Thread's CPU time

Thread's CPU time

[Giulio Moro]

Hi everyone,
 From /proc/xenomai/sched/stat I infer there must be an internal struct somewhere keeping track of how much CPU time is actually used by each thread. Is there any way to access this kind of statistics from a running Xenomai user space program without leaving primary mode and without attempting to read this file? I am implementing an infrastructure to measure the CPU time of a given block of code. Currently I am using Cobalt's clock_gettime(), but this only works reliably if the task does not get preempted or otherwise rescheduled between the two calls.

I have run this kind of tests offline before and that's all nice and good, but I am looking for a lightweight and non invasive way of benchmarking particular sections of large-ish programs with minimal modifications.

Thanks,
Giulio

Re: Thread's CPU time

[Philippe Gerum]

Giulio Moro via Xenomai <xenomai@xenomai.org> writes:

> Hi everyone,
> From /proc/xenomai/sched/stat I infer there must be an internal struct
> somewhere keeping track of how much CPU time is actually used by each
> thread. Is there any way to access this kind of statistics from a
> running Xenomai user space program without leaving primary mode and
> without attempting to read this file? I am implementing an
> infrastructure to measure the CPU time of a given block of
> code. Currently I am using Cobalt's clock_gettime(), but this only
> works reliably if the task does not get preempted or otherwise
> rescheduled between the two calls.
>
> I have run this kind of tests offline before and that's all nice and good, but I am looking for a lightweight and non invasive way of benchmarking particular sections of large-ish programs with minimal modifications.
>
> Thanks,
> Giulio

For xenomai3, you may want to have a look at cobalt_thread_stat() from
the internal API (sys/cobalt.h). The field of interest returned in
struct cobalt_threadstat would be 'xtime' (nsecs).

-- 
Philippe.

Re: Thread's CPU time

[Giulio Moro]

Philippe Gerum wrote on 29/04/2022 17:29:>
> Giulio Moro via Xenomai <xenomai@xenomai.org> writes:
> 
>> Hi everyone,
>>  From /proc/xenomai/sched/stat I infer there must be an internal struct
>> somewhere keeping track of how much CPU time is actually used by each
>> thread. Is there any way to access this kind of statistics from a
>> running Xenomai user space program without leaving primary mode and
>> without attempting to read this file? I am implementing an
>> infrastructure to measure the CPU time of a given block of
>> code. Currently I am using Cobalt's clock_gettime(), but this only
>> works reliably if the task does not get preempted or otherwise
>> rescheduled between the two calls.
>>
>> I have run this kind of tests offline before and that's all nice and good, but I am looking for a lightweight and non invasive way of benchmarking particular sections of large-ish programs with minimal modifications.
>>
>> Thanks,
>> Giulio
> 
> For xenomai3, you may want to have a look at cobalt_thread_stat() from
> the internal API (sys/cobalt.h). The field of interest returned in
> struct cobalt_threadstat would be 'xtime' (nsecs).
> 
Thanks, I tried to implement something based on that and it seems to carry a considerable overhead over calling Cobalt's clock_gettime(): about a 3% CPU increase (as measured by /proc/xenomai/sched/stat) for 6 calls to `cobalt_thread_stat()` during a periodic task with period 360us, which is suboptimal.

A look at COBALT_SYSCALL(thread_getstat, ...) suggests that it would be possible to get just xtime for significantly cheaper than a full clock_gettime() call, considering not only that there would be less memory to be copied around, but also that - afaict - the irqsave / irqrestore calls wouldn't be needed. I am wondering whether, to this effect, it would make sense to implement `CLOCK_THREAD_CPUTIME_ID` for Cobalt's clock_gettime(), returning the 'xtime' field of this struct, e.g. (untested and probably missing includes):

```
diff --git a/kernel/cobalt/posix/clock.c b/kernel/cobalt/posix/clock.c
index 16cc8b20a..434380509 100644
--- a/kernel/cobalt/posix/clock.c
+++ b/kernel/cobalt/posix/clock.c
@@ -134,6 +134,14 @@ int __cobalt_clock_gettime(clockid_t clock_id, struct timespec64 *ts)
                 if (pipeline_get_host_time(ts) != 0)
                         return -EINVAL;
                 break;
+       case CLOCK_THREAD_CPUTIME_ID:
+               {
+                       struct xnthread *thread = xnthread_current();
+                       if (thread == NULL)
+                               return -EPERM;
+                       ns2ts(ts, xnstat_exectime_get_total(&thread->stat.account));
+               }
+               break;
         default:
                 ret = do_ext_clock(clock_id, read_monotonic, ns);
                 if (ret)
```

Or am I oversimplifying here?