[petsc-dev] Comments about PThreads

[Barry Smith]

  Yup, all good ideas, which is why it is open source open community development. :-)

  More seriously if we are lucky Shri will accept a post-doc position at ANL very soon and start moving forward rapidly on the pthread support.

   Barry


  
On Sep 14, 2011, at 10:11 AM, Jose E. Roman wrote:

> I have been doing some tests --with-pthreadclasses. I use
> $ ./ex12 -n 250 -m 250
> (in $PETSC_DIR/src/ksp/ksp/examples/tutorials)
> 
> 1) I tried with -vec_type pthread as explained in the instructions in features/threads.html
> This had a noticeable effect on the run time. However, in another system the execution time did not change, and I realized that it was necessary to also add: -thread_max 2 -main. Question: should these always be used? what is the meaning of -main?
> 
> 2) I thought adding -mat_type aijpthread would yield even better timings, but the response time was the same. Then I noticed that -mat_view_info showed this:
> Matrix Object: 1 MPI processes
>  type: seqaij
>  rows=62500, cols=62500
> so I guess -mat_type aijpthread is not having any effect.
> 
> 3) I understand that this is still in a very initial stage. However, I think that in order to be usable it is necessary to add MatGetVecs_SeqAIJPThreads (similarly to CUSP). For instance, in SLEPc users normally do not pass vectors but just matrices, and the vectors are created with MatGetVecs (no VecSetFromOptions).
> 
> Jose
>