Repository navigation
perf: unmeasured leads from the second whole-repo audit (after #467) #468
Description
Performance leads the second whole-repo static audit (on bcca6c8) found and left without code because none is measured. Each needs a measurement before a change; costs below are estimates.
| Where | Lead | What to measure |
|---|---|---|
pool/hibernate.go idle sweep |
The sweep fans hibernates out wider than the engine's 8 memory dumps; a victim past the 8th holds its claim's transition with the intent committed while it waits for a dump slot, so a request to it waits for the whole hibernate instead of being served | request latency to claims queued behind a large idle sweep |
pool/liveness.go evictDeadWarm |
Each non-live or deleted VM event walks every warm VM under the manager lock; #459 indexed only claims | lock hold per event at the largest warm count in use |
store/dir/dir.go readRecord |
Every listing reads a digest file and labels.json per record on the checkpoint store too, where neither exists (#461 fixed this for s3 only); on a shared FUSE dir store each is a remote lookup |
listing time at a few thousand checkpoints on a shared dir store |
utils/clone_linux.go |
The sparse copy goes through OffsetWriter/SectionReader, which hide the files from copy_file_range, so data moves through a 32 KiB user-space buffer |
reader fetch time of a large disk on ext4 |
pool/setpools.go applyPools |
Every applied pool set trims each pool to its current target, skipping the shrink dwell, so a PUT /v1/pools during a burst's tail destroys warm VMs above the floor that the next burst then re-clones |
claim tier mix around a pool-set change on an adaptive pool |
engine/engine.go DialGuestPort |
The port-forward reply is copied into a string and back into bytes per relay connection | port relay open cost (minor) |
Activity
CMGS commentedon Nov 6, 2026
More unmeasured leads from the static audit of 2026-11-06 (PR #481):
store/dirSweepStagingat startupmeta.jsonbefore the server starts; on a shared FUSEcheckpoint_dirthat is about 2N remote ops on every node's bootGET /v1/checkpointson s3ListenAndServeengine/portconn.goserver/preview.gosdk/pythonframes.pyCMGS commentedon Nov 10, 2026
Status on main (v0.1.31):
copyDatanow copies each data range in 1 MiBReadAt/WriteAtblocks and leaves zero blocks as holes. Hardware A/B on ext4 without reflink: −4 to −6% against the old code. Acopy_file_rangearm is still 4–6% faster than utils: copy data ranges in 1 MiB blocks and leave zero blocks as holes #542 (F1 830.8 vs 877.6 ms, F2 1589.3 vs 1650.9 ms median), so the remaining lead is that gap. On a JuiceFS dir store, the first read of an archived claim no longer writes the zero bytes.ar_records, and the listing usesPrefix=ck_, so they are never listed or read. Eachck_record'smeta.jsonis still read to filter by tenant.runBoundedruns at mostcap(maintSem)workers, so at mostcap(maintSem) − 8victims wait for a dump slot while holding their transition. On a large host that is still many (maintSemisrefill − refill/4). Not re-measured.readRecorddigest and labels reads, theapplyPoolstrim without the shrink dwell, theDialGuestPortstring copy,SweepStagingat startup, reconcile before serve, the port relay copy, the preview 4 KiB frames, and the Pythonframes.pyslice copy.