On Tue, Jul 30, 2024 at 07:13:03PM -0400, David Finkel wrote:
Other mechanisms for querying the peak memory usage of either a process or v1 memory cgroup allow for resetting the high watermark. Restore parity with those mechanisms, but with a less racy API.
For example:
- Any write to memory.max_usage_in_bytes in a cgroup v1 mount resets the high watermark.
- writing "5" to the clear_refs pseudo-file in a processes's proc directory resets the peak RSS.
This change is an evolution of a previous patch, which mostly copied the cgroup v1 behavior, however, there were concerns about races/ownership issues with a global reset, so instead this change makes the reset filedescriptor-local.
Writing any non-empty string to the memory.peak and memory.swap.peak pseudo-files reset the high watermark to the current usage for subsequent reads through that same FD.
Notably, following Johannes's suggestion, this implementation moves the O(FDs that have written) behavior onto the FD write(2) path. Instead, on the page-allocation path, we simply add one additional watermark to conditionally bump per-hierarchy level in the page-counter.
Additionally, this takes Longman's suggestion of nesting the page-charging-path checks for the two watermarks to reduce the number of common-case comparisons.
This behavior is particularly useful for work scheduling systems that need to track memory usage of worker processes/cgroups per-work-item. Since memory can't be squeezed like CPU can (the OOM-killer has opinions), these systems need to track the peak memory usage to compute system/container fullness when binpacking workitems.
Most notably, Vimeo's use-case involves a system that's doing global binpacking across many Kubernetes pods/containers, and while we can use PSI for some local decisions about overload, we strive to avoid packing workloads too tightly in the first place. To facilitate this, we track the peak memory usage. However, since we run with long-lived workers (to amortize startup costs) we need a way to track the high watermark while a work-item is executing. Polling runs the risk of missing short spikes that last for timescales below the polling interval, and peak memory tracking at the cgroup level is otherwise perfect for this use-case.
As this data is used to ensure that binpacked work ends up with sufficient headroom, this use-case mostly avoids the inaccuracies surrounding reclaimable memory.
Suggested-by: Johannes Weiner hannes@cmpxchg.org Suggested-by: Waiman Long longman@redhat.com Acked-by: Johannes Weiner hannes@cmpxchg.org Reviewed-by: Michal Koutný mkoutny@suse.com Signed-off-by: David Finkel davidf@vimeo.com
Reviewed-by: Roman Gushchin roman.gushchin@linux.dev
Thanks!