Execute HPL with the maximum threads

Tuyen__Nguyen — Thu, 31 Oct 2019 10:49:01 GMT

My platform is KNL 7250 ( 1.4GHz, 68 cores).

I used Intel Vtune Amplifier to check the total thread.

I want to execute HPL 2.3 with full threads (272 threads) but can not. I tried 2 way: setup environment variables (OMP_NUM_THREADS=272) or modified HPL_pdgesv.c with add

omp_set_num_threads(272);

#pragma omp parallel

{
    if( ( ALGO->depth == 0 ) || ( GRID->npcol == 1 ) )
   {
      HPL_pdgesv0(  GRID, ALGO, A );
   }
   else
   {
      HPL_pdgesvK2( GRID, ALGO, A );
   }

/*
 * Solve upper triangular system
 */
   if( A->info == 0 ) HPL_pdtrsv( GRID, A );
}

But HPL just works with a maximum 68 threads. If I decrease numbers of threads less than 68, it will work with my number of threads setup. But it does not work with numbers of threads greater than 68.

How can I use full threads with HPL?

Thanks!

topic Execute HPL with the maximum threads in Software Tuning, Performance Optimization & Platform Monitoring

Execute HPL with the maximum threads