of 128k (at a memory efficiency penalty of course).  I prototyped this with a dumbed down block pool allocator at one point and was able to avoid most, if not all, of the fragmentation.  Adding the support to UMA seemed non-trivial, though.

Thanks again for the information.  I hope to get a chance to look at the code this weekend.

- Ben