Evgeny
|
61fee27478
|
congestion: burst-based BW measurement + adaptive inflight (slow start/cong avoidance/loss response)
New features:
- get_time_us() in lib/u_async.c/h for µs-precision timing
- Burst measurement (ETCP_SECTION_MEAS_TS 0x07, MEAS_RESP 0x08, FILLER 0x09)
- 12 MTU packets sent at line rate, receiver measures inter-packet gaps
- Burst triggered when link saturated + min 500ms interval
- Result: measured BW -> bandwidth for shaper, burst_target_bdp for window cap
- 3-phase adaptive inflight in link_stats_timer_cb:
- Slow start: windows doubles each RTT, exits on loss/RTT spike/threshold
- Congestion avoidance: additive increase, proportional RTT shrink, multiplicative loss shrink
- Burst target cap: inflight_lim <= burst_target_bdp * 1.1
- Dynamic optimal_inflight = sum(link->inflight_lim_bytes), recalculated on change
- Adaptive win_timebase = rtt/2 (was hardcoded 10ms)
Fixes:
- dummynet.c: shaper formula fixed (80000 -> 80)
- dummynet.c: queue_data_put call fixed (3 args -> 2)
- etcp.c: SIZE ERROR format string (missing log_name argument)
- etcp_request_pkt: FILLER leaves room for PAYLOAD overhead
Tests:
- test_etcp_congestion.c: single-threaded, 5-phase BW/delay/loss test
- sender + receiver + dummynet in one uasync
- CSV metrics output every 100ms
- All 29 existing tests pass
|
6 months ago |
