Re: [PATCH] xfs: account extra freespace btree splits for multiple allocations

[Dave Chinner]

On Tue, Nov 22, 2022 at 06:03:03PM +0800, Oliver Sang wrote:
> hi Gao Xiang,
> 
> On Tue, Nov 22, 2022 at 09:33:38AM +0800, Gao Xiang wrote:
> > On Tue, Nov 22, 2022 at 09:09:34AM +0800, kernel test robot wrote:
> > > 
> > > please be noted we noticed Gao Xiang and Dave Chinner have already had lots of
> > > discussion around this patch, which seems there is maybe new version later.
> > > we just sent out this report FYI the possible performance impact of this patch.
> > > 
> > > 
> > > Greeting,
> > > 
> > > FYI, we noticed a -15.1% regression of fxmark.ssd_xfs_MWCM_72_directio.works/sec due to commit:
> > 
> > Thanks for your report!
> > 
> > At a glance, I have no idea why this commit can have performance
> > impacts.  Is the result stable?
> 
> in our tests, the result is quite stable.
>      45589           -15.1%      38687 ±  2%  fxmark.ssd_xfs_MWCM_72_directio.works/sec
> 
> and detail data is as below:
> for this commit:
>   "fxmark.ssd_xfs_MWCM_72_directio.works/sec": [
>     39192.224368,
>     39665.690567,
>     38980.680601,
>     37298.99538,
>     37483.256377,
>     39504.606569
>   ],
> 
> for parent:
>   "fxmark.ssd_xfs_MWCM_72_directio.works/sec": [
>     45381.458009,
>     45314.376204,
>     45724.688965,
>     45751.955937,
>     45614.323267,
>     45747.216475
>   ],

This MWCM workload uses a shared directory. Every worker thread (72
of them) iterates creating a new file, writes 4kB of data to it and
then closes it. There is no synchronisation between worker threads.

The worker threads will lockstep on the directory lock for file
creation, they will all attempt to allocate data in the same AG as
the file is created. Hence writeback will race with file creation
for AG locks, too.  Once the first AG is full, they will all attempt
to allocate in the next AG (file creation and writeback).

IOWs, this workload will race to fill AGs, will exercise the "AG
full so skip to next AG" allocator fallbacks, etc.

Changing where/how AGs are considered full will impact how the AG
selection is made. I'm betting that there's a mismatch between the
code that selects the initial AG for allocation (from
xfs_bmap_btalloc() via the nullfb case) and the code that selects
the actual AG for allocation (xfs_alloc_vextent() w/ NEAR_BNO
policy) as a result of this change. This then results in
xfs_alloc_vextent() trying to initially allocate from an AG that
xfs_alloc_fix_freelist() considers to be full, so it skips the
initial selected AG and starts searching for an AG it can allocate
into.

Combine that with AGF lock contention from 70+ tasks all trying to
allocate in the same location...

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx