IORingGroup 1.0.8 allows more than one send in flight per socket. RIO
reports send completion on acknowledgement rather than on copy, so with a
single send in flight a connection was capped at one send per round trip
whenever it had data queued. On a 50ms-RTT shard that showed up as in-game
latency of 100-145ms whenever any data was flowing, against a 50ms idle
baseline; as little as 6 bytes of queued data was enough to hold the gate
shut. Local testing could not surface it: the same measurement at loopback
RTT is microseconds.
Adds two restart-time settings:
network.maxOutstandingSends (default 32) - sends in flight per
connection. Honoured by RIO only; other backends complete sends on copy
and report 1. Costs a request-queue and completion-queue slot per send,
not another buffer, since every outstanding send addresses a different
range of the same registered buffer.
network.sendBufferSize (default 256KB) - per-connection send buffer,
coerced to a power of two of at least the platform allocation
granularity. This is the lever for "send buffer exhausted" disconnects
and the per-connection memory ceiling.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
80aff11a46