On Wed, May 31, 2023 at 10:51:01AM +0800, zhaoyang.huang wrote: > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx> > > This patch fixes unproductive reclaiming of CMA pages by skipping them when they > are not available for current context. It is arise from bellowing OOM issue, which > caused by large proportion of MIGRATE_CMA pages among free pages. Hello, I've been looking into a problem with high memory pressure causing OOMs in some of our workloads, and it seems that this change may have introduced lock contention when there is high memory pressure. I've collected some metrics for my specific workload that suggest this change has increased the lruvec->lru_lock waittime-max by 500x and the waittime-avg by 20x. Experiment ========== The experiment involved 100 hosts, each with 64GB of memory and a single Xeon 8321HC CPU. The experiment ran for over 80 hours. Half of the hosts (50) were configured with the patch reverted and lock stat enabled, while the other half was run against the upstream version. All machines had hugetlb_cma=6G set as a command-line argument. In this context, "upstream" refers to kernel release 6.9 with some minor changes that should not impact the results. Workload ======== The workload is a Java based application that fully utilized the memory, in fact, the JVM runs with `-Xms50735m -Xmx50735m` arguments. Results: ======= A few values from lockstat: waittime-max waittime-total waittime-avg holdtime-max 6.9: 242889 15618873933 715 17485 6.9-with-revert: 487 688563299 34 464 The full data could be seen at: https://docs.google.com/spreadsheets/d/1Dl-8ImlE4OZrfKjbyWAIWWuQtgD3fwEEl9INaZQZ4e8/edit?usp=sharing Possible causes: ================ I've been discussing this with colleagues and we're speculating that the high contention might be linked to the fact that CMA regions are now being skipped. This could potentially extend the duration of the isolate_lru_folios() 'while' loop, resulting in increased pressure on the lock. However, I want to emphasize that I'm not an expert in this area and I am simply sharing the data I collected.
