Cross-platform process cpu % and memory usage of a PID
Ideas from https://github.com/arunoda/node-usage/ but with no C-bindings
Please note that if you need to check a nodejs script process cpu usage, you can use process.cpuUsage
since node v6.1.0. This script remain useful when you have no control over the remote script, or if the process is not a nodejs process.
var pusage = require('pidusage')
pusage.stat(process.pid, function(err, stat) {
expect(err).to.be.null
expect(stat).to.be.an('object')
expect(stat).to.have.property('cpu')
expect(stat).to.have.property('cpuTime')
expect(stat).to.have.property('memory')
console.log('Pcpu: %s', stat.cpu)
console.log('CpuTime: %s', stat.cpuTime) //those are seconds
console.log('Mem: %s', stat.memory) //those are bytes
})
// Unmonitor process
pusage.unmonitor(process.pid);
A check on the os.platform
is done to determine the method to use.
We use /proc/{pid}/stat
in addition to the the PAGE_SIZE
and the CLK_TCK
direclty from getconf()
command. Uptime comes from proc/uptime
file because it's more accurate than the nodejs os.uptime()
.
/!\ As stated in #17, memory will increase when using pidusage.stat
in an interval because of readFile
. Use --expose-gc
and release the garbage collector to avoid such leaking.
Cpu usage is computed by following those instructions. It keeps an history of the current processor time for the given pid so that the computed value gets more and more accurate. Don't forget to do unmonitor(pid)
so that history gets cleared.
Cpu usage does not check the child process tree!
Memory result is representing the RSS (resident set size) only by doing rss*pagesize
, where pagesize
is the result of getconf PAGE_SIZE
.
We use a fallback with the ps -o pcpu,rss -p PID
command to get the same informations.
Memory usage will also display the RSS only, process cpu usage might differ from a distribution to another. Please check the correspoding man ps
for more insights on the subject.
AIX is tricky because I have no AIX test environement, at the moment we use: ps -o pcpu,rssize -p PID
but /proc
results should be more accurate! If you're familiar with the AIX environment and know how to get the same results as we've got with Linux systems, please help.
#4
Windows is really tricky, atm it uses the wmic.exe
: wmic PROCESS {PID} get workingsetsize,usermodetime,kernelmodetime
.
The memory usage here is what windows calls the "Working Set":
Maximum number of bytes in the working set of this process at any point in time. The working set is the set of memory pages touched recently by the threads in the process. If free memory in the computer is above a threshold, pages are left in the working set of a process even if they are not in use. When free memory falls below a threshold, pages are trimmed from working sets. If they are needed, they are then soft-faulted back into the working set before they leave main memory.
The CPU usage is computed the same as it is on linux systems. We have the kernelmodetime
and the usermodetime
processor use. Every time pidusage.stat
is called, we can calculate the processor usage according to the time spent between calls (uses os.uptime()
internally).
Note that before we used wmic path Win32_PerfFormattedData_PerfProc_Process WHERE IDProcess=
(which is slow as hell) and Win32_PerfRawData_PerfProc_Process
(which api breaks on Windows 10 and Windows server 2012). Not every Windows bugged but just some of those. However, the wmic PROCESS
call is faster, and safer as it must be used by internal programs since windows XP (this is clearly an assumption).
This is the safest implementation I've found that works on most Windows version (>= XP). I've tried many other implementations but there was always some failing test case. For example, powershell would be faster but powershell needs to be attached to a console (see this comment). This means it'd have to popup a new cmd.exe
every time we execute pidusage
.
If you know a way that doesn't imply the use of wmic
, please open an issue so that I can try it!
If you want to compute a pidusage tree take a look at pidusage-tree.
MIT