Re: [PATCH v3 3/3] vfs: Use per-cpu list for superblock's inode list

[Jan Kara]

On Wed 24-02-16 09:36:30, Ingo Molnar wrote:
> 
> * Jan Kara <jack@xxxxxxx> wrote:
> 
> > On Tue 23-02-16 14:04:32, Waiman Long wrote:
> > > When many threads are trying to add or delete inode to or from
> > > a superblock's s_inodes list, spinlock contention on the list can
> > > become a performance bottleneck.
> > > 
> > > This patch changes the s_inodes field to become a per-cpu list with
> > > per-cpu spinlocks. As a result, the following superblock inode list
> > > (sb->s_inodes) iteration functions in vfs are also being modified:
> > > 
> > >  1. iterate_bdevs()
> > >  2. drop_pagecache_sb()
> > >  3. wait_sb_inodes()
> > >  4. evict_inodes()
> > >  5. invalidate_inodes()
> > >  6. fsnotify_unmount_inodes()
> > >  7. add_dquot_ref()
> > >  8. remove_dquot_ref()
> > > 
> > > With an exit microbenchmark that creates a large number of threads,
> > > attachs many inodes to them and then exits. The runtimes of that
> > > microbenchmark with 1000 threads before and after the patch on a
> > > 4-socket Intel E7-4820 v3 system (40 cores, 80 threads) were as
> > > follows:
> > > 
> > >   Kernel            Elapsed Time    System Time
> > >   ------            ------------    -----------
> > >   Vanilla 4.5-rc4      65.29s         82m14s
> > >   Patched 4.5-rc4      22.81s         23m03s
> > > 
> > > Before the patch, spinlock contention at the inode_sb_list_add()
> > > function at the startup phase and the inode_sb_list_del() function at
> > > the exit phase were about 79% and 93% of total CPU time respectively
> > > (as measured by perf). After the patch, the percpu_list_add()
> > > function consumed only about 0.04% of CPU time at startup phase. The
> > > percpu_list_del() function consumed about 0.4% of CPU time at exit
> > > phase. There were still some spinlock contention, but they happened
> > > elsewhere.
> > 
> > While looking through this patch, I have noticed that the
> > list_for_each_entry_safe() iterations in evict_inodes() and
> > invalidate_inodes() are actually unnecessary. So if you first apply the
> > attached patch, you don't have to implement safe iteration variants at all.
> > 
> > As a second comment, I'd note that this patch grows struct inode by 1 pointer. 
> > It is probably acceptable for large machines given the speedup but it should be 
> > noted in the changelog. Furthermore for UP or even small SMP systems this is 
> > IMHO undesired bloat since the speedup won't be noticeable.
> > 
> > So for these small systems it would be good if per-cpu list magic would just 
> > fall back to single linked list with a spinlock. Do you think that is reasonably 
> > doable?
> 
> Even many 'small' systems tend to be SMP these days.

Yes, I know. But my tablet with 4 ARM cores is unlikely to benefit from
this change either. It will just have to pay the memory cost. And frankly I
don't care that much myself but if there is some reasonably easy way to
avoid the cost, it would be welcome.

> If you do this then please keep it a separate add-on patch, so that the
> 'UP cost' becomes apparent. Uniprocessor #ifdeffery is really painful in
> places and we might be better off with a single extra pointer.
> Forthermore UP kernels are tested a lot less stringently than SMP
> kernels. It's just 4 bytes for a truly small 32-bit system.

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR