Re: [PATCH 2/2] memcg: dirty pages instrumentation

[Vivek Goyal <vgoyal@redhat.com>]

On Fri, Feb 26, 2010 at 11:21:21PM +0100, Andrea Righi wrote:
> On Fri, Feb 26, 2010 at 04:48:11PM -0500, Vivek Goyal wrote:
> > On Thu, Feb 25, 2010 at 04:12:11PM +0100, Andrea Righi wrote:
> > > On Tue, Feb 23, 2010 at 04:29:43PM -0500, Vivek Goyal wrote:
> > > > On Sun, Feb 21, 2010 at 04:18:45PM +0100, Andrea Righi wrote:
> > > > 
> > > > [..]
> > > > > diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> > > > > index 0b19943..c9ff1cd 100644
> > > > > --- a/mm/page-writeback.c
> > > > > +++ b/mm/page-writeback.c
> > > > > @@ -137,10 +137,11 @@ static struct prop_descriptor vm_dirties;
> > > > >   */
> > > > >  static int calc_period_shift(void)
> > > > >  {
> > > > > -	unsigned long dirty_total;
> > > > > +	unsigned long dirty_total, dirty_bytes;
> > > > >  
> > > > > -	if (vm_dirty_bytes)
> > > > > -		dirty_total = vm_dirty_bytes / PAGE_SIZE;
> > > > > +	dirty_bytes = mem_cgroup_dirty_bytes();
> > > > > +	if (dirty_bytes)
> > > > > +		dirty_total = dirty_bytes / PAGE_SIZE;
> > > > >  	else
> > > > >  		dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
> > > > >  				100;
> > > > 
> > > > Ok, I don't understand this so I better ask. Can you explain a bit how memory
> > > > cgroup dirty ratio is going to play with per BDI dirty proportion thing.
> > > > 
> > > > Currently we seem to be calculating per BDI proportion (based on recently
> > > > completed events), of system wide dirty ratio and decide whether a process
> > > > should be throttled or not.
> > > > 
> > > > Because throttling decision is also based on BDI and its proportion, how
> > > > are we going to fit it with mem cgroup? Is it going to be BDI proportion
> > > > of dirty memory with-in memory cgroup (and not system wide)?
> > > 
> > > IMHO we need to calculate the BDI dirty threshold as a function of the
> > > cgroup's dirty memory, and keep BDI statistics system wide.
> > > 
> > > So, if a task is generating some writes, the threshold to start itself
> > > the writeback will be calculated as a function of the cgroup's dirty
> > > memory. If the BDI dirty memory is greater than this threshold, the task
> > > must start to writeback dirty pages until it reaches the expected dirty
> > > limit.
> > > 
> > 
> > Ok, so calculate dirty per cgroup and calculate BDI's proportion from
> > cgroup dirty? So will you be keeping track of vm_completion events per
> > cgroup or will rely on existing system wide and per BDI completion events
> > to calculate BDI proportion?
> > 
> > BDI proportion are more of an indication of device speed and faster device
> > gets higher share of dirty, so may be we don't have to keep track of
> > completion events per cgroup and can rely on system wide completion events
> > for calculating the proportion of a BDI.
> > 
> > > OK, in this way a cgroup with a small dirty limit may be forced to
> > > writeback a lot of pages dirtied by other cgroups on the same device.
> > > But this is always related to the fact that tasks are forced to
> > > writeback dirty inodes randomly, and not the inodes they've actually
> > > dirtied.
> > 
> > So we are left with following two issues.
> > 
> > - Should we rely on global BDI stats for BDI_RECLAIMABLE and BDI_WRITEBACK
> >   or we need to make these per cgroup to determine actually how many pages
> >   have been dirtied by a cgroup and force writeouts accordingly?
> > 
> > - Once we decide to throttle a cgroup, it should write its inodes and
> >   should not be serialized behind other cgroup's inodes.  
> 
> We could try to save who made the inode dirty
> (inode->cgroup_that_made_inode_dirty) so that during the active
> writeback each cgroup can be forced to write only its own inodes.

Yes, but that will require to store a reference to memcg and will become
little complicated.

I was thinking of just matching the cgroup of task being throttled and
memcg of first dirty page in the inode. So we can possibly implement
something like in memcontroller.

bool memcg_task_inode_cgroup_match(inode)

and this function will retrieve first dirty page and compare the cgroup of
that with task memory cgroup. No hassle of storing a pointer hence
reference to memcg.

Well, we could store css_id, and no need to keep a reference to the
memcg. But I guess not storing anything in inode will be simpler.

Thanks
Vivek