Hi Jan, We are currently getting the total system time and all linux pthread run time using the clock_gettime(). Then we are getting the linux idle time from /proc/stat. We are getting the xenomai thread times and total xenomai time from /proc/xenomai/sched/acct. Subtract the total xenomai time from linux ide. Then calculate as a percentage individual thread CPU utilization and reserve percentage. We can do this in a Linux thread, but it would be nice if we had a single interface/api for getting the per thread CPU utilization for the whole system(Linux and Xenomai).
Rajesh On Fri, Aug 27, 2021 at 5:40 AM Jan Kiszka <jan.kis...@siemens.com> wrote: > On 27.08.21 03:15, Rajesh Venkataraman via Xenomai wrote: > > Hi, > > For the linux pthreads we have used pthread_getcpuclockid() to get the > > threads cpu time clock. Then periodically we do a clock_gettime() on this > > timer to get all threads run time. Then do clock_gettime( > CLOCK_MONOTONIC, > > &ts ) to get the total system. Using the individual thread run time and > > total run time, we calculate the percentage of CPU utilization for each > > thread. Now we have moved the realtime thread to xenomai user space. > > The clock_gettime() on a xenomai user thread returns an error. How do > we > > calculate what percentage of the CPU is utilized by the xenomai thread in > > our Application? > > > > We also get the idle time of the CPU from the /proc/stat and calculate > > the reserve percentage based on the total system time. Does linux idle > time > > also include xenomai run time? > > > > Is this correct method for calculating the CPU utilization percentage > per > > thread in a system/application with xenomai and linux pthreads running > in a > > single process? > > > > Xenomai does not support pthread_getcpuclockid for reading the consumed > time of a thread. There are stats under /proc/xenomai/sched/acct, rtps > uses them, but reading those require switching to Linux (if that is an > issue for your use case). And there is an internal cobalt_thread_stat() > syscall already, used only by some non-posix skins so far. > > I'm sitting on non-upstream changes to export the CPU usage of thread to > userland without leaving RT, actually even without issuing a syscall. > But those need rework in order to attach to some posix interface. > > What are your requirements on a statistic interface? Who is going to > monitor who, and under which constraints? > > Jan > > -- > Siemens AG, T RDA IOT > Corporate Competence Center Embedded Linux >
