On Mon, 2020-02-03 at 15:16 -0800, Tyler Sanderson wrote: > > > On Mon, Feb 3, 2020 at 1:22 PM Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx> wrote: > > On Mon, 2020-02-03 at 12:32 -0800, Tyler Sanderson wrote: > > > There were apparently good reasons for moving away from OOM notifier > > > callback: > > > https://lkml.org/lkml/2018/7/12/314 > > > https://lkml.org/lkml/2018/8/2/322 > > > > > > In particular the OOM notifier is worse than the shrinker because: > > > It is last-resort, which means the system has already gone through > > > heroics to prevent OOM. Those heroic reclaim efforts are expensive and > > > impact application performance. > > > It lacks understanding of NUMA or other OOM constraints. > > > It has a higher potential for bugs due to the subtlety of the callback > > > context. > > > Given the above, I think the shrinker API certainly makes the most sense > > > _if_ the balloon size is static. In that case memory should be reclaimed > > > from the balloon early and proportionally to balloon size, which the > > > shrinker API achieves. > > > > The problem is the shrinker doesn't have any concept of tiering or > > priority. I suspect he reason for using the OOM notification is because in > > practice it should be the last thing we are pulling memory out of with > > things like page cache and slab caches being first. Once we have pages > > that are leaked out of the balloon by the shrinker it will trigger the > > balloon wanting to reinflate. > > Deciding whether to trade IO performance (page cache) for memory-usage > efficiency (balloon) seems use-case dependent. > Deciding when to re-inflate is a similar policy choice. > > If the balloon's shrinker priority is hard-coded to "last-resort" then > there would be no way to implement a policy where page cache growth > could shrink the balloon. > The current balloon implementation allows the host to implement this > policy and tune the tradeoff between balloon and page cache. I'm not saying the priority has to be hard coded, but ideally we should be able to have some sort of priority so we can decide that sort of stuff. The general idea is we want to be able to somehow make it so that if the shrinker is running we aren't trying to inflate the balloon, and us inflating the balloon will not trigger us to trigger the shrinker on ourselves. If our preference is to deflate the balloon rather than maintain page cache then we really should have some way for us to avoid triggering flushing of the page cache if we are inflating the balloon. What we need is a clear definition for what we are willing to flush to fill the balloon and what we are not and then we can use that to decide when we stop inflating and start deflating based on memory pressure. > > Ideally if the shrinker is running we > > shouldn't be able to reinflate the balloon, and if we are reinflating the > > balloon we shouldn't need to run the shrinker. The fact that we can do > > both at the same time is problematic. > > I agree that this is inefficient. > > > > However, if the balloon is inflating and intentionally causing memory > > > pressure then this results in the inefficiency pointed out earlier. > > > > > > If the balloon is inflating but not causing memory pressure then there > > > is no problem with either API. > > > > The entire point of the balloon is to cause memory pressure. Otherwise > > essentially all we are really doing is hinting since the guest doesn't > > need the memory and isn't going to use it any time soon. > > Causing memory pressure is just a mechanism to achieve increased > reclaim. If there was a better mechanism (like the fine-grained-cache- > shrinking one discussed below) then I think the balloon device would be > perfectly justified in using that instead (and maybe "balloon" becomes a > misnomer. Oh well). Agreed. There have been a number of threads on how to go about doing a proactive reclaim in the last month or two. In addition there have been a number of solutions that I have seen going around. I am hoping this is one of the topics we can go over at LSF/MM. > > > This suggests another route: rather than cause memory pressure to shrink > > > the page cache, the balloon could issue the equivalent of "echo 3 > > > > /proc/sys/vm/drop_caches". > > > Of course ideally, we want to be more fine grained than "drop > > > everything". We really want an API that says "drop everything that > > > hasn't been accessed in the last 5 minutes". > > > > > > This would eliminate the need for the balloon to cause memory pressure > > > at all which avoids the inefficiency in question. Furthermore, this > > > pairs nicely with the FREE_PAGE_HINT feature. > > > > Something similar was brought up in the discussion we had about this in my > > patch set. The problem is, by trying to use a value like "5 minutes" it > > implies that we are going to need to track some extra state somewhere to > > determine that value. > > > > An alternative is to essentially just slowly shrink memory for the guest. > > We had some discussion about this in another thread, and the following > > code example was brought up as a way to go about doing that: > > https://github.com/Conan-Kudo/omv-kernel-rc/blob/master/0154-sysctl-vm-Fine-grained-cache-shrinking.patch > > > > The idea is you essentially just slowly bleed the memory from the guest by > > specifying some amount of MB of cache to be freed on some regular > > interval. > > Makes sense. Whatever API is settled on, I'd just propose that we allow > the host to invoke it via the balloon device since the host has a host- > global view of memory and can make decisions that an individual guest > cannot. > > Alex, what is the status of your fine-grained-cache-shrinking patch? It > seems like a really good idea. Well the patch I mentioned above is something that someone else was working on back in 2017. I have been working on the page reporting piece for the last year or so trying to get that sorted before I get to the matter of dealing with how to go about triggering reclaim. I figure I can work on reclaim while getting the QEMU bits accepted. You might take a look at the discussion we had on my free page reporting interface. I think we ended up with something like 3 or 4 different ways to approach reclaim being proposed. The discussion can be found here: https://lore.kernel.org/lkml/20200122173040.6142.39116.stgit@localhost.localdomain/ Thanks. - Alex
