[Xenomai-help] Thread scheduling on mult-processor systems

[Xenomai-help] Thread scheduling on mult-processor systems

[Rob Gubler]

[-- Attachment #1: Type: text/plain, Size: 1305 bytes --]

Hello,

I am trying execute 2 threads on 2 separate processors.  To do this I
created 2 threads, each using the FIFO scheduler where both threads have the
same priority defined by sched_get_priority_max(SCHED_FIFO).  Each thread is
initialized to run on a specific CPU with a call to
pthread_setaffinity_np(...); Thread 1 on CPU 0, and Thread 2 on CPU 1.  I
verified each thread was assigned correctly by calling
pthread_getaffinity_np(...), and also by looking at /proc/xenomai/sched,
while the program was running.

For the purposes of testing, both threads do the same thing:

        for(unsigned long i = 0; i != 9000000; i++)
        {
                if(!(i % 50000))
                {
                        std::cout << "tick" << std::endl;
                }

                sched_yield();
        }

        std::cout << "elapsed time: " << [elapsed time];

My hope was that when these threads would execute simultaneously, they would
truly run in parallel on their respective processor.  But, what I am
noticing is the elapsed time when both threads run is 12 seconds, while when
only one thread runs is 6 seconds.  So, I haven't seem to have bought any
performance improvements by assigning  the thread to run on a specific
processor.

 Where have I gone wrong in my line of thinking?

-Rob

[-- Attachment #2: Type: text/html, Size: 2027 bytes --]

Re: [Xenomai-help] Thread scheduling on mult-processor systems

[Philippe Gerum]

Rob Gubler wrote:
> Hello,
> 
> I am trying execute 2 threads on 2 separate processors.  To do this I
> created 2 threads, each using the FIFO scheduler where both threads have
> the same priority defined by sched_get_priority_max(SCHED_FIFO).  Each
> thread is initialized to run on a specific CPU with a call to
> pthread_setaffinity_np(...); Thread 1 on CPU 0, and Thread 2 on CPU 1. 
> I verified each thread was assigned correctly by calling
> pthread_getaffinity_np(...), and also by looking at /proc/xenomai/sched,
> while the program was running.
> 
> For the purposes of testing, both threads do the same thing:
> 
>         for(unsigned long i = 0; i != 9000000; i++)
>         {
>                 if(!(i % 50000))
>                 {
>                         std::cout << "tick" << std::endl;
>                 }
> 
>                 sched_yield();
>         }
> 
>         std::cout << "elapsed time: " << [elapsed time];
> 
> My hope was that when these threads would execute simultaneously, they
> would truly run in parallel on their respective processor.  But, what I
> am noticing is the elapsed time when both threads run is 12 seconds,
> while when only one thread runs is 6 seconds.  So, I haven't seem to
> have bought any performance improvements by assigning  the thread to run
> on a specific processor.
> 
>  Where have I gone wrong in my line of thinking? 
> 

Most of the code executed in this loop involves serialization, hence the result.

-- 
Philippe.