This design is easier to misuse, because it allows the caller to modify
the contents of the slice after queueing it, but it avoids an extra
allocation + memmove per incoming packet.
Before:
$ go test -bench='Benchmark(QueueIncoming|WriteTo)' -benchtime=2s -benchmem
BenchmarkQueueIncoming-4 7001494 342.4 ns/op 1024 B/op 2 allocs/op
BenchmarkWriteTo-4 3777459 627 ns/op 1024 B/op 2 allocs/op
After:
$ go test -bench=BenchmarkWriteTo -benchtime 2s -benchmem
BenchmarkQueueIncoming-4 13361600 170.1 ns/op 512 B/op 1 allocs/op
BenchmarkWriteTo-4 6702324 373 ns/op 512 B/op 1 allocs/op
Despite the benchmark results, the change in QueueIncoming turns out not
to have an effect in practice. It appears that the compiler had already
been optimizing out the allocation and copy in QueueIncoming.
https://gitlab.torproject.org/tpo/anti-censorship/pluggable-transports/snowflake/-/issues/40187
The WriteTo change, on the other hand, in practice reduces the frequency
of garbage collection.
https://gitlab.torproject.org/tpo/anti-censorship/pluggable-transports/snowflake/-/issues/40199
Rather than use defer. It is only a tiny amount faster, but this
function is frequently called.
Before:
$ go test -bench=BenchmarkSendQueue -benchtime=2s
BenchmarkSendQueue-4 15901834 151 ns/op
After:
$ go test -bench=BenchmarkSendQueue -benchtime=2s
BenchmarkSendQueue-4 15859948 147 ns/op
https://gitlab.torproject.org/tpo/anti-censorship/pluggable-transports/snowflake/-/issues/40177
I had thought to set a buffer size of 2048, half the websocket package
default of 4096. But it turns out when you don't set a buffer size, the
websocket package reuses the HTTP server's read/write buffers, which
empirically already have a size of 2048.
$ go test -bench=BenchmarkUpgradeBufferSize -benchmem -benchtime=5s
BenchmarkUpgradeBufferSize/0-4 25669 234566 ns/op 32604 B/op 113 allocs/op
BenchmarkUpgradeBufferSize/128-4 24739 238283 ns/op 24325 B/op 117 allocs/op
BenchmarkUpgradeBufferSize/1024-4 25352 238885 ns/op 28087 B/op 116 allocs/op
BenchmarkUpgradeBufferSize/2048-4 22660 234890 ns/op 32444 B/op 116 allocs/op
BenchmarkUpgradeBufferSize/4096-4 25668 232591 ns/op 41672 B/op 116 allocs/op
BenchmarkUpgradeBufferSize/8192-4 24908 240755 ns/op 59103 B/op 116 allocs/op
Otherwise the buffers are re-allocated on every iteration, which is a
surprise to me. I thought the compiler would do this transformation
itself.
Now there is just one allocation per client←server read (one
messageReader) and two allocations per server←client read (one
messageReader and one messageWriter).
$ go test -bench=BenchmarkReadWrite -benchmem -benchtime=5s
BenchmarkReadWrite/c←s_150-4 481054 12849 ns/op 11.67 MB/s 8 B/op 1 allocs/op
BenchmarkReadWrite/s←c_150-4 421809 14095 ns/op 10.64 MB/s 56 B/op 2 allocs/op
BenchmarkReadWrite/c←s_3000-4 208564 28003 ns/op 107.13 MB/s 16 B/op 2 allocs/op
BenchmarkReadWrite/s←c_3000-4 186320 30576 ns/op 98.12 MB/s 112 B/op 4 allocs/op
In the client←server direction, this hits a fast path that avoids
allocating a messageWriter.
https://github.com/gorilla/websocket/blob/v1.5.0/conn.go#L760
Cuts the number of allocations in half in the client←server direction:
$ go test -bench=BenchmarkReadWrite -benchmem -benchtime=5s
BenchmarkReadWrite/c←s_150-4 597511 13358 ns/op 11.23 MB/s 33709 B/op 2 allocs/op
BenchmarkReadWrite/s←c_150-4 474176 13756 ns/op 10.90 MB/s 34968 B/op 4 allocs/op
BenchmarkReadWrite/c←s_3000-4 156488 36290 ns/op 82.67 MB/s 68673 B/op 5 allocs/op
BenchmarkReadWrite/s←c_3000-4 190897 34719 ns/op 86.41 MB/s 69730 B/op 8 allocs/op
This continues to asserts the known version while decoding. The client
will only ever generate the latest version while encoding and if the
response needs to change, the impetus will be a new feature, set in the
deserialized request, which can be used as a distinguisher.
