net: use splice(2) on Linux when reading from UnixConn, rework splice tests
Rework the splice tests and benchmarks. Move the reading and writing of
the spliced connections to child processes so that the I/O is not part
of benchmarks or profiles.
Enable the use of splice(2) when reading from a unix connection and
writing to a TCP connection. The updated benchmarks show a performance
gain when using splice(2) to copy large chunks of data that the original
benchmark did not capture.
name old time/op new time/op delta
Splice/tcp-to-tcp/1024-8 5.01µs ± 2% 5.08µs ± 3% ~ (p=0.068 n=8+10)
Splice/tcp-to-tcp/2048-8 4.76µs ± 5% 4.65µs ± 3% -2.36% (p=0.015 n=9+8)
Splice/tcp-to-tcp/4096-8 4.91µs ± 2% 4.98µs ± 5% ~ (p=0.315 n=9+10)
Splice/tcp-to-tcp/8192-8 5.50µs ± 4% 5.44µs ± 3% ~ (p=0.758 n=7+9)
Splice/tcp-to-tcp/16384-8 7.65µs ± 7% 6.53µs ± 3% -14.65% (p=0.000 n=10+9)
Splice/tcp-to-tcp/32768-8 15.3µs ± 7% 8.5µs ± 5% -44.21% (p=0.000 n=10+10)
Splice/tcp-to-tcp/65536-8 30.0µs ± 6% 15.7µs ± 1% -47.58% (p=0.000 n=10+8)
Splice/tcp-to-tcp/131072-8 59.2µs ± 2% 27.4µs ± 5% -53.75% (p=0.000 n=9+9)
Splice/tcp-to-tcp/262144-8 121µs ± 4% 54µs ±19% -55.56% (p=0.000 n=9+10)
Splice/tcp-to-tcp/524288-8 247µs ± 6% 108µs ±12% -56.34% (p=0.000 n=10+10)
Splice/tcp-to-tcp/
1048576-8 490µs ± 4% 199µs ±12% -59.31% (p=0.000 n=8+10)
Splice/unix-to-tcp/1024-8 1.20µs ± 2% 1.35µs ± 7% +12.47% (p=0.000 n=10+10)
Splice/unix-to-tcp/2048-8 1.33µs ±12% 1.57µs ± 4% +17.85% (p=0.000 n=9+10)
Splice/unix-to-tcp/4096-8 2.24µs ± 4% 1.67µs ± 4% -25.14% (p=0.000 n=9+10)
Splice/unix-to-tcp/8192-8 4.59µs ± 8% 2.20µs ±10% -52.01% (p=0.000 n=10+10)
Splice/unix-to-tcp/16384-8 8.46µs ±13% 3.48µs ± 6% -58.91% (p=0.000 n=10+10)
Splice/unix-to-tcp/32768-8 18.5µs ± 9% 6.1µs ± 9% -66.99% (p=0.000 n=10+10)
Splice/unix-to-tcp/65536-8 35.9µs ± 7% 13.5µs ± 6% -62.40% (p=0.000 n=10+9)
Splice/unix-to-tcp/131072-8 79.4µs ± 6% 25.7µs ± 4% -67.62% (p=0.000 n=10+9)
Splice/unix-to-tcp/262144-8 157µs ± 4% 54µs ± 8% -65.63% (p=0.000 n=10+10)
Splice/unix-to-tcp/524288-8 311µs ± 3% 107µs ± 8% -65.74% (p=0.000 n=10+10)
Splice/unix-to-tcp/
1048576-8 643µs ± 4% 185µs ±32% -71.21% (p=0.000 n=10+10)
name old speed new speed delta
Splice/tcp-to-tcp/1024-8 204MB/s ± 2% 202MB/s ± 3% ~ (p=0.068 n=8+10)
Splice/tcp-to-tcp/2048-8 430MB/s ± 5% 441MB/s ± 3% +2.39% (p=0.014 n=9+8)
Splice/tcp-to-tcp/4096-8 833MB/s ± 2% 823MB/s ± 5% ~ (p=0.315 n=9+10)
Splice/tcp-to-tcp/8192-8 1.49GB/s ± 4% 1.51GB/s ± 3% ~ (p=0.758 n=7+9)
Splice/tcp-to-tcp/16384-8 2.14GB/s ± 7% 2.51GB/s ± 3% +17.03% (p=0.000 n=10+9)
Splice/tcp-to-tcp/32768-8 2.15GB/s ± 7% 3.85GB/s ± 5% +79.11% (p=0.000 n=10+10)
Splice/tcp-to-tcp/65536-8 2.19GB/s ± 5% 4.17GB/s ± 1% +90.65% (p=0.000 n=10+8)
Splice/tcp-to-tcp/131072-8 2.22GB/s ± 2% 4.79GB/s ± 4% +116.26% (p=0.000 n=9+9)
Splice/tcp-to-tcp/262144-8 2.17GB/s ± 4% 4.93GB/s ±17% +127.25% (p=0.000 n=9+10)
Splice/tcp-to-tcp/524288-8 2.13GB/s ± 6% 4.89GB/s ±13% +130.15% (p=0.000 n=10+10)
Splice/tcp-to-tcp/
1048576-8 2.09GB/s ±10% 5.29GB/s ±11% +153.36% (p=0.000 n=10+10)
Splice/unix-to-tcp/1024-8 850MB/s ± 2% 757MB/s ± 7% -10.94% (p=0.000 n=10+10)
Splice/unix-to-tcp/2048-8 1.54GB/s ±11% 1.31GB/s ± 3% -15.32% (p=0.000 n=9+10)
Splice/unix-to-tcp/4096-8 1.83GB/s ± 4% 2.45GB/s ± 4% +33.59% (p=0.000 n=9+10)
Splice/unix-to-tcp/8192-8 1.79GB/s ± 9% 3.73GB/s ± 9% +108.05% (p=0.000 n=10+10)
Splice/unix-to-tcp/16384-8 1.95GB/s ±13% 4.68GB/s ± 3% +139.80% (p=0.000 n=10+9)
Splice/unix-to-tcp/32768-8 1.78GB/s ± 9% 5.38GB/s ±10% +202.71% (p=0.000 n=10+10)
Splice/unix-to-tcp/65536-8 1.83GB/s ± 8% 4.85GB/s ± 6% +165.70% (p=0.000 n=10+9)
Splice/unix-to-tcp/131072-8 1.65GB/s ± 6% 5.10GB/s ± 4% +208.77% (p=0.000 n=10+9)
Splice/unix-to-tcp/262144-8 1.67GB/s ± 4% 4.87GB/s ± 7% +191.19% (p=0.000 n=10+10)
Splice/unix-to-tcp/524288-8 1.69GB/s ± 3% 4.93GB/s ± 7% +192.38% (p=0.000 n=10+10)
Splice/unix-to-tcp/
1048576-8 1.63GB/s ± 3% 5.60GB/s ±44% +243.26% (p=0.000 n=10+9)
Change-Id: I1eae4c3459c918558c70fc42283db22ff7e0442c
Reviewed-on: https://go-review.googlesource.com/113997
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>