Throughput scaling
ops/s vs server cores — the headline curves (cores filter not applied here)
Latency percentiles
client-side p50 / p99 / p99.9 at selected core count (log ms scale)
Per-core efficiency
ops per server core vs cores — is it magic, or just using cores well?
Pipeline effect
no pipelining vs depth 16 at selected core count (pipeline filter not applied here)
Value-size effect
100 B vs 1 KB at selected core count — gaps compress when the network takes over (value-size filter not applied here)
Run health
minimum hit rate and connection errors behind every plotted point
