[systemd-devel] Systemd, cgroups and short-lived processes

[Boris Dolgov]

Hi Lennart,

thanks for your answer! Please, see some more questions below.

On Wed, May 3, 2017 at 11:54 PM, Lennart Poettering
<lennart at poettering.net> wrote:
> On Wed, 03.05.17 18:02, Boris Dolgov (boris at dolgov.name) wrote:
>
>> Hello!
>>
>> I am using systemd and repeatedly need to do the following:
>> - Start a process with memory limit and cpu affinity;
>> - Be notified when it and all its children are finished (typically in
>> 0.01-1 second);
>> - Get cpu and memory usage statistics.
>>
>> Cgroups look perfect for me, but what is the best way of managing them?
>>
>> As far as I understand, ControlGroupInterface article advises to
>> create a transient scope unit for each process I run, but I need to
>> preserve cpu accounting information, so scope units will not work.
>
> They will. Just make sure to use a recent systemd version that
> provides you with the Ref/Unref() bus calls, which permits you to ref
> specific units, so that you can read statistics off them.

Looks like Fedora 26 has these calls, that's nice.

>> Alternatively, I could create service units, but I am concerned with
>> performance issues and would like to create processes by myself.
>>
>> Are any lower-level APIs available and advised to use?
>> For example, I'd be happy to create a slice using systemd, create its
>> subtrees manually and manage them myself, but the aforementioned
>> article does not recommend it.
>
> It should be sufficient to use scopes with Ref/Unref and the CPUShares
> setting.
>
> However, the cpuset controller is still not exposed in systemd, for
> the reason that its semantics are really really broken, still. We'll
> add support as soon as that is cleaned up in the kernel.
>
> However, given that the per-process CPU affinitty stuff exposed by
> libc is usually enough this should not be a big limitation.
Yes, looks like sched_setaffinity is enough for me.

But I am still concerned with performance issues for a lot of
short-lived processes.
I've tried to run the following program:
https://gist.github.com/bdolgov/539214149042830318e0f8e5073ea274
With CMD=nosd_cmd it does 200 runs per second, but with CMD=sd_cmd, it
does only 120 runs per second. It looks like a 40% slowdown, which is
not very good for me.

Thanks,
Boris.