Get number of busy CPUs in Python

Question

I am writing a multiprocessing routine to run on a server with plenty of CPUs. However, the server has multiple users and its usage may vary. So I would like to adapt the number of processors being used according to the current load.

Is there a way to estimate the amount of CPUs currently busy in Python? I only found multiprocessing.cpu_count()
Bonus question: Is it possible to change multiprocessing.Pool(processes=no_cpus) during activity, in case the load on the server has changed after a while?

If several such programms run they might mutually misleading themselves. — NoDataDumpNoContribution, Jul 14 '14 at 09:15
@alKid: good question, maybe 80% load, active threads? unfortunately I have little experience in this... — n1000, Jul 14 '14 at 09:26
@Trilarion: good point :D I'll assume I am the only one implementing such a check. — n1000, Jul 14 '14 at 09:27

score 7 · Accepted Answer · edited May 23 '17 at 10:27

There are a number of complications...

you can't determine which CPUs are busy

Processes (and threads) are scheduled by the Linux kernel on any CPU. Even determining the "current CPU" is awkward -- see How can I see which CPU core a thread is running in?

multiprocessing.Pool is designed to start up N workers, which run "forever." Each accepts a task from a queue, does some work, then outputs data. A Pool doesn't change size.

Two suggestions:

the uptime command outputs something like this:

19:05:07 up 4 days, 20:43, 3 users, load average:0.99, 1.01, 0.82

The last three numbers are the "load average" over the last minute, five minutes, and 15 minutes. Consider using the first number to load-balance your application.

consider having your application do time.sleep(factor) after completing each piece of work.

Thus you can increase the factor when the system is busy (high load average), and make the delay shorter when the system is more idle (low load; ie surfing). Pool stays same size.

interesting approach! however my script is also producing a lot of data and therefore the disks are quite busy while cpu is is not fully used. i will give it a try. — n1000, Jul 15 '14 at 07:40
Accepted this answer over the other, since it shows an actual approach how to solve my problem. — n1000, Oct 12 '14 at 07:40

FooBar · Answer 2 · 2016-02-29T16:52:54.730

I followed johntellsall's solution, here's some actual simple schematic. There is one, as Python (for me) confuses virtual with actual cpu. I've decided to calibrate against the average load of the past 15 minutes.

Sleeping numbers are quite arbitrary.

def sleepForMultiCore():
    # divide by 2 since Python does not distinguish physical and virtual core
    cores = 0.5*mp.cpu_count()
    loadAvg = os.getloadavg()[2]

    if loadAvg > cores*1.3:
        sleepTime = 5*60
    elif loadAvg > cores:
        sleepTime = 2*60
    elif loadAvg > cores*0.9:
        sleepTime = 1*60
    else:
        sleepTime = 0
    print ('sleeping for ', sleepTime)
    time.sleep(sleepTime)

score 1 · Answer 3 · edited Jun 20 '20 at 09:12

I think the best library for your question is psutil with which you can manage some CPU and Memory usage in some different OS':

psutil is a module providing an interface for retrieving information on running processes and system utilization (CPU, memory) in a portable way by using Python, implementing many functionalities offered by tools like ps, top and Windows task manager.

It currently supports Linux, OS X, FreeBSD and Windows with Python versions from 2.4 to 3.1 by using a unique code base.

Get number of busy CPUs in Python

3 Answers3