memcg: sc.nr_to_reclaim should be initialized
in [Kernel]
|
Prev: [PATCH 1/7] memcg: sc.nr_to_reclaim should be initialized
Next: memcg: nid and zid can be calculated from zone
From: Balbir Singh on 25 Jul 2010 04:30 On Fri, Jul 23, 2010 at 1:03 PM, KOSAKI Motohiro <kosaki.motohiro(a)jp.fujitsu.com> wrote: >> * KOSAKI Motohiro <kosaki.motohiro(a)jp.fujitsu.com> [2010-07-16 19:13:31]: >> >> > Currently, mem_cgroup_shrink_node_zone() initialize sc.nr_to_reclaim as 0. >> > It mean shrink_zone() only scan 32 pages and immediately return even if >> > it doesn't reclaim any pages. >> > >> > This patch fixes it. >> > >> > Signed-off-by: KOSAKI Motohiro <kosaki.motohiro(a)jp.fujitsu.com> >> > --- >> > �mm/vmscan.c | � �1 + >> > �1 files changed, 1 insertions(+), 0 deletions(-) >> > >> > diff --git a/mm/vmscan.c b/mm/vmscan.c >> > index 1691ad0..bd1d035 100644 >> > --- a/mm/vmscan.c >> > +++ b/mm/vmscan.c >> > @@ -1932,6 +1932,7 @@ unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem, >> > � � � � � � � � � � � � � � � � � � � � � � struct zone *zone, int nid) >> > �{ >> > � � struct scan_control sc = { >> > + � � � � � .nr_to_reclaim = SWAP_CLUSTER_MAX, >> > � � � � � � .may_writepage = !laptop_mode, >> > � � � � � � .may_unmap = 1, >> > � � � � � � .may_swap = !noswap, >> >> Could you please do some additional testing on >> >> 1. How far does this push pages (in terms of when limit is hit)? > > 32 pages per mem_cgroup_shrink_node_zone(). > > That said, the algorithm is here. > > 1. call mem_cgroup_largest_soft_limit_node() > � calculate largest cgroup > 2. call mem_cgroup_shrink_node_zone() and shrink 32 pages > 3. goto 1 if limit is still exceed. > > If it's not your intention, can you please your intended algorithm? We set it to 0, since we care only about a single page reclaim on hitting the limit. IIRC, in the past we saw an excessive pushback on reclaiming SWAP_CLUSTER_MAX pages, just wanted to check if you are seeing the same behaviour even now after your changes. Balbir -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: KOSAKI Motohiro on 25 Jul 2010 05:50 > >> 1. How far does this push pages (in terms of when limit is hit)? > > > > 32 pages per mem_cgroup_shrink_node_zone(). > > > > That said, the algorithm is here. > > > > 1. call mem_cgroup_largest_soft_limit_node() > > � calculate largest cgroup > > 2. call mem_cgroup_shrink_node_zone() and shrink 32 pages > > 3. goto 1 if limit is still exceed. > > > > If it's not your intention, can you please your intended algorithm? > > We set it to 0, since we care only about a single page reclaim on > hitting the limit. IIRC, in the past we saw an excessive pushback on > reclaiming SWAP_CLUSTER_MAX pages, just wanted to check if you are > seeing the same behaviour even now after your changes. Actually, we have 32 pages reclaim batch size. (see nr_scan_try_batch() and related functions) thus <32 value doesn't works as your intended. But, If you run your test again, and (if there is) report any bugs. I'm very glad and fix it soon. Thanks. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Balbir Singh on 25 Jul 2010 12:50
On Sun, Jul 25, 2010 at 3:18 PM, KOSAKI Motohiro <kosaki.motohiro(a)jp.fujitsu.com> wrote: >> >> 1. How far does this push pages (in terms of when limit is hit)? >> > >> > 32 pages per mem_cgroup_shrink_node_zone(). >> > >> > That said, the algorithm is here. >> > >> > 1. call mem_cgroup_largest_soft_limit_node() >> > � calculate largest cgroup >> > 2. call mem_cgroup_shrink_node_zone() and shrink 32 pages >> > 3. goto 1 if limit is still exceed. >> > >> > If it's not your intention, can you please your intended algorithm? >> >> We set it to 0, since we care only about a single page reclaim on >> hitting the limit. IIRC, in the past we saw an excessive pushback on >> reclaiming SWAP_CLUSTER_MAX pages, just wanted to check if you are >> seeing the same behaviour even now after your changes. > > Actually, we have 32 pages reclaim batch size. (see nr_scan_try_batch() and related functions) > thus <32 value doesn't works as your intended. > > But, If you run your test again, and (if there is) report any bugs. I'm very glad and fix it soon. > I understand that, the point is when to do stop the reclaim (do we really need 32 pages to stop the reclaim, when we hit the limit, something as low as a single page can help). This is quite a subtle thing, I'd mark it as low priority. I'll definitely come back if I see unexpected behaviour. Balbir -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |