Investiagion: relay response time grows too quickly in relation to the number of connections

[kamirr]

1.2s response time for barely 50k connections and relatively low load (25k req/s). See #191 (comment) for more data.

[kamirr]

Investigation conculuded: the main reason for the ~linear latency is the fact that the time required to process a request by the relay server (from dispatch to socket send) is roughly proportional to sqrt(# nodes).

spreadsheet with my measurements

Possible approaches for fixing the issue:

• Measuring main hashmap access times.
• Checking TLB/cache misses via perf.
• Re-checking all locks (I've found no contentions myself).
• Other?