TCP Optimization in Production

TCP tuning is not a search for the largest buffer or the fastest-looking kernel setting. A useful change removes a measured limit for a specific workload.

A short API response, a cross-region database replication stream, and a server accepting thousands of new connections per second stress different parts of TCP. The API may be dominated by connection setup and small writes. The replication stream may need several megabytes in flight. The busy listener may be limited by application accept rate rather than data transfer.

Production optimization therefore begins with three questions:

The last question matters because throughput, latency, memory use, CPU use, and fairness can move in different directions.