M9: parking (scale) #10

New issue

Open

opened 2026-10-02 00:16:05 +00:00 by jhgaylor · 0 comments

jhgaylor commented

2026-10-02 00:16:05 +00:00

Owner

Depends on: #3

Trigger: Panes with real history still blow the memory budget after step 1 (rerun the S9 benchmark).

Step 1 (cheap memory wins) is split out as a separate issue and comes first.

From PLAN.md, section "M9".

Superlogical's server numbers are about 400KB per terminal against 5MB for
tmux, and unparking takes about 200µs.

Trigger: any one of these:
- geek's daemon holds more than about 50 panes;
- an agent fleet runs;
- RSS per idle pane is measured above 2MB.
Measure first. The milestone starts with a benchmark (RSS per pane at
empty, full screen and 10k scrollback; per attached client) committed as a
test.
Step 2: parking, only if panes with real history still blow the budget.
S9 found nothing that argues for PTY or client-buffer parking. Parking
can't save the shells and shims either (0.75–0.87 GB at 500 panes).
Stalled clients: a client that stops reading cost about 28 MB in one
burst, and up to about 64 MB per client from the 1,024-frame queue. Cap the
queue in bytes, not frames.
CI benchmark: S9's bench.py at 50 idle panes (3.07–3.11 MB per pane
over four runs) is the regression test M9 wants. Leave headroom.
Terminal parking.
- After 60s with no PTY reads, write the VT state as a GHOSTSNP checkpoint,
  using the M2 path and S5's format, and free the engine.
- Typing doesn't unpark it.
- Attaching streams the parked snapshot from disk.
- Parked state is encrypted with the key decided for M4c.
PTY parking. Idle or unwatched PTYs move off their own read tasks onto
one shared epoll task.
Client buffer parking. Free an idle client's per-pane buffers.
Done when:
- 500 idle shells on geek cost under 1MB each in daemon RSS;
- attaching to a parked pane draws it in under 50ms;
- the benchmark guards against regressions in CI.

**Depends on:** #3 **Trigger:** Panes with real history still blow the memory budget after step 1 (rerun the S9 benchmark). Step 1 (cheap memory wins) is split out as a separate issue and comes first. _From PLAN.md, section "M9"._ --- Superlogical's server numbers are about 400KB per terminal against 5MB for tmux, and unparking takes about 200µs. - **Trigger:** any one of these: - geek's daemon holds more than about 50 panes; - an agent fleet runs; - RSS per idle pane is measured above 2MB. Measure first. The milestone starts with a benchmark (RSS per pane at empty, full screen and 10k scrollback; per attached client) committed as a test. - **Step 2: parking, only if panes with real history still blow the budget.** S9 found nothing that argues for PTY or client-buffer parking. Parking can't save the shells and shims either (0.75–0.87 GB at 500 panes). - **Stalled clients:** a client that stops reading cost about 28 MB in one burst, and up to about 64 MB per client from the 1,024-frame queue. Cap the queue in bytes, not frames. - **CI benchmark:** S9's `bench.py` at 50 idle panes (3.07–3.11 MB per pane over four runs) is the regression test M9 wants. Leave headroom. - **Terminal parking.** - After 60s with no PTY reads, write the VT state as a GHOSTSNP checkpoint, using the M2 path and S5's format, and free the engine. - Typing doesn't unpark it. - Attaching streams the parked snapshot from disk. - Parked state is encrypted with the key decided for M4c. - **PTY parking.** Idle or unwatched PTYs move off their own read tasks onto one shared epoll task. - **Client buffer parking.** Free an idle client's per-pane buffers. - **Done when:** - 500 idle shells on geek cost under 1MB each in daemon RSS; - attaching to a parked pane draws it in under 50ms; - the benchmark guards against regressions in CI.

jhgaylor added the

milestone

gated

labels

2026-10-02 00:16:05 +00:00

jhgaylor added a new dependency

2026-10-02 00:16:06 +00:00

#3 M9 step 1: memory cheap wins (S9)

jhgaylor referenced this issue

2026-10-02 00:16:06 +00:00

M9 step 1: memory cheap wins (S9) #3