Improve WSL2 guest memory reclaim by benhillis · Pull Request #41096 · microsoft/WSL

Ben Hillis (benhillis) · 2026-07-16T19:43:33Z

What this changes

This replaces the existing memory reduction loop with a shared idle detector and simpler reclaim policy.

CPU idle detection uses aggregate busy time instead of user time only.
Free pages are compacted when the VM enters idle and after every reclaim operation.
Cache reclaim waits for two minutes of sustained idle, protecting normal pauses between commands.
Reclaim and compaction CPU are excluded from the next idle sample so the thread does not reset its own grace period.

Modes

Mode	Behavior after two minutes idle
Disabled	No memory reduction thread
DropCache	Run `drop_caches` once for the idle period
Gradual	Reclaim cold cache toward a 128 MB floor using RAM-scaled 256 MB to 1 GB steps

Gradual uses memory.reclaim with swappiness=0, which prevents proactive reclaim from swapping anonymous memory. If memory.reclaim is unavailable, it falls back to DropCache.

The default remains DropCache.

Why

The current implementation waits approximately ten minutes before dropping cache and only measures user CPU time. This leaves vmmemWSL elevated long after real workloads finish and can treat kernel or I/O work as idle.

The new policy returns memory sooner while preserving cache during short user pauses.

DropCache results

Measured with two clean 16-way Linux kernel builds, two read-heavy source archive runs, and three warm-cache reuse runs per configuration.

Metric	Existing DropCache	New DropCache
Kernel build: 50% working-set reduction	591s	177s
Source archive: 50% working-set reduction	542s	142s
Warm reread after 60 seconds idle	0.233s	0.220s
Kernel working set after 15 minutes	2.14 GiB	2.12 GiB
Source archive working set after 15 minutes	1.99 GiB	1.98 GiB

The new DropCache policy reaches the existing eventual footprint roughly seven minutes sooner without affecting the 60-second warm-cache case.

Gradual results

A focused test using the same two-minute grace period showed:

warm rereads after 60 seconds remained at 0.22-0.23 seconds
1.83 GiB of reclaimable cache remained intact through 120 seconds
reclaim began at approximately 129 seconds
cache fell to 0.14 GiB by 140 seconds
vmmemWSL fell from 3.23 GiB to 1.63 GiB by 150 seconds

Gradual reaches a lower sustained footprint because it continues reclaiming regrown cache toward the floor. DropCache remains the default for compatibility and Gradual remains available for users who prioritize minimum memory usage.

Validation

cmake --build . --config Debug -- -m
built and deployed the generated MSI
verified memory.reclaim swappiness=0 support on the WSL 6.18 kernel
collected host vmmemWSL, guest memory, reclaim, compaction, workload timing, and warm-cache reuse data

Copilot

Pull request overview

This PR updates WSL2’s guest-side memory reduction logic to reclaim memory more efficiently and safely while the VM is idle, aiming to reduce host working set without hurting short-pause resume performance.

Changes:

Replaces the user-CPU ring-buffer heuristic with aggregate busy/idle CPU sampling from /proc/stat.
Adds reclaim and compaction behavior tuned for idle periods, including a 2-minute grace period and RAM-scaled reclaim requests via memory.reclaim (with fallback to drop_caches).
Improves robustness of /proc/stat and /proc/meminfo parsing to avoid terminating the background thread on malformed input.

Copilot-Session: ccc721ed-5fa2-41b2-a5f9-5a8502cdf364

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 2 comments.

Copilot-Session: ccc721ed-5fa2-41b2-a5f9-5a8502cdf364

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

David Bennett (dkbennett)

Looks good, recommend adding unit tests for CpuIdleTracker. It would need to be pulled out of the anonymous namespace but could be unit tested for confidence it is correctly tracking these states as expected.

Blue (OneBlue)

LGTM, minor suggestions

Blue (OneBlue) · 2026-07-24T20:10:54Z

+        return false;
+    }
+
+    unsigned long long fields[8] = {};


nit: This might be a good candidate for using std::regex (can be done in a followup though)

Blue (OneBlue) · 2026-07-24T20:13:27Z

+                }
+                else if (!droppedThisIdlePeriod)
+                {
+                    if (WriteToFile("/proc/sys/vm/drop_caches", "1\n") == 0)


Should we switch this to 3 while we're at it to drop metadata as well ?

Ben Hillis (benhillis) · 2026-07-24T22:20:45Z

Good calls, will address as a follow-up.

* Address follow-ups from PR #41096 - Parse the /proc/stat cpu line with std::regex instead of a manual strtoull cursor loop, while retaining the O_CLOEXEC bounded read. - Use drop_caches=3 in DropCache mode to also drop reclaimable slab (dentries/inodes), matching the SReclaimable slab counted by GetReclaimableCacheBytes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 0e37f3d5-68e7-4973-8c27-bf4329c8fd9b * Restrict /proc/stat field separators to spaces/tabs ECMAScript \s matches newlines, so on a truncated aggregate cpu line the optional irq/softirq/steal groups could consume digits from the next line in the read buffer. Use [ \t] so the match cannot span lines. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 0e37f3d5-68e7-4973-8c27-bf4329c8fd9b --------- Co-authored-by: Ben Hillis Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 0e37f3d5-68e7-4973-8c27-bf4329c8fd9b

Ben Hillis (benhillis) requested a review from a team as a code owner July 16, 2026 19:43

Copilot AI review requested due to automatic review settings July 16, 2026 19:43

Copilot started reviewing on behalf of Ben Hillis (benhillis) July 16, 2026 19:44 View session

Copilot AI reviewed Jul 16, 2026

View reviewed changes

Comment thread src/linux/init/util.cpp

Improve WSL2 guest memory reclaim behavior

f4c4747

Ben Hillis (benhillis) force-pushed the user/benhill/improve-memory-reclaim branch from b1f6752 to f4c4747 Compare July 16, 2026 20:14

Use rolling CPU window for memory reclaim

34f8d33

Copilot-Session: ccc721ed-5fa2-41b2-a5f9-5a8502cdf364

Copilot AI review requested due to automatic review settings July 17, 2026 23:52

Copilot started reviewing on behalf of Ben Hillis (benhillis) July 17, 2026 23:53 View session

Copilot AI reviewed Jul 17, 2026

View reviewed changes

Comment thread src/linux/init/util.cpp

Comment thread src/linux/init/util.cpp

Harden memory reclaim sampling

dc0264b

Copilot-Session: ccc721ed-5fa2-41b2-a5f9-5a8502cdf364

Copilot AI review requested due to automatic review settings July 20, 2026 16:55

Copilot started reviewing on behalf of Ben Hillis (benhillis) July 20, 2026 16:55 View session

Copilot AI reviewed Jul 20, 2026

View reviewed changes

Comment thread src/linux/init/util.cpp

Ben Hillis (benhillis) mentioned this pull request Jul 21, 2026

Rework the WSL2 guest memory reduction thread around explicit reclaim helpers #40777

Closed

David Bennett (dkbennett) approved these changes Jul 21, 2026

View reviewed changes

Blue (OneBlue) approved these changes Jul 24, 2026

View reviewed changes

Ben Hillis (benhillis) merged commit 49f841c into master Jul 24, 2026
12 checks passed

Ben Hillis (benhillis) deleted the user/benhill/improve-memory-reclaim branch July 24, 2026 22:20

Ben Hillis (benhillis) mentioned this pull request Jul 24, 2026

Address memory reclaim follow-ups from #41096 #41171

Merged

Navigation Menu

Repository navigation

Improve WSL2 guest memory reclaim - #41096

Improve WSL2 guest memory reclaim#41096
Ben Hillis (benhillis) merged 3 commits into
masterfrom
user/benhill/improve-memory-reclaim

Ben Hillis (benhillis) commented Jul 16, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

David Bennett (dkbennett) left a comment

Uh oh!

Blue (OneBlue) left a comment

Uh oh!

Blue (OneBlue) Jul 24, 2026

Uh oh!

Blue (OneBlue) Jul 24, 2026

Uh oh!

Ben Hillis (benhillis) commented Jul 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Ben Hillis (benhillis) commented Jul 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this changes

Modes

Why

DropCache results

Gradual results

Validation

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

David Bennett (dkbennett) left a comment

Choose a reason for hiding this comment

Uh oh!

Blue (OneBlue) left a comment

Choose a reason for hiding this comment

Uh oh!

Blue (OneBlue) Jul 24, 2026

Choose a reason for hiding this comment

Uh oh!

Blue (OneBlue) Jul 24, 2026

Choose a reason for hiding this comment

Uh oh!

Ben Hillis (benhillis) commented Jul 24, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Ben Hillis (benhillis) commented Jul 16, 2026 •

edited

Loading