You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
--rate no longer sends in per-thread bursts. Each thread's connections were scheduled inside one contiguous slice of the interval (1/T of it, T being the thread count), so one thread at a time carried the whole fleet's rate. Once that rate was more than a thread can send, sends ran late, the responses queued behind them, and the latency figures measured gcannon rather than the server: p50 read ~150 ms at 1M connections, -t 64 and 250K req/s, against a server whose own p50 was ~70 µs. Each thread now spreads its connections over the whole interval, interleaved with the other threads. At 62,500 connections, 250K req/s and 16 threads, p50 went from 14.2 ms to 18 µs and late sends from 2.1M of 5.0M to none. Rates a single thread can carry alone measure as before. (#30)