[petsc-users] Pthread support

[Shri]

On Jul 23, 2012, at 8:00 PM, Alexander Goncharov wrote:

> Hi Jed,
> 
> thank you for a quick reply. If I run ex2 with -mat_type seqaijpthread
> -nthreads 4, everything runs fine. I see 400% utilization.

I've taken out all this code from the development version. Are you using petsc-3.3 while running this?
> 
> The problem persists if I just start two instances of the ex2 from the
> console without seqaijpthread/seqpthread options. Just two single
> processor jobs. They show up as two 50% jobs in the "top" output and run
> slower. I hope I made my problem more clear. 
> 
> Thank you,
> Alexander.
> 
> On Mon, 2012-07-23 at 21:41 -0500, Jed Brown wrote:
>> On Mon, Jul 23, 2012 at 9:08 PM, Alexander Goncharov
>> <alexvg77 at gmail.com> wrote:
>>        Hello!
>> 
>>        I have a question about pthread support in PETSC. I could not
>>        find it in
>>        the development version. Is it going to be supported in the
>>        new release?
>> 
>>        The reason I tried development version is because in petsc-3.3
>>        compiled
>>        with pthreadclasses=1 all jobs would be sent to one core, if I
>>        start
>>        executables separately. Same for MPI, mpirun -np 4 would show
>>        up as 4
>>        processes with 25% CPU usage each, in the "top" output.
>>        Without
>>        pthreadclasses all is ok. I would have processes with each
>>        utilizing
>>        separate core at 100%. Have you come across such a behavior?
>> 
>> 
>> Sounds like an affinity problem. When running with multiple MPI
>> processes and threads, you generally have to specify the number of
>> threads manually.
>> 
>> 
>> The threading code is in the process of being overhauled to a cleaner
>> and more flexible design. It's not ready for use yet, but should be in
>> a couple months.
> 
>