Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S933221Ab1CaJxK (ORCPT ); Thu, 31 Mar 2011 05:53:10 -0400 Received: from cantor2.suse.de ([195.135.220.15]:55461 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752884Ab1CaJxI (ORCPT ); Thu, 31 Mar 2011 05:53:08 -0400 Date: Thu, 31 Mar 2011 11:53:06 +0200 From: Michal Hocko To: Ying Han Cc: Balbir Singh , KAMEZAWA Hiroyuki , linux-mm@kvack.org, linux-kernel@vger.kernel.org Subject: Re: [RFC 0/3] Implementation of cgroup isolation Message-ID: <20110331095306.GA30290@tiehlicka.suse.cz> References: <20110328093957.089007035@suse.cz> <20110328200332.17fb4b78.kamezawa.hiroyu@jp.fujitsu.com> <4D920066.7000609@gmail.com> <20110330081853.GC15394@tiehlicka.suse.cz> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.20 (2009-06-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4993 Lines: 110 On Wed 30-03-11 10:59:21, Ying Han wrote: > On Wed, Mar 30, 2011 at 1:18 AM, Michal Hocko wrote: > > On Tue 29-03-11 21:23:10, Balbir Singh wrote: > >> On 03/28/11 16:33, KAMEZAWA Hiroyuki wrote: > >> > On Mon, 28 Mar 2011 11:39:57 +0200 > >> > Michal Hocko wrote: > > [...] > >> > Isn't it the same result with the case where no cgroup is used ? > >> > What is the problem ? > >> > Why it's not a problem of configuration ? > >> > IIUC, you can put all logins to some cgroup by using cgroupd/libgcgroup. > >> > > >> > >> I agree with Kame, I am still at loss in terms of understand the use > >> case, I should probably see the rest of the patches > > > > OK, it looks that I am really bad at explaining the usecase. Let's try > > it again then (hopefully in a better way). > > > > Consider a service which serves requests based on the in-memory > > precomputed or preprocessed data. > > Let's assume that getting data into memory is rather costly operation > > which considerably increases latency of the request processing. Memory > > access can be considered random from the system POV because we never > > know which requests will come from outside. > > This workflow will benefit from having the memory resident as long as > > and as much as possible because we have higher chances to be used more > > often and so the initial costs would pay off. > > Why is mlock not the right thing to do here? Well, if the memory would > > be locked and the working set would grow (again this depends on the > > incoming requests) then the application would have to unlock some > > portions of the memory or to risk OOM because it basically cannot > > overcommit. > > On the other hand, if the memory is not mlocked and there is a global > > memory pressure we can have some part of the costly memory swapped or > > paged out which will increase requests latencies. If the application is > > placed into an isolated cgroup, though, the global (or other cgroups) > > activity doesn't influence its cgroup thus the working set of the > > application. > > > If we compare that to mlock we will benefit from per-group reclaim when > > we get over the limit (or soft limit). So we do not start evicting the > > memory unless somebody makes really pressure on the _application_. > > Cgroup limits would, of course, need to be selected carefully. > > > > There might be other examples when simply kernel cannot know which > > memory is important for the process and the long unused memory is not > > the ideal choice. > > Michal, > > Reading through your example, sounds to me you can accomplish the > "guarantee" of the high priority service using existing > memcg mechanisms. > > Assume you have the service named cgroup-A which needs memory > "guarantee". Meantime we want to launch cgroup-B with no memory > "guarantee". What you want is to have cgroup-B uses the slack memory > (not being allocated by cgroup-A), but also volunteer to give up under > system memory pressure. This would require a "guarantee" that no pages are reclaimed from a group if that group is under its soft limit, right? I am thinking if we can achieve that without too many corner cases when cgroups (process's accounted memory) don't leave out much for other memory used by the kernel. That was my concern so I made that isolation rather opt-in without modifying the current reclaim logic too much (there are, of course, parts that can be improved). > So continue w/ my previous post, you can consider the following > configuration in 32G machine. We can only have resident size of > cgroup-A as much as the machine capacity. > > cgroup-A : limit_in_bytes =32G soft_limit_in_bytes = 32G > cgroup-B : limit_in_bytes =20G soft_limit_in_bytes = 0G > > To be a little bit extreme, there shouldn't be memory pressure on > cgroup-A unless it grows above the machine capacity. If the global > memory contention is triggered by cgroup-B, we should steal pages from > it always. > > However, the current implementation of soft_limit needs to be improved > for the example above. Especially when we start having lots of cgroups > running w/ different limit setting, we need to have soft_limit being > efficient and we can eliminate the global lru scanning. Lots of groups is really an issue because we can end up in a situation when everybody is under the limit while there is not much memory left for the kernel. Maybe sum(soft_limit) < kernel_treshold condition would solve this. > The later one breaks the isolation. Sorry, I don't understand. Why would elimination of the global lru scanning break isolation? Or am I misreading you? Thanks -- Michal Hocko SUSE Labs SUSE LINUX s.r.o. Lihovarska 1060/12 190 00 Praha 9 Czech Republic -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/