Michal Hocko <mhocko@xxxxxxxx> writes: > On Thu 15-12-22 13:50:14, Huang, Ying wrote: >> Michal Hocko <mhocko@xxxxxxxx> writes: >> >> > On Tue 13-12-22 11:29:45, Mina Almasry wrote: >> >> On Tue, Dec 13, 2022 at 6:03 AM Michal Hocko <mhocko@xxxxxxxx> wrote: >> >> > >> >> > On Tue 13-12-22 14:30:40, Johannes Weiner wrote: >> >> > > On Tue, Dec 13, 2022 at 02:30:57PM +0800, Huang, Ying wrote: >> >> > [...] >> >> > > > After these discussion, I think the solution maybe use different >> >> > > > interfaces for "proactive demote" and "proactive reclaim". That is, >> >> > > > reconsider "memory.demote". In this way, we will always uncharge the >> >> > > > cgroup for "memory.reclaim". This avoid the possible confusion there. >> >> > > > And, because demotion is considered aging, we don't need to disable >> >> > > > demotion for "memory.reclaim", just don't count it. >> >> > > >> >> > > Hm, so in summary: >> >> > > >> >> > > 1) memory.reclaim would demote and reclaim like today, but it would >> >> > > change to only count reclaimed pages against the goal. >> >> > > >> >> > > 2) memory.demote would only demote. >> >> > > >> >> >> >> If the above 2 points are agreeable then yes, this sounds good to me >> >> and does address our use case. >> >> >> >> > > a) What if the demotion targets are full? Would it reclaim or fail? >> >> > > >> >> >> >> Wei will chime in if he disagrees, but I think we _require_ that it >> >> fails, not falls back to reclaim. The interface is asking for >> >> demotion, and is called memory.demote. For such an interface to fall >> >> back to reclaim would be very confusing to userspace and may trigger >> >> reclaim on a high priority job that we want to shield from proactive >> >> reclaim. >> > >> > But what should happen if the immediate demotion target is full but >> > lower tiers are still usable. Should the first one demote before >> > allowing to demote from the top tier? >> > >> >> > > 3) Would memory.reclaim and memory.demote still need nodemasks? >> >> >> >> memory.demote will need a nodemask, for sure. Today the nodemask would >> >> be useful if there is a specific node in the top tier that is >> >> overloaded and we want to reduce the pressure by demoting. In the >> >> future there will be N tiers and the nodemask says which tier to >> >> demote from. >> > >> > OK, so what is the exact semantic of the node mask. Does it control >> > where to demote from or to or both? >> > >> >> I don't think memory.reclaim would need a nodemask anymore? At least I >> >> no longer see the use for it for us. >> >> >> >> > > Would >> >> > > they return -EINVAL if a) memory.reclaim gets passed only toptier >> >> > > nodes or b) memory.demote gets passed any lasttier nodes? >> >> > >> >> >> >> Honestly it would be great if memory.reclaim can force reclaim from a >> >> top tier nodes. It breaks the aginig pipeline, yes, but if the user is >> >> specifically asking for that because they decided in their usecase >> >> it's a good idea then the kernel should comply IMO. Not a strict >> >> requirement for us. Wei will chime in if he disagrees. >> > >> > That would require a nodemask to say which nodes to reclaim, no? The >> > default behavior should be in line with what standard memory reclaim >> > does. If the demotion is a part of that process so should be >> > memory.reclaim part of it. If we want to have a finer control then a >> > nodemask is really a must and then the nodemaks should constrain both >> > agining and reclaim. >> > >> >> memory.demote returning -EINVAL for lasttier nodes makes sense to me. >> >> >> >> > I would also add >> >> > 4) Do we want to allow to control the demotion path (e.g. which node to >> >> > demote from and to) and how to achieve that? >> >> >> >> We care deeply about specifying which node to demote _from_. That >> >> would be some node that is approaching pressure and we're looking for >> >> proactive saving from. So far I haven't seen any reason to control >> >> which nodes to demote _to_. The kernel deciding that based on the >> >> aging pipeline and the node distances sounds good to me. Obviously >> >> someone else may find that useful. >> > >> > Please keep in mind that the interface should be really prepared for >> > future extensions so try to abstract from your immediate usecases. >> >> I see two requirements here, one is to control the demotion source, that >> is, which nodes to free memory. The other is to control the demotion >> path. I think that we can use two different parameters for them, for >> example, "from=<demotion source nodes>" and "to=<demotion target >> nodes>". In most cases we don't need to control the demotion path. >> Because in current implementation, the nodes in the lower tiers in the >> same socket (local nodes) will be preferred. I think that this is >> the desired behavior in most cases. > > Even if the demotion path is not really required at the moment we should > keep in mind future potential extensions. E.g. when a userspace based > balancing is to be implemented because the default behavior cannot > capture userspace policies (one example would be enforcing a > prioritization of containers when some container's demoted pages would > need to be demoted further to free up a space for a different > workload). Yes. We should consider the potential requirements. >> >> > 5) Is the demotion api restricted to multi-tier systems or any numa >> >> > configuration allowed as well? >> >> > >> >> >> >> demotion will of course not work on single tiered systems. The >> >> interface may return some failure on such systems or not be available >> >> at all. >> > >> > Is there any strong reason for that? We do not have any interface to >> > control NUMA balancing from userspace. Why cannot we use the interface >> > for that purpose? >> >> Do you mean to demote the cold pages from the specified source nodes to >> the specified target nodes in different sockets? We don't do that to >> avoid loop in the demotion path. If we prevent the target nodes from >> demoting cold pages to the source nodes at the same time, it seems >> doable. > > Loops could be avoid by properly specifying from and to nodes if this is > going to be a fine grained interface to control demotion. Yes. Best Regards, Huang, Ying
