Periodic 2MB pushes over one persistent TCP connection

A Linux server pushes a 2MB risk snapshot to a client over one persistent TCP connection on a 40ms-RTT path. In a benchmark that sends snapshots back-to-back, every snapshot after the first completes in roughly one RTT. In production the snapshots go out every 5 seconds with the connection idle in between, and every one of them takes about seven RTTs, as if the connection were brand new. There is no loss, no reconnection, and receive windows are ample. Kernel defaults are untouched. What explains the difference?

36 of 45 in Networking