Re: [patch 04/11] mm: memcg: per-priority per-zone hierarchy scangenerations

[KAMEZAWA Hiroyuki]

On Tue, 13 Sep 2011 13:03:01 +0200
Johannes Weiner <jweiner@xxxxxxxxxx> wrote:

> On Tue, Sep 13, 2011 at 07:27:59PM +0900, KAMEZAWA Hiroyuki wrote:
> > On Mon, 12 Sep 2011 12:57:21 +0200
> > Johannes Weiner <jweiner@xxxxxxxxxx> wrote:
> > 
> > > Memory cgroup limit reclaim currently picks one memory cgroup out of
> > > the target hierarchy, remembers it as the last scanned child, and
> > > reclaims all zones in it with decreasing priority levels.
> > > 
> > > The new hierarchy reclaim code will pick memory cgroups from the same
> > > hierarchy concurrently from different zones and priority levels, it
> > > becomes necessary that hierarchy roots not only remember the last
> > > scanned child, but do so for each zone and priority level.
> > > 
> > > Furthermore, detecting full hierarchy round-trips reliably will become
> > > crucial, so instead of counting on one iterator site seeing a certain
> > > memory cgroup twice, use a generation counter that is increased every
> > > time the child with the highest ID has been visited.
> > > 
> > > Signed-off-by: Johannes Weiner <jweiner@xxxxxxxxxx>
> > 
> > I cannot image how this works. could you illustrate more with easy example ?
> 
> Previously, we did
> 
> 	mem = mem_cgroup_iter(root)
> 	  for each priority level:
> 	    for each zone in zonelist:
> 
> and this would reclaim memcg-1-zone-1, memcg-1-zone-2, memcg-1-zone-3
> etc.
> 
yes.

> The new code does
> 
> 	for each priority level
> 	  for each zone in zonelist
>             mem = mem_cgroup_iter(root)
> 
> but with a single last_scanned_child per memcg, this would scan
> memcg-1-zone-1, memcg-2-zone-2, memcg-3-zone-3 etc, which does not
> make much sense.
> 
> Now imagine two reclaimers.  With the old code, the first reclaimer
> would pick memcg-1 and scan all its zones, the second reclaimer would
> pick memcg-2 and reclaim all its zones.  Without this patch, the first
> reclaimer would pick memcg-1 and scan zone-1, the second reclaimer
> would pick memcg-2 and scan zone-1, then the first reclaimer would
> pick memcg-3 and scan zone-2.  If the reclaimers are concurrently
> scanning at different priority levels, things are even worse because
> one reclaimer may put much more force on the memcgs it gets from
> mem_cgroup_iter() than the other reclaimer.  They must not share the
> same iterator.
> 
> The generations are needed because the old algorithm did not rely too
> much on detecting full round-trips.  After every reclaim cycle, it
> checked the limit and broke out of the loop if enough was reclaimed,
> no matter how many children were reclaimed from.  The new algorithm is
> used for global reclaim, where the only exit condition of the
> hierarchy reclaim is the full roundtrip, because equal pressure needs
> to be applied to all zones.
> 
Hm, ok, maybe good for global reclam.
Is this used for both of reclaim-by-limit and global-reclaim ?
If so, I need to abandon node-selection-logic for reclaim-by-limit
and nodemask-for-memcg which shows me very good result. 
I'll be sad ;)

Thanks,
-Kame



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/