How many cores I am using on a Linux Server?

Question

I am just wondering how I can find out in bash how many CPU cores a user is now using on a Linux Server?

I am submitting a fair amount of background jobs to the server, so I'd like to write a bash script to check before I submit a job if the running jobs (processes) submitted by me are too many, because I have to let other users have enough cores to use for their jobs. I'd like to know what bash command can tell me how many cores my jobs are using now.

Thanks and regards!

Solution 1

To get the number of CPU cores per CPU:

grep "^core id" /proc/cpuinfo | sort -u | wc -l

Or to get the number of physical CPUs:

grep "^physical id" /proc/cpuinfo | sort -u | wc -l

Solution 2

I don't know if it helps, but you can use the mpstat utility to get a breakdown of CPU usage by individual processor (or core). For example:

$ mpstat -P ALL 1

12:49:59 PM  CPU   %user   %nice    %sys %iowait    %irq   %soft  %steal   %idle    intr/s
12:50:00 PM  all    7.89    0.00    1.25    0.88    0.00    0.00    0.00   89.97   1359.00
12:50:00 PM    0   14.00    0.00    0.00    0.00    0.00    0.00    0.00   86.00   1043.00
12:50:00 PM    1   15.84    0.00    7.92    3.96    0.00    0.99    0.00   71.29    297.00
12:50:00 PM    2    3.96    0.00    0.00    1.98    0.00    0.99    0.00   93.07      0.00
12:50:00 PM    3    3.96    0.00    0.99    2.97    0.00    0.00    0.00   92.08      0.00
12:50:00 PM    4    4.00    0.00    0.00    0.00    0.00    0.00    0.00   96.00      0.00
12:50:00 PM    5    4.95    0.00    0.99    0.00    0.00    0.00    0.00   94.06     18.00
12:50:00 PM    6   10.89    0.00    0.99    0.00    0.00    0.00    0.00   88.12      0.00
12:50:00 PM    7    5.05    0.00    0.00    0.00    0.00    0.00    0.00   94.95      0.00

In this example, you can see that CPU's 0, 1, and 6 are doing more work than the rest of them. Sometimes you will see that a single CPU is near (or at) 100% while others are at zero. This can be an indicator of program (or portion of a program) that is single-threaded and only able to use a single CPU at a time.

To install mpstat on a Fedora, RHEL, or CentOS system, use yum install sysstat.

Solution 3

So you'll see there are responses here that will tell you how your cores are being utilized.

HOWEVER - this isn't really doing you a service. You've made a basic assumption that just doesn't hold - that your jobs will tend to group themselves onto some subset of the cores.

Instead, your jobs will be spread out across all of the cores, unless you implement something that keeps them "penned in" somehow. (Note: I'm not recommending that; just saying, "unless")

Here's an alternative strategy: Identify for your particular system what the LOAD level is when you feel it's "acceptable" for other users to be adding more jobs. Then, build something that only submits a new background job when the load level is down below that limit.

That way, the solution will be core-count-independent, more portable, more flexible, and easier to "tweak" as well.

Solution 4

If you want to do this so it works on linux and OS X, you can do:

CORES=$(grep -c ^processor /proc/cpuinfo 2>/dev/null || sysctl -n hw.ncpu)

Solution 5

Unless explicitly configured not to (ie. pinning a process to a specific CPU), all cores can be assumed to be in use at all times. The scheduler will allocate processes the next available core. Case in point, "System Monitor" (part of GNOME) is showing my load almost the same across all 4 cores of my machine.

View more solutions

Comments

[baz]

I am submitting a fair amount of background jobs to the server, so I like to write a bash script to check before I submit a job if the running jobs(processes) submit by me are too many, because I have to let other users have enough cores to use for their jobs. So I like to know what bash command can tell me how many cores my jobs are using now.

[baz]

I like to know a bash command that can tell me the number, since I want to use it in a bash script.

[Marc Novakowski]

@Tim top can run in batch mode with the -b switch. Combine it with the -n switch to get the number of iterations you need.

[baz]

can you give an example of code for obtaining the number of cores used by all the jobs (processes) submit by a user?

[Marc Novakowski]

@Tim Not off the bat. I'd either have to write such a script by parsing the top output or google to see if someone's already done it. However, I'm not convinced its worth the effort because I think I'll get different outputs for every run given the speed at which context switching occurs. Isn't the system load a good enough indicator? It effectively tells you how many virtual cores are required to handle the process queue (processes waiting on the CPU).

[baz]

@ nagul: Yes system load is good and I also check that too. However some people keep telling me that I can not use up all the cores because who knows others will run their jobs soon. So I guess I have to limit the cores I am using. By the way, have you seen somewhere online scripts that check the CPU and memory load and dynamically submit jobs based on the usage? I have been looking them for so long. Thanks!

[baz]

Thank you! I do check the overall system load. I have been told by one of my colleagues that I can not use up all the cores because who knows others will run their jobs soon and my jobs will last for quite a while. Do you also think I have to limit the cores I am using?

[dmckee --- ex-moderator kitten]

Trying to count cores is the wrong way to allocate CPU. That is the kernel's job. To tweak it, use [re]nice, or if that isn't enough get a queue manager like pbs.

[dmckee --- ex-moderator kitten]

Relevant link: superuser.com/questions/49020/linux-queuing-system

[Marc Novakowski]

@Tim I'd suggest a different approach. Like EmmEff says, both cores will be used as long as system load is above 1.0. Have a script that parses top output to get %CPU utilization for your processes, and if the cumulative usage goes above a specified threshold, lowers their priority using nice. Or better still, if you absolutely have to make sure you don't hog all cores, bind all processes to one core: cyberciti.biz/tips/…

[baz]

@ nagul: Thanks! I finally get it done by parsing the top

[pbr]

I just now noticed a comment you added to another poster's answer, indicating this is because of something a colleague told you. So, indeed, it appears that the misconception is his, not yours. Sorry about that!

[baz]

Thank you! As to "Identify for your particular system what the LOAD level is when you feel it's "acceptable" for other users to be adding more jobs", if the current load is not mainly caused by my running jobs then I would feel more free to use the rest CPUs, but if in the current load my running jobs take a fair amount then I probably have to give up using the rest CPUs. So do you still think I don't have to estimate the CPU usage of my total running jobs?

[pbr]

No, I really don't think it makes sense to distinguish between how much of the current load is from your jobs vs. other jobs. Consider the simplicity: your job tries to be a good citizen, and only submits itself if the load is below a level which you consider acceptable. The idea is NEVER to utilize all the remaining capacity; that leaves other future users in a bind. Ideally, all of the users would be using the same logic for this.

[Nav]

Thanks a million for providing this answer! My colleagues were using the "top" command and making all the wrong conclusions about processor utilization until I showed them the mpstat command.

[Alix Axel]

Not available on Ubuntu: # aptitude search mpstat #.

[Bad Boy]

@AlixAxel try aptitude search systat instead.

[Alix Axel]

@MattSolnit: Still no luck.

[user]

mpstat is provided by the package sysstat on Debian and on Ubuntu. I would expect most Debian-derived distributions use the same package name. On Debian-derived distributions, start by apt-cache search --full mpstat.

[itoctopus]

This is the best answer - as it shows what each core is doing. It works right out of the box on a Centos 6 server.

[seralouk]

how to check this for a given user?