Adapt the gradual reclaim floor to the working set using refaults#40779
Closed
benhillis wants to merge 3 commits into
Closed
Adapt the gradual reclaim floor to the working set using refaults#40779benhillis wants to merge 3 commits into
benhillis wants to merge 3 commits into
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
This PR (stacked PR 3/4) evolves the WSL2 mini-init memory reduction thread’s Gradual reclaim policy by making the reclaim “floor” adaptive to the guest’s working set, using file refaults (/proc/vmstat workingset_refault_file) and PSI memory pressure (/proc/pressure/memory) as feedback signals to back off reclaim when it starts impacting actively-used pages.
Changes:
- Adds a robust procfs reader (
ReadProcFile) and refactors CPU-idle detection to use aggregate busy vs idle jiffies from/proc/stat. - Implements an adaptive Gradual reclaim floor that raises quickly on refault/pressure backoff and decays after sustained calm; scales the max floor to a fraction of total guest RAM.
- Updates DropCache + compaction behavior to run as best-effort, with compaction gated by free-memory growth.
Comment on lines
+3538
to
+3539
| { | ||
| wil::unique_fd Fd{open("/proc/stat", O_RDONLY)}; | ||
| wil::unique_fd Fd{open(Path, O_RDONLY)}; |
Comment on lines
+3845
to
+3851
| const char* Marker = strstr(Buffer, "workingset_refault_file "); | ||
| if (Marker == nullptr) | ||
| { | ||
| return -1; | ||
| } | ||
|
|
||
| return strtoll(Marker + (sizeof("workingset_refault_file ") - 1), nullptr, 10); |
Replace the ring-buffer idle detector and user-CPU-only sampling in the mini-init memory reduction thread with a clearer, helper-based design: - Sample aggregate non-idle CPU time (user, system, irq, softirq, steal) so kernel-bound work keeps the VM out of the idle state, instead of looking at user time alone. - ReadProcFile reads a full procfs snapshot into a caller buffer (close-on-exec, partial-read safe); GetReclaimableCacheBytes / GetFreeMemoryBytes read the relevant counters through it. - Gradual mode reclaims cold page cache (cgroup memory.reclaim) above a fixed floor while CPU-idle, with a hysteresis margin so it does not churn near the floor. - DropCache mode stays gated on sustained CPU idle, drops once, and re-drops only after the reclaimable cache grows meaningfully. - Compaction is gated on free-memory growth so it runs only when there are newly-freed pages worth coalescing. RequestCgroupReclaim performs the memory.reclaim write best-effort: it treats the kernel's expected EAGAIN (some, but not all, pages evicted) as success without logging, and never throws so a transient write error cannot tear down the long-lived reduction thread. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
A CPU-bound workload can sit on gigabytes of cold page cache that a CPU-idle check would never reclaim. Read the PSI "some avg10" memory pressure from /proc/pressure/memory and reclaim cold cache toward the fixed floor whenever pressure is low, even while the VM is busy, backing off once the workload starts stalling on memory. A busy interval reclaims at most a bounded step (c_gradualStepBusyBytes) so a large backlog is drained gently; an idle interval drains the full excess at once. When PSI is unavailable (kernel built without CONFIG_PSI), gradual reclaim falls back to gating on CPU idle. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Reclaiming toward a single fixed floor either gives back too little on large VMs or evicts pages a larger working set immediately faults back in. Make the floor adaptive and self-regulating: - Track file refaults (/proc/vmstat workingset_refault_file) as a signal that reclaim is cutting into the working set. When refaults spike (or PSI pressure rises into the backoff band), raise the floor to protect what the workload is actually using and stop reclaiming. - After sustained calm, decay the floor back toward the base so a shrunken working set is eventually re-probed downward. - Scale the floor's upper bound to a fraction of guest RAM so large working sets on large VMs can be fully protected, falling back to a fixed cap when total RAM is unavailable. The refault counter is parsed unsigned and clamped so it cannot wrap negative and disable the brake. The PSI-unavailable path keeps the same refault brake and floor decay so behavior degrades gracefully on kernels without CONFIG_PSI. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
benhillis
force-pushed
the
benhill/mem-reclaim-3-adaptive
branch
from
3c2691b to
4e60f35
Compare
Member
Author
|
Closing for now. These changes are being submitted one PR at a time; this will be reopened against master once the preceding PR in the series has merged, so it shows only its own incremental change. The branch remains pushed. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked PR 3 of 4 — builds on #40777 and the PSI change. Diff shown is cumulative.
What this does
Reclaiming toward a single fixed floor either gives back too little on large VMs or evicts pages a larger working set immediately faults back in. This makes the floor adaptive and self-regulating:
/proc/vmstat workingset_refault*) as a signal that reclaim is cutting into the working set. When refaults spike (or PSI pressure rises into the backoff band), the floor is raised to protect what the workload is actually using and reclaim stops.CONFIG_PSI.Stack