mirror of
https://github.com/tailscale/tailscale.git
synced 2026-11-20 16:51:18 -08:00
RTT under a TCP load had grown to about 11ms at p99 (wireguard-go: 3.7). Instrumenting the queues found three causes: Backpressure. Upstream encrypts in place, so a packet stays charged to its socket (for TCP, counted by its small queues, TSQ) until the NIC has sent it, and TCP is held to the NIC's pace. Our encrypted packets are new (and earlier ones were orphaned), so TCP could fill the NIC's qdisc as fast as we encrypt: on a slow uplink with a FIFO qdisc, that would be massive bufferbloat. Now each encrypted packet takes over its packets' socket ownership and charge (tswg_move_owner), for destructors that just uncharge truesize (sock_wfree, and TCP's), settled to its real size (segments' truesize is about twice their length). TSQ then allowed TCP only about 1ms in flight, which cost 15-20% of throughput, so sockets sending over tswg get sk_pacing_shift 9 (about 2ms), as Wi-Fi drivers set 8 for their aggregation. The tx worker. Upstream pins a peer's tx worker to one CPU by peer ID. When that's the CPU taking the NIC's interrupts, softirq work starves it, and encrypted packets waited up to 20ms to be sent. Queue it on the CPU that finished encrypting instead (also saving an IPI). Queue length. The peer queue's limit counted entries, but a GRO entry is up to 64 packets; count packets, with a limit of 2048. Medians of interleaved runs in the lean harness, tswg vs. wireguard-go: TCP 1 stream 10.1 vs. 9.7 Gbps, 4 streams 12.3 vs. 10.8; RTT under load p50 1.7 vs. 1.7ms. p99 under load is still about 10ms on this testbed: at full load, the receiving VM's workers sometimes don't run for tens of milliseconds. That's next to investigate. Updates #426 Signed-off-by: Brad Fitzpatrick <[email protected]> Change-Id: Ief80aac9a1e443727c89b990c86c7c8f71690954