How to get CPU usage

Question

My Go program needs to know the current cpu usage percentage of all system and user processes.

How can I obtain that?

Vainglorious

which language and OS you use? – Skeptical

I use Go (from golang.org) under Linux, but I would like to use something portable to other *nix if it's possible. – Sevilla

Not exactly an answer, but I tried multiple ways to do it (including reading the proc/stat) file and then realised that the service is deployed on Kubernetes and we were already reporting the usage from the pod. So I went with that one instead. – Indoiranian

Answer 1

I had a similar issue and never found a lightweight implementation. Here is a slimmed down version of my solution that answers your specific question. I sample the /proc/stat file just like tylerl recommends. You'll notice that I wait 3 seconds between samples to match top's output, but I have also had good results with 1 or 2 seconds. I run similar code in a loop within a go routine, then I access the cpu usage when I need it from other go routines.

You can also parse the output of top -n1 | grep -i cpu to get the cpu usage, but it only samples for half a second on my linux box and it was way off during heavy load. Regular top seemed to match very closely when I synchronized it and the following program:

package main

import (
    "fmt"
    "io/ioutil"
    "strconv"
    "strings"
    "time"
)

func getCPUSample() (idle, total uint64) {
    contents, err := ioutil.ReadFile("/proc/stat")
    if err != nil {
        return
    }
    lines := strings.Split(string(contents), "\n")
    for _, line := range(lines) {
        fields := strings.Fields(line)
        if fields[0] == "cpu" {
            numFields := len(fields)
            for i := 1; i < numFields; i++ {
                val, err := strconv.ParseUint(fields[i], 10, 64)
                if err != nil {
                    fmt.Println("Error: ", i, fields[i], err)
                }
                total += val // tally up all the numbers to get total ticks
                if i == 4 {  // idle is the 5th field in the cpu line
                    idle = val
                }
            }
            return
        }
    }
    return
}

func main() {
    idle0, total0 := getCPUSample()
    time.Sleep(3 * time.Second)
    idle1, total1 := getCPUSample()

    idleTicks := float64(idle1 - idle0)
    totalTicks := float64(total1 - total0)
    cpuUsage := 100 * (totalTicks - idleTicks) / totalTicks

    fmt.Printf("CPU usage is %f%% [busy: %f, total: %f]\n", cpuUsage, totalTicks-idleTicks, totalTicks)
}

It seems like I'm allowed to link to the full implementation that I wrote on bitbucket; if it's not, feel free to delete this. It only works on linux so far, though: systemstat.go

Mauk

Answer 2

Check out this package http://github.com/c9s/goprocinfo, goprocinfo package does the parsing stuff for you.

stat, err := linuxproc.ReadStat("/proc/stat")
if err != nil {
    t.Fatal("stat read fail")
}

for _, s := range stat.CPUStats {
    // s.User
    // s.Nice
    // s.System
    // s.Idle
    // s.IOWait
}

Indeterminate

Answer 3

The mechanism for getting CPU usage is OS-dependent, since the numbers mean slightly different things to different OS kernels.

On Linux, you can query the kernel to get the latest stats by reading the pseudo-files in the /proc/ filesystem. These are generated on-the-fly when you read them to reflect the current state of the machine.

Specifically, the /proc/<pid>/stat file for each process contains the associated process accounting information. It's documented in proc(5). You're interested specifically in fields utime, stime, cutime and cstime (starting at the 14th field).

You can calculate the percentage easily enough: just read the numbers, wait some time interval, and read them again. Take the difference, divide by the amount of time you waited, and there's your average. This is precisely what the top program does (as well as all other programs that perform the same service). Bear in mind that you can have over 100% cpu usage if you have more than 1 CPU.

If you just want a system-wide summary, that's reported in /proc/stat -- calculate your average using the same technique, but you only have to read one file.

Nimbostratus

Answer 4

You can use the os.exec package to execute the ps command and get the result.

Here is a program issuing the ps aux command, parsing the result and printing the CPU usage of all processes on linux :

package main

import (
    "bytes"
    "log"
    "os/exec"
    "strconv"
    "strings"
)

type Process struct {
    pid int
    cpu float64
}

func main() {
    cmd := exec.Command("ps", "aux")
    var out bytes.Buffer
    cmd.Stdout = &out
    err := cmd.Run()
    if err != nil {
        log.Fatal(err)
    }
    processes := make([]*Process, 0)
    for {
        line, err := out.ReadString('\n')
        if err!=nil {
            break;
        }
        tokens := strings.Split(line, " ")
        ft := make([]string, 0)
        for _, t := range(tokens) {
            if t!="" && t!="\t" {
                ft = append(ft, t)
            }
        }
        log.Println(len(ft), ft)
        pid, err := strconv.Atoi(ft[1])
        if err!=nil {
            continue
        }
        cpu, err := strconv.ParseFloat(ft[2], 64)
        if err!=nil {
            log.Fatal(err)
        }
        processes = append(processes, &Process{pid, cpu})
    }
    for _, p := range(processes) {
        log.Println("Process ", p.pid, " takes ", p.cpu, " % of the CPU")
    }
}

Uranus

String parsing from a command output is not a great solution as the command can change and it's also heavy on cpu/mem/etc. – Chilson

@EricAnderson Did you notice the accepted answer (written more than one year after mine) does the same ? Did you choose to answer only the first one to provide a solution to the question ? – Helve

This is a good answer as it does work on Mac OS X 10.10 with just copy& paste (while the accepted answer doesn't) – Catnap

The OP tagged Linux - not MacOS. So while this is a nice multi-OS solution - shelling out is not really the go-way. – Overripe

Answer 5

Here is an OS independent solution using Cgo to harness the clock() function provided by C standard library:

//#include <time.h>
import "C"
import "time"

var startTime = time.Now()
var startTicks = C.clock()

func CpuUsagePercent() float64 {
    clockSeconds := float64(C.clock()-startTicks) / float64(C.CLOCKS_PER_SEC)
    realSeconds := time.Since(startTime).Seconds()
    return clockSeconds / realSeconds * 100
}

Grande

You could simplify time.Now().Sub(startTime) to time.Since(startTime). – Cannikin

Sidenote: clock() gives the time of the current proccess only. – Gera

Answer 6

I recently had to take CPU usage measurements from a Raspberry Pi (Raspbian OS) and used github.com/c9s/goprocinfo combined with what is proposed here:

• Accurate calculation of CPU usage given in percentage in Linux?

The idea comes from the htop source code and is to have two measurements (previous / current) in order to calculate the CPU usage:

func calcSingleCoreUsage(curr, prev linuxproc.CPUStat) float32 {

  PrevIdle := prev.Idle + prev.IOWait
  Idle := curr.Idle + curr.IOWait

  PrevNonIdle := prev.User + prev.Nice + prev.System + prev.IRQ + prev.SoftIRQ + prev.Steal
  NonIdle := curr.User + curr.Nice + curr.System + curr.IRQ + curr.SoftIRQ + curr.Steal

  PrevTotal := PrevIdle + PrevNonIdle
  Total := Idle + NonIdle
  // fmt.Println(PrevIdle, Idle, PrevNonIdle, NonIdle, PrevTotal, Total)

  //  differentiate: actual value minus the previous one
  totald := Total - PrevTotal
  idled := Idle - PrevIdle

  CPU_Percentage := (float32(totald) - float32(idled)) / float32(totald)

  return CPU_Percentage
}

For more you can also check https://github.com/tgogos/rpi_cpu_memory

Kastroprauxel