mirror of
https://github.com/tailscale/tailscale.git
synced 2026-11-20 08:41:19 -08:00
Recruiting fewer crypt workers and going inline for slow peers (change 17 and 18) cost one TCP stream in the QEMU lab (4 vCPUs) a third of its throughput, 11.5 to 7-8 Gbps. EC2 didn't show it, as both data planes were at its per-flow limit there. The sender, held to a few GSO packets in flight by TCP Small Queues, never had recruit_factor's worth waiting, so it got one encrypt worker. Now the first recruit_free (default 2) draining workers are recruited at the old backlog, and only past that does recruit_factor apply. And tswg_tx_slow reset its count every 10ms window, so the first tx_inline_mbps' worth of every window was encrypted inline however fast the peer was sending. Now, as on receive, a peer that went over the rate counts as fast for 100ms. Also start the per-peer until-jiffies at the peer's creation rather than 0, which is in the future for the first 5 minutes after boot. One stream in the lab: 7.1 to 11.3 Gbps, 1.23x wireguard-go's, at 0.66x / 0.80x its sender / receiver CPU per GB. Updates #426 Signed-off-by: Brad Fitzpatrick <[email protected]> Change-Id: I59c20b2d005ff3125cd4055fc1d607db3ca18ab1