On Tue, Jan 23, 2024 at 8:48 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > > The revert isn't a straight-forward solution. > > The patch you're reverting fixed conventional reclaim and broke > MGLRU. Your revert fixes MGLRU and breaks conventional reclaim. > > On Tue, Jan 23, 2024 at 05:58:05AM -0800, T.J. Mercier wrote: > > They both are able to make progress. The main difference is that a > > single iteration of try_to_free_mem_cgroup_pages with MGLRU ends soon > > after it reclaims nr_to_reclaim, and before it touches all memcgs. So > > a single iteration really will reclaim only about SWAP_CLUSTER_MAX-ish > > pages with MGLRU. WIthout MGLRU the memcg walk is not aborted > > immediately after nr_to_reclaim is reached, so a single call to > > try_to_free_mem_cgroup_pages can actually reclaim thousands of pages > > even when sc->nr_to_reclaim is 32. (I.E. MGLRU overreclaims less.) > > https://lore.kernel.org/lkml/20221201223923.873696-1-yuzhao@xxxxxxxxxx/ > > Is that a feature or a bug? Feature! > * 1. Memcg LRU only applies to global reclaim, and the round-robin incrementing > * of their max_seq counters ensures the eventual fairness to all eligible > * memcgs. For memcg reclaim, it still relies on mem_cgroup_iter(). > > If it bails out exactly after nr_to_reclaim, it'll overreclaim > less. But with steady reclaim in a complex subtree, it will always hit > the first cgroup returned by mem_cgroup_iter() and then bail. This > seems like a fairness issue. Right. Because the memcg LRU is maintained in pg_data_t and not in each cgroup, I think we are currently forced to have the iteration across all child memcgs for non-root memcg reclaim for fairness. > We should figure out what the right method for balancing fairness with > overreclaim is, regardless of reclaim implementation. Because having > two different approaches and reverting dependent things back and forth > doesn't make sense. > > Using an LRU to rotate through memcgs over multiple reclaim cycles > seems like a good idea. Why is this specific to MGLRU? Shouldn't this > be a generic piece of memcg infrastructure? It would be pretty sweet if it were. I haven't tried to measure this part in isolation, but I know we had to abandon attempts to use per-app memcgs in the past (2018?) because the perf overhead was too much. In recent tests where this feature is used, I see some perf gains which I think are probably attributable to this. > Then there is the question of why there is an LRU for global reclaim, > but not for subtree reclaim. Reclaiming a container with multiple > subtrees would benefit from the fairness provided by a container-level > LRU order just as much; having fairness for root but not for subtrees > would produce different reclaim and pressure behavior, and can cause > regressions when moving a service from bare-metal into a container. > > Figuring out these differences and converging on a method for cgroup > fairness would be the better way of fixing this. Because of the > regression risk to the default reclaim implementation, I'm inclined to > NAK this revert. In the meantime, instead of a revert how about changing the batch size geometrically instead of the SWAP_CLUSTER_MAX constant: reclaimed = try_to_free_mem_cgroup_pages(memcg, - min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX), + (nr_to_reclaim - nr_reclaimed)/2, GFP_KERNEL, reclaim_options); I think that should address the overreclaim concern (it was mentioned that the upper bound of overreclaim was 2 * request), and this should also increase the reclaim rate for root reclaim with MGLRU closer to what it was before.
