Benchmarks

All runs: Node 22, headless, eth_http firmware (DHCP + TCP + HTTP round trip), Linux. One "round" = one full DHCP → TCP connect → HTTP body → FIN cycle of the firmware. MIPS = emulated instructions / wall time.

Current headline numbers (2026-08-09, post wedge-fix)

Run	Budget	Rounds	TCP fail	Time	MIPS	rounds/s
100M soak, no restart	100M inst	604	0	47.9 s	~2.1	12.6
200M soak (2026-08-08)	200M inst	1012	0	870.8 s	~1.0	1.16*
20M quick check (this doc)	20M inst	121	0	10.1 s	~1.97	12.0

* The 200M soak predates the throughput work and used per-round gateway restarts (~0.7 s per restart), which is why rounds/s is much lower than the 100M run above.

Typical run (20M instructions):

real  0m10.138s   user 0m10.569s
121 rounds, 0 TCP fail, ~2 MIPS, ~12 rounds/s

Throughput tuning history (2026-08-08)

20M-instruction --connect runs (eth_http, fresh gateway):

Config	Steps	Rounds	Wall	MIPS	rounds/s
Baseline (old cli.mjs)	~2600	101	~20.0 s	~1.00	5.05
TICK_EVERY=1000 (no poll split)	1939	67	11.05 s	1.81	6.06
TICK_EVERY=5000 (no poll split)	2473	95	13.9–14.14 s	1.44	6.8
POLL_EVERY=1000 + TICK_EVERY=5000	3061	118	15.8–16.0 s	1.25	7.46

Key wins: batching WASM calls (tick_n(5000) instead of per-instruction tick), splitting the cheap all-JS poll (every instruction) from the expensive WASM poll (every 1000), and removing the per-round gateway restart (default since 2026-08-09).

Soak results

100M-instruction soak (2026-08-09, final config)

• 604 TCP connected, 0 TCP fail, 0 fl=18 stale frames, 0 timeouts
• 47.9 s wall, ~12.6 rounds/s
• No per-round gateway restart (gVisor opens a fresh session per connection; the firmware ignores stale fl=18 frames)

200M-instruction soak (2026-08-08)

• 1012 TCP connections, 0 TCP fail, 0 SYN-ACK timeouts in 870.8 s
• RSS grew linearly 153→214 MB (~0.06 MB/round; not runaway, but not a plateau either)

Connect-mode soaks (external gateway process, --connect)

• Two consecutive 100M runs against the same gateway process: 1055 TCP rounds total (526 + 529 connected), 0 fail, 0 SYN-ACK timeouts, 525 RESET messages. RSS ~193 MB at end of each run — the abandoned-stack leak resets with the room, doesn't accumulate.
• Two consecutive 15M runs into 74 stale sessions: 74/77 TCP connected, 0 fail, 0 timeouts.

Restart mode (legacy, GW_RESTART=1)

• ~1.27 rounds/s (each kill+spawn+reconnect costs ~0.7 s). Only for pathological stale-session cases.

Environment / reproducibility

• Deterministic: the model clock is instruction-count based, so a fixed instruction budget is reproducible across machines (wall time varies with CPU).
• The gateway must be reachable: check ss -tlnp | grep <port> before trusting a 0-round result — a dead gateway burns the whole budget in the DHCP wait (documented in AGENTS.md §10).
• After a long soak, restart the gateway (kill <pid>, relaunch openhw-gw -port 5099) if the next run shows TCP fl=10 + recv-wait stall — that's stale-session pollution, not a code regression.

How to measure

# CLI throughput (gateway mode)
cd stm32-periph-wasm/pkg
time node cli.mjs ../eth_http/eth_http.bin 20000000 \
  --gateway --config=../../eth_http/config.yaml
grep -c "TCP connected" <outfile>   # rounds
grep -c "TCP fail" <outfile>        # must be 0

# Soak with stats
SOAK_STATS=1 node cli.mjs ../eth_http/eth_http.bin 200000000 \
  --gateway --config=../../eth_http/config.yaml

# Browser-mode throughput (runs in rAF steps; ~a round per few frames)
node site/test_flow.mjs

Tunables

Env	Default	Effect
MAX_BATCH	20000	instructions per emu_start; must stay < ~40k (Unicorn WASM wedge)
TICK_EVERY	5000	instruction interval for tick_n() + watchdog + interrupt checks
POLL_EVERY	1000	instruction interval for DMA/ETH poll checks
GW_RESTART	0	1 = restart gateway per round (legacy, ~10x slower)
SOAK_STATS	—	1 = emit soak statistics at the end
DBG_TX/DBG_RX	—	1 = trace TX/RX frames
RX_HEX	—	1 = dump first 64 B of each injected RX frame
DBG_FLAG/DBG_IRQF/DBG_PC/DBG_GW/DBG_DMA	—	diagnostic traces (see cli.mjs)