[PATCH -mmotm 0/5] memcg: per cgroup dirty limit (v6)

[Andrea Righi]

On Thu, Mar 11, 2010 at 10:03:07AM -0500, Vivek Goyal wrote:
> On Thu, Mar 11, 2010 at 06:25:00PM +0900, KAMEZAWA Hiroyuki wrote:
> > On Thu, 11 Mar 2010 10:14:25 +0100
> > Peter Zijlstra <peterz at infradead.org> wrote:
> > 
> > > On Thu, 2010-03-11 at 10:17 +0900, KAMEZAWA Hiroyuki wrote:
> > > > On Thu, 11 Mar 2010 09:39:13 +0900
> > > > KAMEZAWA Hiroyuki <kamezawa.hiroyu at jp.fujitsu.com> wrote:
> > > > > > The performance overhead is not so huge in both solutions, but the impact on
> > > > > > performance is even more reduced using a complicated solution...
> > > > > > 
> > > > > > Maybe we can go ahead with the simplest implementation for now and start to
> > > > > > think to an alternative implementation of the page_cgroup locking and
> > > > > > charge/uncharge of pages.
> > > 
> > > FWIW bit spinlocks suck massive.
> > > 
> > > > > 
> > > > > maybe. But in this 2 years, one of our biggest concerns was the performance.
> > > > > So, we do something complex in memcg. But complex-locking is , yes, complex.
> > > > > Hmm..I don't want to bet we can fix locking scheme without something complex.
> > > > > 
> > > > But overall patch set seems good (to me.) And dirty_ratio and dirty_background_ratio
> > > > will give us much benefit (of performance) than we lose by small overheads.
> > > 
> > > Well, the !cgroup or root case should really have no performance impact.
> > > 
> > > > IIUC, this series affects trgger for background-write-out.
> > > 
> > > Not sure though, while this does the accounting the actual writeout is
> > > still !cgroup aware and can definately impact performance negatively by
> > > shrinking too much.
> > > 
> > 
> > Ah, okay, your point is !cgroup (ROOT cgroup case.)
> > I don't think accounting these file cache status against root cgroup is necessary.
> > 
> 
> I think what peter meant was that with memory cgroups created we will do
> writeouts much more aggressively.
> 
> In balance_dirty_pages()
> 
> 	if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
> 		break;
> 
> Now with Andrea's patches, we are calculating bdi_thres per memory cgroup
> (almost)
> 
> bdi_thres ~= per_memory_cgroup_dirty * bdi_fraction
> 
> But bdi_nr_reclaimable and bdi_nr_writeback stats are still global.

Correct. More exactly:

 bdi_thresh = memcg dirty memory limit * BDI's share of the global dirty memory

Before:

 bdi_thresh = global dirty memory limit * BDI's share of the global dirty memory

> 
> So for the same number of dirty pages system wide on this bdi, we will be
> triggering writeouts much more aggressively if somebody has created few
> memory cgroups and tasks are running in those cgroups.

Right, if we don't touch per-cgroup dirty limits.

> 
> I guess it might cause performance regressions in case of small file
> writeouts because previously one could have written the file to cache and
> be done with it but with this patch set, there are higher changes that
> you will be throttled to write the pages back to disk.
> 
> I guess we need two pieces to resolve this.
> 	- BDI stats per cgroup.
> 	- Writeback of inodes from same cgroup.
> 
> I think BDI stats per cgroup will increase the complextiy.

There'll be the opposite problem I think, the number of dirty pages
(system-wide) will increase, because in this way we'll consider BDI
shares of memcg dirty memory. So I think we need both: per memcg BDI
stats and system-wide BDI stats, then we need to take the min of the two
when evaluating bdi_thresh. Maybe... I'm not really sure about this, and
need to figure better this part. So I started with the simplest
implementation: global BDI stats, and per-memcg dirty memory.

I totally agree about the other point, writeback of inodes per cgroup is
another feature that we need.

> I am still setting up the system to test whether we see any speedup in
> writeout of large files with-in a memory cgroup with small memory limits.
> I am assuming that we are expecting a speedup because we will start
> writeouts early and background writeouts probably are faster than direct
> reclaim?

mmh... speedup? I think with a large file write + reduced dirty limits
you'll get a more uniform write-out (more frequent small writes),
respect to few and less frequent large writes. The system will be more
reactive, but I don't think you'll be able to see a speedup in the large
write itself.

Thanks,
-Andrea