• AutoCannon's -c option sets total open connections, -d sets duration, -p sets pipelining, and -W sets worker threads for the load generator. 20:39
  • Worker threads matter when the load generator machine has many cores, because the test client must also generate enough traffic. 20:39
  • With -c 6, -p 2, and -W 2, each worker opens three connections and each connection pipelines two requests. 21:00 22:07
  • Concurrent in-flight requests are computed as connections times pipelining, so 6 * 2 = 12 requests in flight. 22:47
  • The main metric the uploader watches is average requests per second, while the table also reports percentiles and min/max samples. 23:08 23:32
  • The local /simple run averaged around 18,000 requests per second, with worst samples around 16,000 and best samples around 18,000. 23:32 23:49

Benchmark reading checklist

  • Check average requests per second first. 23:08
  • Check whether errors occurred, because a high rate with failures is not a working result. 20:04
  • Check bytes transferred, because a small JSON response and a 30 KB response stress very different parts of the system. 20:04
  • Check the tester's CPU too, since a weak load generator can hide the server's real capacity. 20:39