You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Direct transfers are faster on fast links, for less CPU. Between
two 8-vCPU machines in one cloud zone (0.2 ms, 15.6 Gbit/s on one TCP
stream), a 10 GiB direct transfer ran at 2.71 Gbit/s (median of 3;
v0.1.10 on the same machines, run in turn with it: 1.54) and 100 GiB
at 2.55, BLAKE3 equal. The sender spent 6.8 CPU s per GiB (v0.1.10:
10.3) and the receiver 8.4 (10.1). The relayed path is unchanged:
3.26 Gbit/s on both versions (docs/scale.md test 7).
From the path switch, whose send and receive buffers are now rings: an
acknowledgement no longer moves every unacknowledged byte to the front
of its buffer (pathswitch v0.1.4); and from the direct path's QUIC
library: on Linux the receiver reads datagrams in batches (UDP GRO), a
send batch holds up to 44 packets (was 14), and a connection queues up
to 4096 received packets (quiclink v0.7.4).
A closed session no longer leaves a write hanging: a direct
session's write blocked on the peer's window returns an error once the
session is closed; before, it waited (quiclink v0.7.4).
CPU benchmarks: make bench-line measures one direct connection
on loopback (2 GiB, 6 runs) with a CPU profile; make bench-ab runs
the direct and relayed CPU benchmarks (BenchmarkDirectCPU, BenchmarkRelayedCPU) of builds that differ only in their
dependencies, in turn.
The protocol and the wire are unchanged from v0.1.10.
Built on link v0.13.3, pathswitch v0.1.4, quiclink v0.7.4, relay client
v0.24.4, relay proto v0.22.1 and relay-core v0.19.4.