[franz1981]

HandoffHttpServer opens one Apache HttpClient 5.6.1 connection to the mock per client connection (~10K at 10K connections). HttpClient's default SocketConfig enables TCP keepalive (5 s idle, 5 s interval, 3 probes). At 50K req/s most of these connections are idle, and after any load-generator stall (e.g. a wrk2 GC pause) thousands of them go idle at the same moment, so their probes fire as one batch: 5.6-8K probes within 100 ms, every ~5 s.

On loopback every packet is queued on the sending CPU's backlog (net.core.netdev_max_backlog, 1000 by default), so each batch overflows it:

• 2-4.6K probes dropped per round (kfree_skb reason CPU_BACKLOG, matching the /proc/net/softnet_stat dropped deltas);
• connections whose probes are dropped three times in a row are aborted by the server's kernel (RST reason TCP_KEEPALIVE_TIMEOUT, TcpExtTCPAbortOnTimeout); the mock logs Connection reset by peer;
• the same backlog also drops server responses to the load generator, which then wait for a retransmit (>= 200 ms).

Unchanged server, 4 carriers, 10K connections, fixed 50K req/s: keepalive aborts in 3 of 4 runs (149, 1102, 517, 0); per-second p99 in the probe-batch seconds +15-20 ms.

This disables TCP keepalive on those connections; on loopback it has no purpose, and HttpClient still validates idle connections before reuse. With it, 2 runs: 0 aborts, 0 backlog drops, 0 resets.