DrumsBench

Experiments / EXP-033

Warm fork: a prefork hands out 4 copies of a desktop with Chrome and Calc open in about 0.08 to 0.14 s at the call on two hosts, 0 of 360 copies invalid; the first replication failed and the 0.139 s did not reproduce exactly

Meet the fork bar (p50 under 250 ms, p99 under 600 ms, 0 invalid children) for fork(4) of a desktop with real apps open, using a one-action-ahead prefork, and replicate it on fresh hosts.

preliminaryA real observation that needs more runs, a positive control, or an independent check before it counts.
Date
2026-11-10 to 2026-11-19
Arms
drums-api
Hardware
m8i.16xlarge spot (64 vCPU), us-east-1a, 250 GiB gp3; main host, first replication host, rep2 host
OS
Ubuntu 24.04, kernel 7.0.0-1014-aws
Workload
W1p: fork(4) of a parent with Chrome and LibreOffice Calc open and typed state, 30 forks per run, one desktop, client on the host
Code
sota/t2b at 4a80f0d (private research archive)
Evidence
runbooks/sota/t2b/20261110T0400Z/PREREGISTRATION.md and FREEZE.txt (frozen 2026-11-10T06:59:26Z, sha256 6efb4f9e...)
runbooks/sota/t2b/20261110T0400Z/REPORT.md, raw/main/tables.md, raw/replication/tables.md, raw/explore/, raw/main/crash-events.txt
runbooks/sota/t2b-rep2/20261119T0628Z/PREREGISTRATION.md and FREEZE.txt (frozen 2026-11-19T06:31:02Z, sha256 15019ce8...)
runbooks/sota/t2b-rep2/20261119T0628Z/REPORT.md, tables.md, DEVIATIONS.md, raw/A, raw/B, raw/explore, raw/SHA256SUMS
Limitation
A prefork child is the parent as it was 0.1 s after the previous action, prepared during think time; a cold fork is about 2.0 s and misses the bar, and p99 is about 2.8 s whenever one fork in 30 is not served by the prefork.

In one paragraph

A cold fork of a desktop with Chrome and LibreOffice Calc open into 4 copies takes about 2.0 s on every build, above the 250 ms bar, and its floor is the 70-process dump. The prefork moves that work into the agent's think time: about 0.1 s after each action Drums captures the parent and prepares 4 children (about 2.1 s), and hands them out at the next fork call only if nothing has changed (no input, no task start or exit, same desktop state hash). On the pre-registered main run the wait at the call was 0.139 s p50 (CI [0.114, 0.168]), 29 of 30 forks served, 0 of 120 copies invalid. The first fresh-host replication failed: one Chrome crash exit the crash watch could not attribute tainted every running episode, and 96 of 120 copies were correctly refused. With the attribution fix, a second replication on a fresh host (2026-11-19) gave 0.080 s on the frozen release and 0.106 s on drumsd v0.2.1, 0 of 120 invalid each. Both were faster than 0.139 s and outside the pre-registered plus-or-minus 20%, so the honest line is "about 0.08 to 0.14 s p50 across two hosts", not a reproduced 0.139 s.

Setup

m8i.16xlarge spot, us-east-1a, kernel 7.0.0-1014-aws, 250 GiB gp3. Main run: the T2b release (drumsd sota/t2b, release 4a80f0d). Rep2 arm A: the same frozen release (identical drumsd binary); arm B: drumsd v0.2.1, which carries the crash-attribution fix fe3291c and a freeze before the cgroup kill.

What ran

2026-11-10: the pre-registered main run (30 W1p forks), then a replication on a fresh host from REPRODUCE.md alone, then a labelled post-hoc rerun of 60 forks with fe3291c on a fresh host (60 of 60 served, 0 of 240 invalid, p50 0.156 s, p99 0.644 s). 2026-11-19: rep2, both arms on one fresh host, plus an exploratory arm B run with its own origin checkpoint (0.092 s, 29 of 30, 0 of 120; does not count).

Results

T2b: warm fork (prefork) of a desktop with Chrome and LibreOffice Calc open into 4 copies, main run and two fresh-host replications preliminary2026-11-10 · T2b main host; rep2 host

MetricValueUnitnNote
Main run (2026-11-10): prefork fork(4) at the call0.139 (p95 0.647)s3095% CI [0.114, 0.168]; p99 2.816 s (the first fork); 29/30 claimed; 0/120 copies invalid; bar met on p50, p99 inconclusive
First replication (2026-11-10, fresh host): prefork fork(4)2.35s30NOT replicated: 6/30 claimed, 96/120 copies refused as tainted after one unattributed Chrome SIGTRAP exit (fail-closed); fixed post hoc in fe3291c
rep2 arm A (2026-11-19, frozen T2b release): prefork fork(4)0.08 (p95 0.501)s3095% CI [0.049, 0.097]; p99 0.535 s; 30/30 claimed; 0/120 invalid; met bar but outside +-20% (faster)
rep2 arm B (2026-11-19, drumsd v0.2.1): prefork fork(4)0.106 (p95 1.58)s3095% CI [0.080, 0.117]; p99 2.789 s (one fork in 30 not served by the prefork ran cold); 29/30 claimed; 0/120 invalid
Copies invalid, main run and both rep2 arms0 / 360copies0/120 in each run; 95% upper bound 0.030 per run
Cold fork(4), no prefork (rep2 arm B, secondary)2.01 (p95 2.43)s3095% CI [1.981, 2.029]; main run 2.040 s; the 250 ms bar is not met by a cold fork on any build

Host: m8i.16xlarge spot, us-east-1a, 250 GiB gp3, Ubuntu 24.04, kernel 7.0.0-1014-aws. client on the host; one desktop at a time. Workload: W1p: fork(4) of a parent with Chrome and LibreOffice Calc open and typed state; the prefork pipeline (capture plus 4 children, about 2.1 s) runs during think time and is claimed at the next fork call. Method: T2b pre-registered (frozen 2026-11-10T06:59:26Z, sha256 6efb4f9e...); rep2 pre-registered (frozen 2026-11-19T06:31:02Z, sha256 15019ce8...). 30 forks per arm, bootstrap CIs 10,000 resamples. Bar: fork p50 < 250 ms, p99 < 600 ms, 0 invalid children. Replication tolerance +-20%. A prefork is claimed only if there was no input, no task start or exit since the capture, and the desktop state hash matches; children are the parent as it was 0.1 s after the previous action.. Source: research archive: runbooks/sota/t2b/20261110T0400Z/REPORT.md; research archive: runbooks/sota/t2b/20261110T0400Z/raw/main/tables.md (line 11); research archive: runbooks/sota/t2b/20261110T0400Z/raw/replication/tables.md (line 5); research archive: runbooks/sota/t2b-rep2/20261119T0628Z/REPORT.md; research archive: runbooks/sota/t2b-rep2/20261119T0628Z/tables.md.

  • The exact 0.139 s did not reproduce within +-20% on either rep2 arm (both faster); quote it as about 0.08 to 0.14 s p50 across the main host and the rep2 host, not as a reproduced 0.139 s. Cause of the faster replication not established.
  • p99 is not under the bar whenever a prefork goes unserved: one cold fork in 30 makes it about 2.7 to 2.8 s.
  • A prefork child is the parent as it was 0.1 s after the previous action. It is not comparable to a vendor fork of a running sandbox without that semantics.
  • Rep2 deviation 4: arm B's parent was restored from the origin checkpoint arm A made; an exploratory run with B's own origin gave 0.092 s, 29/30, 0/120.

Failures and surprises

Limitations

Single desktop on an idle host; client on the host. Prefork semantics differ from a vendor fork of a running sandbox, so the 0.08 to 0.14 s is not comparable to E2B's 5.15 s or Prime's 11.0 s without saying so. The v0.2.0 integration re-checked the cold fork (1.949 s), not the prefork.

Reproduce

REPRODUCE.md and bench/reproduce.sh in the T2b folder; bench/rep2.sh and run_all.sh in the rep2 folder (analyzer unchanged, seed 20261110, 10,000 resamples).

Sources

Both pre-registrations and freezes, both reports, the raw tables and crash-event files, listed under Evidence. Research archive, internal for now.