Coordinated omission in latency benchmarks

[buybackoff]

An eye-opening talk by Gil Tene. We already use HDR Histogram but need to make sure GC doesn't pauses result collection. https://www.youtube.com/watch?v=YpBwt-rO07o

One idea is to use Spreads.IPC/Aeron to write and read values into a IPC channel and then process them on Spreads side. Then we will measure correct response time. Run an elevated GC-free proc with a precise timer (e.g. volatile incrementing long counter or performance counter) in a separate thread. That would be a realistic test for data arriving -> processed -> returned.

Good news is that most focus so far was on throughput, latency percentiles were just to understand orders of magnitude.

Upd. Same talk, another version: https://www.infoq.com/presentations/latency-response-time