Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
6974 Discussions

MKL does not respect affinity on master thread

Tue_B_
Novice
‎08-05-2015 06:42 AM
260 Views

This is follow-up/semi-related post to this rather long post: 

https://software.intel.com/en-us/comment/1825717#comment-1825717

So like the title says:

MKL does not respect affinity on the master thread. Even though I enforce MKL to spawn threads on the NUMA node the master thread is running on, only one thread runs on this NUMA node and the rest gets pushed to other NUMA nodes. This is only a problem for the masterthread, since all the other threads actually respecet the affinity they have been given.

The code below is made to run on systems with a minimum of 6 NUMA nodes with at least 6 cores on each thread. (In my case I run it on a system with 8 NUMA nodes with 6 cores on each.) 

It is very clear to see what happens if you comment/uncommet  case (1) in the first nested region aka this line:

       CASE(1)
          !call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C1,dim)<--- comment out this one

This bug or whatever it is, is extremely problematic for us, since it in practice makes it impossible to use nested MKL on NUMA systems. Does anyone have a solution to it? Or does Intel have an idea about when this could be resolved? 

program NumaAwareDGEMM

 use IFPORT
 use omp_lib
 use mkl_service
 implicit none

 logical(4) :: Success
 integer :: NoNUMANodes, blocksize,dim
 integer :: N,I,J
 integer :: ID
 real*8,allocatable,dimension(:,:) :: A, B,C1,c2,c3,c4,c5,c6,c7,c8
    

NoNUMANodes=6                     !How many NUMA nodes to distribute calculations over
success = SETENVQQ("OMP_DISPLAY_ENV=TRUE")
success=SETENVQQ("OMP_PLACES={0:6},{6:6},{12:6},{18:6},{24:6},{30:6}")
 

 blocksize=1000
 dim=blocksize*NoNUMANodes
 allocate(A(dim,dim))
 allocate(B(dim,dim))
 allocate(C1(dim,dim))
 allocate(C2(dim,dim))
 allocate(C3(dim,dim))
 allocate(C4(dim,dim))
 allocate(C5(dim,dim))
 allocate(C6(dim,dim))
 allocate(C7(dim,dim))
 allocate(C8(dim,dim))
 call KMP_SET_STACKSIZE_S(990000000)
 call omp_set_dynamic(0)
 call mkl_set_dynamic(0)
 call omp_set_nested(1)
 call MKL_SET_NUM_THREADS(5)

   !intialization region
   call omp_set_num_threads(NoNumaNodes) !Outer parallelization across all numanodes
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i,ID)
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      ID=omp_get_thread_num()
      print *,'Thread binding for socket=',ID
      SELECT CASE (ID)
        CASE(1)
          success=SETENVQQ("OMP_PLACES={0:6}")
        CASE(2)
          success=SETENVQQ("OMP_PLACES={6:6}")
        CASE(3)
          success=SETENVQQ("OMP_PLACES={12:6}")
        CASE(4)
          success=SETENVQQ("OMP_PLACES={18:6}")
        CASE(5)
          success=SETENVQQ("OMP_PLACES={24:6}")
        CASE(6)
          success=SETENVQQ("OMP_PLACES={30:6}")
        CASE(7)
          success=SETENVQQ("OMP_PLACES={36:6}")
        CASE(8)
          success=SETENVQQ("OMP_PLACES={42:6}")
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
    print*,'Initialization over'   
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i)  
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      SELECT CASE (i)
        CASE(1)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C1,dim)
        CASE(2)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C2,dim)
        CASE(3)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C3,dim)
        CASE(4)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C4,dim)
        CASE(5)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C5,dim)
        CASE(6)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C6,dim)
        CASE(7)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C7,dim)
        CASE(8)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C8,dim)
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
   print*,'First MKL call done'
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i)  
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      SELECT CASE (i)
        CASE(1)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C1,dim)
        CASE(2)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C2,dim)
        CASE(3)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C3,dim)
        CASE(4)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C4,dim)
        CASE(5)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C5,dim)
        CASE(6)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C6,dim)
        CASE(7)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C7,dim)
        CASE(8)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C8,dim)
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
  end program NumaAwareDGEMM

 

dasfasd

 

 

0 Kudos

Link Copied

2 Replies
TimP
Honored Contributor III
‎08-05-2015 11:18 AM
260 Views
Assuming master thread sets id to 0, it looks at first glance you're not attempting to set its affinity.
0 Kudos
Copy link
Tue_B_
Novice
‎08-05-2015 11:50 AM
260 Views

Hey Tim

You are of course right!

I posted the wrong code (I tried to change that in an attempt to fix it), anyway here is the code, now with affinity settings for the master thread. 

Now the problem has changed to the one I also observed in the previous post, namely that the master thread no longer spawns any additional threads. 

Does anyone have a solution to this bug/problem?

 

 use IFPORT
 use omp_lib
 use mkl_service
 implicit none

 logical(4) :: Success
 integer :: NoNUMANodes, blocksize,dim
 integer :: N,I,J
 integer :: ID
 real*8,allocatable,dimension(:,:) :: A, B,C1,c2,c3,c4,c5,c6,c7,c8
    

NoNUMANodes=6                     !How many NUMA nodes to distribute calculations over
success = SETENVQQ("OMP_DISPLAY_ENV=TRUE")
success=SETENVQQ("OMP_PLACES={0:6},{6:6},{12:6},{18:6},{24:6},{30:6}")
 

 blocksize=1000
 dim=blocksize*NoNUMANodes
 allocate(A(dim,dim))
 allocate(B(dim,dim))
 allocate(C1(dim,dim))
 allocate(C2(dim,dim))
 allocate(C3(dim,dim))
 allocate(C4(dim,dim))
 allocate(C5(dim,dim))
 allocate(C6(dim,dim))
 allocate(C7(dim,dim))
 allocate(C8(dim,dim))
 call KMP_SET_STACKSIZE_S(990000000)
 call omp_set_dynamic(0)
 call mkl_set_dynamic(0)
 call omp_set_nested(1)
 call MKL_SET_NUM_THREADS(5)

   !intialization region
   call omp_set_num_threads(NoNumaNodes) !Outer parallelization across all numanodes
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i,ID)
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      ID=omp_get_thread_num()
      print *,'Thread binding for socket=',ID
      SELECT CASE (ID)
        CASE(0)
          success=SETENVQQ("OMP_PLACES={0:6}")
        CASE(1)
          success=SETENVQQ("OMP_PLACES={6:6}")
        CASE(2)
          success=SETENVQQ("OMP_PLACES={12:6}")
        CASE(3)
          success=SETENVQQ("OMP_PLACES={18:6}")
        CASE(4)
          success=SETENVQQ("OMP_PLACES={24:6}")
        CASE(5)
          success=SETENVQQ("OMP_PLACES={30:6}")
        CASE(6)
          success=SETENVQQ("OMP_PLACES={36:6}")
        CASE(7)
          success=SETENVQQ("OMP_PLACES={42:6}")
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
    print*,'Initialization over'   
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i)  
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      SELECT CASE (i)
        CASE(1)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C1,dim)
        CASE(2)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C2,dim)
        CASE(3)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C3,dim)
        CASE(4)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C4,dim)
        CASE(5)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C5,dim)
        CASE(6)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C6,dim)
        CASE(7)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C7,dim)
        CASE(8)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C8,dim)
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
   print*,'First MKL call done'
   !$OMP PARALLEL DEFAULT(SHARED) PRIVATE(i)  
   !$OMP DO SCHEDULE(STATIC)
   do i = 1,NoNumanodes
      SELECT CASE (i)
        CASE(1)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C1,dim)
        CASE(2)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C2,dim)
        CASE(3)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C3,dim)
        CASE(4)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C4,dim)
        CASE(5)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C5,dim)
        CASE(6)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C6,dim)
        CASE(7)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C7,dim)
        CASE(8)
          call dgemm('N','N',dim,dim,dim,1.d0,A,dim,B,dim,0.d0,C8,dim)
      END SELECT 
   end do
   !$OMP END DO
   !$OMP END PARALLEL  
  end program NumaAwareDGEMM

 

0 Kudos
Copy link
Reply

Community support is provided during standard business hours (Monday to Friday 7AM - 5PM PST). Other contact methods are available here.

Intel does not verify all solutions, including but not limited to any file transfers that may appear in this community. Accordingly, Intel disclaims all express and implied warranties, including without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement, as well as any warranty arising from course of performance, course of dealing, or usage in trade.

For more complete information about compiler optimizations, see our Optimization Notice.