Achieving high throughput with Go TCP client-server - linux

Question

I'm going to develop a simple TCP client and server and I want to achieve high throughput (300000 Requests/Second) which is easy to reach with Cpp or C TCP client and server on a server hardware. I mean a server with 48 Cores and 64G Memory.
On my testbed, both client and server have 10G network interface card and I have receive-side-scaling at server side and transmit-packet-steering enabled at the client.
I configure the client to send 10 thousand requests per second. I just run multiple instances of Go go run client.go from a bash script to increase the throughput. However, in this way, Go is going to create lots of threads at the operating systems and a large number of threads results in high context switching cost, and I could not approach such throughputs. I suspected the number of Go instances I'm running from the command line. The code below is the code snippet for the client in the approach:
func Main(cmd_rate_int int, cmd_port string) {
//runtime.GOMAXPROCS(2) // set maximum number of processes to be used by this applications
//var rate float64 = float64(rate_int)
rate := float64(cmd_rate_int)
port = cmd_port
conn, err := net.Dial("tcp", port)
if err != nil {
fmt.Println("ERROR", err)
os.Exit(1)
}
var my_random_number float64 = nextTime(rate) * 1000000
var my_random_int int = int(my_random_number)
var int_message int64 = time.Now().UnixNano()
byte_message := make([]byte, 8)
go func(conn net.Conn) {
buf := make([]byte, 8)
for true {
_, err = io.ReadFull(conn, buf)
now := time.Now().UnixNano()
if err != nil {
return
}
last := int64(binary.LittleEndian.Uint64(buf))
fmt.Println((now - last) / 1000)
}
return
}(conn)
for true {
my_random_number = nextTime(rate) * 1000000
my_random_int = int(my_random_number)
time.Sleep(time.Microsecond * time.Duration(my_random_int))
int_message = time.Now().UnixNano()
binary.LittleEndian.PutUint64(byte_message, uint64(int_message))
conn.Write(byte_message)
}
}
So I try to run all my Go threads by calling go client() in the main so I do not run multiple instances in the Linux command line. I thought it may be a better idea. And it is really a better idea basically and the number of threads doesn't increase toward 700 or so in the operating system. But the throughput still is low and it seems it doesn't employ all capability of the underlying hardware. Actually, you may want to see the code I have run in the second approach:
func main() {
//runtime.GOMAXPROCS(2) // set maximum number of processes to be used by this applications
args := os.Args[1:]
rate_int, _ := strconv.Atoi(args[0])
client_size, _ := strconv.Atoi(args[1])
port := args[2]
i := 0
for i <= client_size {
go client.Main(rate_int, port)
i = i + 1
}
for true {
}
}
I was wondering what is the best practice for in order to reach high throughput? I have always heard that Go is lightweight and performant and pretty comparable with C/Cpp pthread. However, I think in terms of performance still C/Cpp is far far better than Go. I might do something really wrong on this issue, so I would be happy if anybody can help to achieve high throughput with Go.

Answer 1

this is a quick rework of the op code.
As the original source code is working, it does not provide a solution, however it illustrates bucket token usage, and few other small go tips.
It does re use similar default values as op source code.
It demonstrates you do not need two files / programs, to provide both client and server.
It demonstrates usage of flag package.
It shows how to parse unix nano timestamp appropriately using time.Unix(x,y)
It shows how to take advantage of io.Copy to write-what-you-read on the same net.Conn. Rather than manual writing.
Still, this is improper for production delivery.
package main
import (
"encoding/binary"
"flag"
"fmt"
"io"
"log"
"math"
"math/rand"
"net"
"os"
"sync/atomic"
"time"
"github.com/juju/ratelimit"
)
var total_rcv int64
func main() {
var cmd_rate_int float64
var cmd_port string
var client_size int
flag.Float64Var(&cmd_rate_int, "rate", 400000, "change rate of message reading")
flag.StringVar(&cmd_port, "port", ":9090", "port to listen")
flag.IntVar(&client_size, "size", 20, "number of clients")
flag.Parse()
t := flag.Arg(0)
if t == "server" {
server(cmd_port)
} else if t == "client" {
for i := 0; i < client_size; i++ {
go client(cmd_rate_int, cmd_port)
}
// <-make(chan bool) // infinite wait.
<-time.After(time.Second * 2)
fmt.Println("total exchanged", total_rcv)
} else if t == "client_ratelimit" {
bucket := ratelimit.NewBucketWithQuantum(time.Second, int64(cmd_rate_int), int64(cmd_rate_int))
for i := 0; i < client_size; i++ {
go clientRateLimite(bucket, cmd_port)
}
// <-make(chan bool) // infinite wait.
<-time.After(time.Second * 3)
fmt.Println("total exchanged", total_rcv)
}
}
func server(cmd_port string) {
ln, err := net.Listen("tcp", cmd_port)
if err != nil {
panic(err)
}
for {
conn, err := ln.Accept()
if err != nil {
panic(err)
}
go io.Copy(conn, conn)
}
}
func client(cmd_rate_int float64, cmd_port string) {
conn, err := net.Dial("tcp", cmd_port)
if err != nil {
log.Println("ERROR", err)
os.Exit(1)
}
defer conn.Close()
go func(conn net.Conn) {
buf := make([]byte, 8)
for {
_, err := io.ReadFull(conn, buf)
if err != nil {
break
}
// int_message := int64(binary.LittleEndian.Uint64(buf))
// t2 := time.Unix(0, int_message)
// fmt.Println("ROUDNTRIP", time.Now().Sub(t2))
atomic.AddInt64(&total_rcv, 1)
}
return
}(conn)
byte_message := make([]byte, 8)
for {
wait := time.Microsecond * time.Duration(nextTime(cmd_rate_int))
if wait > 0 {
time.Sleep(wait)
fmt.Println("WAIT", wait)
}
int_message := time.Now().UnixNano()
binary.LittleEndian.PutUint64(byte_message, uint64(int_message))
_, err := conn.Write(byte_message)
if err != nil {
log.Println("ERROR", err)
return
}
}
}
func clientRateLimite(bucket *ratelimit.Bucket, cmd_port string) {
conn, err := net.Dial("tcp", cmd_port)
if err != nil {
log.Println("ERROR", err)
os.Exit(1)
}
defer conn.Close()
go func(conn net.Conn) {
buf := make([]byte, 8)
for {
_, err := io.ReadFull(conn, buf)
if err != nil {
break
}
// int_message := int64(binary.LittleEndian.Uint64(buf))
// t2 := time.Unix(0, int_message)
// fmt.Println("ROUDNTRIP", time.Now().Sub(t2))
atomic.AddInt64(&total_rcv, 1)
}
return
}(conn)
byte_message := make([]byte, 8)
for {
bucket.Wait(1)
int_message := time.Now().UnixNano()
binary.LittleEndian.PutUint64(byte_message, uint64(int_message))
_, err := conn.Write(byte_message)
if err != nil {
log.Println("ERROR", err)
return
}
}
}
func nextTime(rate float64) float64 {
return -1 * math.Log(1.0-rand.Float64()) / rate
}

Answer 2

Edit This is a pretty bad answer. Check mh-cbon comments for the reasons.
I don't fully understand how you're trying to do so, but if I want to control the rate on Go, I usually do 2 nested for loops:
for ;; time.Sleep(time.Second) {
go func (){
for i:=0; i<rate; i++ {
go func (){
// Do whatever
}()
}
}()
}
I'm starting a goroutine inside each loop to:
on the outer loop, to ensure it's only 1 second between iterations
on the inner loop, to ensure I can start all the requests I want
Putting this on a problem like yours, it would look something like:
package main
import (
"net"
"os"
"time"
)
const (
rate = 100000
address = "localhost:8090"
)
func main() {
conn, err := net.Dial("tcp", address)
if err != nil {
os.Stderr.Write([]byte(err.Error() + "\n"))
os.Exit(1)
}
for ; err == nil; time.Sleep(time.Second) {
go func() {
for i := 0; i < rate; i++ {
go func(conn net.Conn) {
if _, err := conn.Write([]byte("01234567")); err != nil {
os.Stderr.Write([]byte("\nConnection closed: " + err.Error() + "\n"))
}
}(conn)
}
}()
}
}
To verify that this is actually sending the target request rate, you can have a test TCP listener like this:
package main
import (
"fmt"
"net"
"os"
"time"
)
const (
address = ":8090"
payloadSize = 8
)
func main() {
count := 0
b := make([]byte, payloadSize)
l, err := net.Listen("tcp", address)
if err != nil {
fmt.Fprintf(os.Stdout, "\nCan't listen to address %v: %v\n", address, err)
return
}
defer l.Close()
go func() {
for ; ; time.Sleep(time.Second) {
fmt.Fprintf(os.Stdout, "\rRate: %v/s ", count)
count = 0
}
}()
for {
conn, err := l.Accept()
if err != nil {
fmt.Fprintf(os.Stderr, "\nFailed to accept connection: %v\n", err)
}
for {
_, err := conn.Read(b)
if err != nil {
fmt.Fprintf(os.Stderr, "\nConnection closed: %v\n", err)
break
}
count = count + 1
}
}
}
I found some issues due to not being able to write concurrently into the connection with an error inconsistent fdMutex. This is due to reaching over 0xfffff concurrent writes, which fdMutex does not support. To mitigate this issue, make sure you don't go over that number of concurrent writes. In my system, it was >100k/s. This is not the 300k/s you're expecting, but my system is not prepared for that.