On Mon, Apr 16, 2012 at 8:15 AM, Michal Hocko <mhocko@xxxxxxx> wrote: > On Thu 12-04-12 04:22:33, Johannes Weiner wrote: >> On Wed, Apr 11, 2012 at 08:45:33PM -0400, Rik van Riel wrote: >> > On 04/11/2012 06:00 PM, Ying Han wrote: >> > >1. If soft_limit are all set to MAX, it wastes first three periority iterations >> > >without scanning anything. >> > > >> > >2. By default every memcg is eligibal for softlimit reclaim, and we can also >> > >set the value to MAX for special memcg which is immune to soft limit reclaim. >> > > >> > >This idea is based on discussion with Michal and Johannes from LSF. >> > >> > Combined with patch 2/5, would this not result in always >> > returning "reclaim from this memcg" for groups without a >> > configured softlimit, while groups with a configured >> > softlimit only get reclaimed from when they are over >> > their limit? >> > >> > Is that the desired behaviour when a system has some >> > cgroups with a configured softlimit, and some without? >> >> Yes, in general I think this new behaviour is welcome. >> >> In the past, soft limits were only used to give excess memory a lower >> priority and there was no particular meaning associated with "being >> below your soft limit". This change makes it so that soft limits are >> actually a minimum guarantee, too, so you wouldn't get reclaimed if >> you behaved (if possible): >> >> A-unconfigured B-below-softlimit >> old: reclaim reclaim >> new: reclaim no reclaim (if possible) >> >> The much less obvious change here, however, is that we no longer put >> extra pressure on groups above their limit compared to unconfigured >> groups: >> >> A-unconfigured B-above-softlimit >> old: reclaim reclaim twice >> new: reclaim reclaim > > Agreed and I guess that the above should be a part of the changelog. > This is changing previous behavior and we should rather be explicit > about that. Ok, I will include it on next post. Thanks ! --Ying > >> I still think that it's a reasonable use case to put a soft limit on a >> workload to "nice" it memory-wise, without looking at the machine as a >> whole and configuring EVERY cgroup based on global knowledge and >> static partitioning of the machine. > > -- > Michal Hocko > SUSE Labs > SUSE LINUX s.r.o. > Lihovarska 1060/12 > 190 00 Praha 9 > Czech Republic -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href