Date: Fri, 22 Feb 2019 12:19:35 +0100 From: Pieter de Goeje <pieter@degoeje.nl> To: Eugene Grosbein <eugen@grosbein.net>, Farhan Khan <farhan@farhan.codes>, freebsd-hackers@freebsd.org Subject: Re: Optimize execution of processes by CPU core Message-ID: <7ff616b5-03c5-7650-d28b-6e85e8aa5f0e@degoeje.nl> In-Reply-To: <b0292ddf-642f-750e-6970-12492c0ced1b@grosbein.net> References: <2fcd9aee092730e11880c3ae88de4898@farhan.codes> <b0292ddf-642f-750e-6970-12492c0ced1b@grosbein.net>
next in thread | previous in thread | raw e-mail | index | archive | help
Op 20-2-2019 om 19:19 schreef Eugene Grosbein:
> 21.02.2019 0:35, Farhan Khan via freebsd-hackers wrote:
>> Hi all,
>>
>> I am trying to optimize the execution of a CPU-intensive workload where I am running multiple instances of a program.
>> The moment the program ends (expected behavior), the calling shell script verifies the results and if its good it reruns the program.
>> The machines I am running this on have 8 cores, but ps reports that some of the processes frequently run on the same CPU,
>> so I suspect I am not getting optimized performance.
>
> System scheduler switches useland processed from one CPU core to another often enough
> so each CPU has even load and you cannot see that with naked. But you can easily verify
> if load is even or not checking sysctl kern.cp_times that shows five monotonically increasing
> counters per each CPU core. For example, in case of dual-core system:
>
> $ sysctl kern.cp_times
> kern.cp_times: 14789486 132229 14016113 327160 949428773 14374865 139326 12056998 2941012 949179638
> $ sysctl kern.clockrate
> kern.clockrate: { hz = 1000, tick = 1000, profhz = 8126, stathz = 127 }
>
> There are "stathz" ticks per second and for each core exactly one of five counters is incremented by one:
> user, nice, system, interrupt, idle. That is, each 5th counter is incremented
> if corresponding CPU core was idle during "the tick".
>
> You can save output of sysctl kern.cp_times, run your test, stop it and save output again.
> Then compare difference of each 5th counter and they should be approximately equal.
>
> You can even draw graphs if you periodically get samples of the sysctl, compute diffs with previous samples,
> divide diffs by period length in seconds and then divide again by "stathz" value.
> Multiply by 100 to get idle time of single CPU code in percents for the period.
> Repeat for each core.
>
> I use net-mgmt/mrtg to draw such per-CPU graphs for my servers, it works just fine.
`top -P` does the same thing. It displays a
user/nice/system/interrupt/idle line for each CPU.
- Pieter
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?7ff616b5-03c5-7650-d28b-6e85e8aa5f0e>