Merge branch 'task-pinning' into 'master_pre_stage'

Task pinning

See merge request !53

Makefile.am

@@ -69,6 +69,8 @@ libelpa@SUFFIX@_private_la_SOURCES = \
   src/elpa1/mod_distribute_global_column.F90 \
   src/elpa1/mod_v_add_s.F90 \
   src/elpa1/mod_solve_secular_equation.F90 \
+  src/helpers/mod_thread_affinity.F90 \
+  src/helpers/check_thread_affinity.c \
   src/elpa_index.c
 
 libelpa@SUFFIX@_private_la_SOURCES += src/elpa_c_interface.c

ci_test_scripts/generate_gitlab_ci_tests.py

@@ -610,9 +610,14 @@ coverage = {
         #"knl" : "--enable-avx512",
         #"power8" : " --enable-vsx --disable-sse --disable-sse-assembly --disable-avx --disable-avx2 --disable-avx512 --disable-mpi-module --with-GPU-compute-capability=sm_60 ",
 
+#instruction_set = {
+#        "sse" : " --enable-sse --enable-sse-assembly",
+#        "avx" : " --enable-avx",
+#        "avx2" : " --enable-avx2",
+#        "avx512" : "--enable-avx512",
+#}
+
 instruction_set = {
-        "sse" : " --enable-sse --enable-sse-assembly",
-        "avx" : " --enable-avx",
         "avx2" : " --enable-avx2",
         "avx512" : "--enable-avx512",
 }

src/elpa1/elpa1_template.F90

@@ -83,6 +83,7 @@ function elpa_solve_evp_&
    use elpa_scalapack_interfaces
 #endif
    use solve_tridi
+   use thread_affinity
    implicit none
 #include "../general/precision_kinds.F90"
    class(elpa_abstract_impl_t), intent(inout)                         :: obj
@@ -176,6 +177,8 @@ function elpa_solve_evp_&
    MATH_DATATYPE(kind=rck), allocatable, target               :: aIntern(:,:)
    MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable, target   :: qIntern(:,:)
 #endif
+   integer(kind=c_int)                             :: pinningInfo
+
    logical                                         :: do_tridiag, do_solve, do_trans_ev
    integer(kind=ik)                                :: nrThreads
    integer(kind=ik)                                :: global_index
@@ -216,8 +219,19 @@ function elpa_solve_evp_&
 #ifdef WITH_NVTX
    call nvtxRangePush("elpa1")
 #endif
+   call obj%get("output_pinning_information", pinningInfo, error)
+   if (error .ne. ELPA_OK) then
+     print *,"Problem setting option for debug. Aborting..."
+     stop
+   endif
+   
+   if (pinningInfo .eq. 1) then
+     call init_thread_affinity(nrThreads)
 
-
+     call check_thread_affinity()
+     if (my_pe .eq. 0) call print_thread_affinity(my_pe)
+     call cleanup_thread_affinity()
+   endif
    success = .true.
 
 #ifdef REDISTRIBUTE_MATRIX
@@ -619,7 +633,6 @@ function elpa_solve_evp_&
      call blacs_gridexit(blacs_ctxt_)
    endif
 #endif /* REDISTRIBUTE_MATRIX */
-
    call obj%timer%stop("elpa_solve_evp_&
    &MATH_DATATYPE&
    &_1stage_&

src/elpa2/elpa2_template.F90

@@ -89,6 +89,7 @@
    use elpa_scalapack_interfaces
 #endif
    use solve_tridi
+   use thread_affinity
    use, intrinsic :: iso_c_binding
    implicit none
 #include "../general/precision_kinds.F90"
@@ -200,6 +201,7 @@
 #endif
    integer(kind=ik)                                                   :: global_index
    logical                                                            :: reDistributeMatrix, doRedistributeMatrix
+   integer(kind=ik)                                                   :: pinningInfo
 
 #if REALCASE == 1
 #undef GPU_KERNEL
@@ -238,6 +240,20 @@
     nrThreads = 1
 #endif
 
+   call obj%get("output_pinning_information", pinningInfo, error)
+   if (error .ne. ELPA_OK) then
+     print *,"Problem setting option for debug. Aborting..."
+     stop
+   endif
+
+   if (pinningInfo .eq. 1) then
+     call init_thread_affinity(nrThreads)
+
+     call check_thread_affinity()
+     if (my_pe .eq. 0) call print_thread_affinity(my_pe)
+     call cleanup_thread_affinity()
+   endif
+
     success = .true.
 
 #ifdef REDISTRIBUTE_MATRIX

src/elpa_index.c

@@ -271,6 +271,7 @@ static const elpa_index_int_entry_t int_entries[] = {
         BOOL_ENTRY("print_flops", "Print FLOP rates on task 0", 0, ELPA_AUTOTUNE_NOT_TUNABLE, 0, PRINT_YES),
         BOOL_ENTRY("measure_performance", "Also measure with flops (via papi) with the timings", 0, ELPA_AUTOTUNE_NOT_TUNABLE, 0, PRINT_YES),
         BOOL_ENTRY("check_pd", "Check eigenvalues to be positive", 0, ELPA_AUTOTUNE_NOT_TUNABLE, 0, PRINT_YES),
+        BOOL_ENTRY("output_pinning_information", "Print the pinning information", 0, ELPA_AUTOTUNE_NOT_TUNABLE, 0, PRINT_YES),
         BOOL_ENTRY("cannon_for_generalized", "Whether to use Cannons algorithm for the generalized EVP", 1, ELPA_AUTOTUNE_NOT_TUNABLE, 0, PRINT_YES),
 };
 

src/helpers/check_thread_affinity.c

+//    Copyright 2021, A. Marek
+//
+//    This file is part of ELPA.
+//
+//    The ELPA library was originally created by the ELPA consortium,
+//    consisting of the following organizations:
+//
+//    - Max Planck Computing and Data Facility (MPCDF), formerly known as
+//      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
+//    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
+//      Informatik,
+//    - Technische Universität München, Lehrstuhl für Informatik mit
+//      Schwerpunkt Wissenschaftliches Rechnen ,
+//    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
+//    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
+//      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
+//      and
+//    - IBM Deutschland GmbH
+//
+//
+//    This particular source code file contains additions, changes and
+//    enhancements authored by Intel Corporation which is not part of
+//    the ELPA consortium.
+//
+//    More information can be found here:
+//    http://elpa.mpcdf.mpg.de/
+//
+//    ELPA is free software: you can redistribute it and/or modify
+//    it under the terms of the version 3 of the license of the
+//    GNU Lesser General Public License as published by the Free
+//    Software Foundation.
+//
+//    ELPA is distributed in the hope that it will be useful,
+//    but WITHOUT ANY WARRANTY; without even the implied warranty of
+//    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+//    GNU Lesser General Public License for more details.
+//
+//    You should have received a copy of the GNU Lesser General Public License
+//    along with ELPA. If not, see <http://www.gnu.org/licenses/>
+//
+//    ELPA reflects a substantial effort on the part of the original
+//    ELPA consortium, and we ask you to respect the spirit of the
+//    license that we chose: i.e., please contribute any changes you
+//    may have back to the original ELPA library distribution, and keep
+//    any derivatives of ELPA under the same license that we chose for
+//    the original distribution, the GNU Lesser General Public License.
+//
+// Author: Andreas Marek, MPCDF
+
+#define  _GNU_SOURCE
+#include <stdio.h>
+#include <stdlib.h>
+#include <sched.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+
+void get_thread_affinity(int *cpu_id) {
+  *cpu_id = sched_getcpu();
+}
+
+void get_process_affinity(int cpu_id) {
+  cpu_set_t set;
+  int ret, i;
+  int cpu;
+  cpu_id = 9999999 ;
+  ret = sched_getaffinity(0, sizeof(cpu_set_t), &set);
+  for (i=0; i < CPU_SETSIZE; i++)
+    {
+      cpu = CPU_ISSET(i, &set);
+      if (cpu == 1) { cpu_id = i; }
+    }
+}
+
+void get_process_id(int *process_id, int *pprocess_id) {
+  int id;
+  *process_id = 0;
+  *pprocess_id = 0;
+  id = getpid();
+  *process_id = id ;
+  id = getppid();
+  *pprocess_id = id ; 
+  //printf("My pid %d \n",*process_id);
+  //printf("My ppid %d \n",*pprocess_id);
+}

src/helpers/mod_thread_affinity.F90

+!    Copyright 2021, A. Marek
+!
+!    This file is part of ELPA.
+!
+!    The ELPA library was originally created by the ELPA consortium,
+!    consisting of the following organizations:
+!
+!    - Max Planck Computing and Data Facility (MPCDF), formerly known as
+!      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
+!    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
+!      Informatik,
+!    - Technische Universität München, Lehrstuhl für Informatik mit
+!      Schwerpunkt Wissenschaftliches Rechnen ,
+!    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
+!    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
+!      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
+!      and
+!    - IBM Deutschland GmbH
+!
+!
+!    This particular source code file contains additions, changes and
+!    enhancements authored by Intel Corporation which is not part of
+!    the ELPA consortium.
+!
+!    More information can be found here:
+!    http://elpa.mpcdf.mpg.de/
+!
+!    ELPA is free software: you can redistribute it and/or modify
+!    it under the terms of the version 3 of the license of the
+!    GNU Lesser General Public License as published by the Free
+!    Software Foundation.
+!
+!    ELPA is distributed in the hope that it will be useful,
+!    but WITHOUT ANY WARRANTY; without even the implied warranty of
+!    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+!    GNU Lesser General Public License for more details.
+!
+!    You should have received a copy of the GNU Lesser General Public License
+!    along with ELPA. If not, see <http://www.gnu.org/licenses/>
+!
+!    ELPA reflects a substantial effort on the part of the original
+!    ELPA consortium, and we ask you to respect the spirit of the
+!    license that we chose: i.e., please contribute any changes you
+!    may have back to the original ELPA library distribution, and keep
+!    any derivatives of ELPA under the same license that we chose for
+!    the original distribution, the GNU Lesser General Public License.
+!
+! Author: Andreas Marek, MPCDF
+
+#include "config-f90.h"
+module thread_affinity
+  use precision
+
+  implicit none
+
+  public :: check_thread_affinity, &
+            init_thread_affinity, cleanup_thread_affinity, print_thread_affinity
+  private
+! integer(kind=ik) :: thread_num
+  integer(kind=ik) :: thread_max
+  integer(kind=ik) :: process_cpu_id
+  integer(kind=ik), allocatable :: cpu_ids(:)
+
+  interface
+    subroutine get_process_id_c(process_id, pprocess_id) bind(C, name="get_process_id")
+      use, intrinsic :: iso_c_binding
+      implicit none
+      integer(kind=C_INT), intent(out) :: process_id, pprocess_id
+    end subroutine
+  end interface
+
+  interface
+    subroutine get_thread_affinity_c(cpu_id) bind(C, name="get_thread_affinity")
+      use, intrinsic :: iso_c_binding
+      implicit none
+      integer(kind=C_INT) :: cpu_id
+    end subroutine
+  end interface
+  interface
+    subroutine get_process_affinity_c(cpu_id) bind(C, name="get_process_affinity")
+      use, intrinsic :: iso_c_binding
+      implicit none
+      integer(kind=C_INT), value :: cpu_id
+    end subroutine
+  end interface
+
+contains
+  subroutine get_thread_affinity(cpu_id)
+    use, intrinsic :: iso_c_binding
+    implicit none
+    integer(kind=ik), intent(out) :: cpu_id
+    integer(kind=C_INT) :: cpu_id_c
+    call get_thread_affinity_c(cpu_id_c)
+    cpu_id = int(cpu_id_c, kind=ik)
+  end subroutine
+  subroutine get_process_affinity(cpu_id)
+    use, intrinsic :: iso_c_binding
+    implicit none
+    integer(kind=ik), intent(out) :: cpu_id
+    integer(kind=C_INT) :: cpu_id_c
+    call get_process_affinity_c(cpu_id_c)
+    cpu_id = int(cpu_id_c, kind=ik)
+  end subroutine
+  subroutine get_process_id(process_id, pprocess_id)
+    use, intrinsic :: iso_c_binding
+    implicit none
+    integer(kind=ik), intent(out) :: process_id, pprocess_id
+    integer(kind=C_INT) :: process_id_c, pprocess_id_c
+    call get_process_id_c(process_id_c, pprocess_id_c)
+    process_id  = int(process_id_c,  kind=ik)
+    pprocess_id = int(pprocess_id_c, kind=ik)
+  end subroutine
+
+
+  subroutine init_thread_affinity(nrThreads)
+    use precision
+    use omp_lib
+
+    implicit none
+    integer(kind=ik)             :: istat
+    integer(kind=ik), intent(in) :: nrThreads
+
+    thread_max = nrThreads
+#ifdef WITH_OPENMP_TRADITIONAL
+    if(.not.(allocated(cpu_ids))) then
+       allocate(cpu_ids(0:thread_max-1), stat=istat)
+       if (istat .ne. 0) then
+         print *,"Error when allocating init_thread_affinity"
+       endif
+    endif
+#endif
+  end subroutine init_thread_affinity
+
+  subroutine cleanup_thread_affinity
+    use precision
+    implicit none
+    integer(kind=ik) :: istat
+
+    if((allocated(cpu_ids))) then
+       deallocate(cpu_ids, stat=istat)
+       if (istat .ne. 0) then
+         print *,"Error when deallocating init_thread_affinity"
+       endif
+    endif
+
+  end subroutine cleanup_thread_affinity
+
+  subroutine check_thread_affinity()
+    use precision
+    use omp_lib
+    implicit none
+    integer(kind=ik)             :: thread_cpu_id
+    integer(kind=ik)             :: i, actuall_num
+
+    call get_process_affinity(process_cpu_id)
+
+#ifdef WITH_OPENMP_TRADITIONAL
+
+!$OMP  PARALLEL DO &
+!$OMP  DEFAULT(NONE) &
+!$OMP  PRIVATE(i,thread_cpu_id,actuall_num) &
+!$OMP  SHARED(thread_max,cpu_ids) &
+!$OMP  SCHEDULE(STATIC)
+
+    do i=0,thread_max-1
+       call get_thread_affinity(thread_cpu_id)
+       actuall_num=omp_get_thread_num()
+       cpu_ids(actuall_num)=thread_cpu_id
+    enddo
+#endif
+
+  end subroutine check_thread_affinity
+
+  subroutine print_thread_affinity(mype)
+
+    use precision
+    implicit none
+    integer(kind=ik) :: i
+    integer(kind=ik), intent(in) :: mype
+    integer(kind=ik) :: pid, ppid
+
+    call get_process_id(pid, ppid)
+    write(*,'("Task ",i4," runs on process id: ",i4," with pid ",i4," and ppid ",i4)') mype, process_cpu_id,pid,ppid
+#ifdef WITH_OPENMP_TRADITIONAL
+    write(*,'("Each task uses ",i4," threads")') thread_max
+       do i=0,thread_max-1
+          write(*,'("Thread ",i4," is running on logical CPU-ID ",i4)') i,cpu_ids(i)
+          print *,i,cpu_ids(i)
+       enddo
+#endif
+  end subroutine print_thread_affinity
+
+end module thread_affinity