You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
tailscale/wgengine/bench/bench.go

410 lines
8.7 KiB
Go

// Copyright (c) Tailscale Inc & AUTHORS
// SPDX-License-Identifier: BSD-3-Clause
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
// Create two wgengine instances and pass data through them, measuring
// throughput, latency, and packet loss.
package main
import (
"bufio"
"io"
"log"
"net"
"net/http"
"net/http/pprof"
"net/netip"
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
"os"
"strconv"
"sync"
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
"time"
"tailscale.com/types/logger"
)
const PayloadSize = 1000
const ICMPMinSize = 24
var Addr1 = netip.MustParsePrefix("100.64.1.1/32")
var Addr2 = netip.MustParsePrefix("100.64.1.2/32")
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
func main() {
var logf logger.Logf = log.Printf
log.SetFlags(0)
debugMux := newDebugMux()
go runDebugServer(debugMux, "0.0.0.0:8999")
mode, err := strconv.Atoi(os.Args[1])
if err != nil {
log.Fatalf("%q: %v", os.Args[1], err)
}
traf := NewTrafficGen(nil)
// Sample test results below are using GOMAXPROCS=2 (for some
// tests, including wireguard-go, higher GOMAXPROCS goes slower)
// on apenwarr's old Linux box:
// Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz
// My 2019 Mac Mini is about 20% faster on most tests.
switch mode {
// tx=8786325 rx=8786326 (0 = 0.00% loss) (70768.7 Mbits/sec)
case 1:
setupTrivialNoAllocTest(logf, traf)
// tx=6476293 rx=6476293 (0 = 0.00% loss) (52249.7 Mbits/sec)
case 2:
setupTrivialTest(logf, traf)
// tx=1957974 rx=1958379 (0 = 0.00% loss) (15939.8 Mbits/sec)
case 11:
setupBlockingChannelTest(logf, traf)
// tx=728621 rx=701825 (26620 = 3.65% loss) (5525.2 Mbits/sec)
// (much faster on macOS??)
case 12:
setupNonblockingChannelTest(logf, traf)
// tx=1024260 rx=941098 (83334 = 8.14% loss) (7516.6 Mbits/sec)
// (much faster on macOS??)
case 13:
setupDoubleChannelTest(logf, traf)
// tx=265468 rx=263189 (2279 = 0.86% loss) (2162.0 Mbits/sec)
case 21:
setupUDPTest(logf, traf)
// tx=1493580 rx=1493580 (0 = 0.00% loss) (12210.4 Mbits/sec)
case 31:
setupBatchTCPTest(logf, traf)
// tx=134236 rx=133166 (1070 = 0.80% loss) (1088.9 Mbits/sec)
case 101:
setupWGTest(nil, logf, traf, Addr1, Addr2)
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
default:
log.Fatalf("provide a valid test number (0..n)")
}
logf("initialized ok.")
traf.Start(Addr1.Addr(), Addr2.Addr(), PayloadSize+ICMPMinSize, 0)
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
var cur, prev Snapshot
var pps int64
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
i := 0
for {
i += 1
time.Sleep(10 * time.Millisecond)
if (i % 100) == 0 {
prev = cur
cur = traf.Snap()
d := cur.Sub(prev)
if prev.WhenNsec == 0 {
logf("tx=%-6d rx=%-6d", d.TxPackets, d.RxPackets)
} else {
logf("%v @%7d pkt/s", d, pps)
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
}
}
pps = traf.Adjust()
}
}
func newDebugMux() *http.ServeMux {
mux := http.NewServeMux()
mux.HandleFunc("/debug/pprof/", pprof.Index)
mux.HandleFunc("/debug/pprof/cmdline", pprof.Cmdline)
mux.HandleFunc("/debug/pprof/profile", pprof.Profile)
mux.HandleFunc("/debug/pprof/symbol", pprof.Symbol)
mux.HandleFunc("/debug/pprof/trace", pprof.Trace)
return mux
}
func runDebugServer(mux *http.ServeMux, addr string) {
srv := &http.Server{
Addr: addr,
Handler: mux,
}
if err := srv.ListenAndServe(); err != nil {
log.Fatal(err)
}
}
// The absolute minimal test of the traffic generator: have it fill
// a packet buffer, then absorb it again. Zero packet loss.
func setupTrivialNoAllocTest(logf logger.Logf, traf *TrafficGen) {
go func() {
b := make([]byte, 1600)
for {
n := traf.Generate(b, 16)
if n == 0 {
break
}
traf.GotPacket(b[0:n+16], 16)
}
}()
}
// Almost the same, but this time allocate a fresh buffer each time
// through the loop. Still zero packet loss. Runs about 2/3 as fast for me.
func setupTrivialTest(logf logger.Logf, traf *TrafficGen) {
go func() {
for {
b := make([]byte, 1600)
n := traf.Generate(b, 16)
if n == 0 {
break
}
traf.GotPacket(b[0:n+16], 16)
}
}()
}
// Pass packets through a blocking channel between sender and receiver.
// Still zero packet loss since the sender stops when the channel is full.
// Max speed depends on channel length (I'm not sure why).
func setupBlockingChannelTest(logf logger.Logf, traf *TrafficGen) {
ch := make(chan []byte, 1000)
go func() {
// transmitter
for {
b := make([]byte, 1600)
n := traf.Generate(b, 16)
if n == 0 {
close(ch)
break
}
ch <- b[0 : n+16]
}
}()
go func() {
// receiver
for b := range ch {
traf.GotPacket(b, 16)
}
}()
}
// Same as setupBlockingChannelTest, but now we drop packets whenever the
// channel is full. Max speed is about the same as the above test, but
// now with nonzero packet loss.
func setupNonblockingChannelTest(logf logger.Logf, traf *TrafficGen) {
ch := make(chan []byte, 1000)
go func() {
// transmitter
for {
b := make([]byte, 1600)
n := traf.Generate(b, 16)
if n == 0 {
close(ch)
break
}
select {
case ch <- b[0 : n+16]:
default:
}
}
}()
go func() {
// receiver
for b := range ch {
traf.GotPacket(b, 16)
}
}()
}
// Same as above, but at an intermediate blocking channel and goroutine
// to make things a little more like wireguard-go. Roughly 20% slower than
// the single-channel version.
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
func setupDoubleChannelTest(logf logger.Logf, traf *TrafficGen) {
ch := make(chan []byte, 1000)
ch2 := make(chan []byte, 1000)
go func() {
// transmitter
for {
b := make([]byte, 1600)
n := traf.Generate(b, 16)
if n == 0 {
close(ch)
break
}
select {
case ch <- b[0 : n+16]:
default:
}
}
}()
go func() {
// intermediary
for b := range ch {
ch2 <- b
}
close(ch2)
}()
go func() {
// receiver
for b := range ch2 {
traf.GotPacket(b, 16)
}
}()
}
// Instead of a channel, pass packets through a UDP socket.
func setupUDPTest(logf logger.Logf, traf *TrafficGen) {
la, err := net.ResolveUDPAddr("udp", ":0")
if err != nil {
log.Fatalf("resolve: %v", err)
}
s1, err := net.ListenUDP("udp", la)
if err != nil {
log.Fatalf("listen1: %v", err)
}
s2, err := net.ListenUDP("udp", la)
if err != nil {
log.Fatalf("listen2: %v", err)
}
a2 := s2.LocalAddr()
// On macOS (but not Linux), you can't transmit to 0.0.0.0:port,
// which is what returns from .LocalAddr() above. We have to
// force it to localhost instead.
a2.(*net.UDPAddr).IP = net.ParseIP("127.0.0.1")
s1.SetWriteBuffer(1024 * 1024)
s2.SetReadBuffer(1024 * 1024)
go func() {
// transmitter
b := make([]byte, 1600)
for {
n := traf.Generate(b, 16)
if n == 0 {
break
}
s1.WriteTo(b[16:n+16], a2)
}
}()
go func() {
// receiver
b := make([]byte, 1600)
for traf.Running() {
// Use ReadFrom instead of Read, to be more like
// how wireguard-go does it, even though we're not
// going to actually look at the address.
n, _, err := s2.ReadFrom(b)
if err != nil {
log.Fatalf("s2.Read: %v", err)
}
traf.GotPacket(b[:n], 0)
}
}()
}
// Instead of a channel, pass packets through a TCP socket.
// TCP is a single stream, so we can amortize one syscall across
// multiple packets. 10x amortization seems to make it go ~10x faster,
// as expected, getting us close to the speed of the channel tests above.
// There's also zero packet loss.
func setupBatchTCPTest(logf logger.Logf, traf *TrafficGen) {
sl, err := net.Listen("tcp", ":0")
if err != nil {
log.Fatalf("listen: %v", err)
}
var slCloseOnce sync.Once
slClose := func() {
slCloseOnce.Do(func() {
sl.Close()
})
}
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
s1, err := net.Dial("tcp", sl.Addr().String())
if err != nil {
log.Fatalf("dial: %v", err)
}
s2, err := sl.Accept()
if err != nil {
log.Fatalf("accept: %v", err)
}
s1.(*net.TCPConn).SetWriteBuffer(1024 * 1024)
s2.(*net.TCPConn).SetReadBuffer(1024 * 1024)
ch := make(chan int)
go func() {
// transmitter
defer slClose()
defer s1.Close()
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
bs1 := bufio.NewWriterSize(s1, 1024*1024)
b := make([]byte, 1600)
i := 0
for {
i += 1
n := traf.Generate(b, 16)
if n == 0 {
break
}
if i == 1 {
ch <- n
}
bs1.Write(b[16 : n+16])
// TODO: this is a pretty half-baked batching
// function, which we'd never want to employ in
// a real-life program.
//
// In real life, we'd probably want to flush
// immediately when there are no more packets to
// generate, and queue up only if we fall behind.
//
// In our case however, we just want to see the
// technical benefits of batching 10 syscalls
// into 1, so a fixed ratio makes more sense.
if (i % 10) == 0 {
bs1.Flush()
}
}
}()
go func() {
// receiver
defer slClose()
defer s2.Close()
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
bs2 := bufio.NewReaderSize(s2, 1024*1024)
// Find out the packet size (we happen to know they're
// all the same size)
packetSize := <-ch
b := make([]byte, packetSize)
for traf.Running() {
// TODO: can't use ReadFrom() here, which is
// unfair compared to UDP. (ReadFrom for UDP
// apparently allocates memory per packet, which
// this test does not.)
n, err := io.ReadFull(bs2, b)
if err != nil {
log.Fatalf("s2.Read: %v", err)
}
traf.GotPacket(b[:n], 0)
}
}()
}