IV31593: PERFORMANCE PROBLEM WITH XLSMPOPTS BINDING OPTION

November 2012 PTF for XL C for AIX, V12.1
November 2012 PTF for XL C/C++ for AIX, V12.1
November 2012 PTF for XL Fortran for AIX, V14.1
November 2012 PTF for XL Fortran Runtime for AIX, V14.1
March 2013 PTF for XL C for AIX, V12.1
March 2013 PTF for XL C/C++ for AIX, V12.1
March 2013 PTF for XL Fortran for AIX, V14.1
March 2013 PTF for XL Fortran Runtime for AIX, V14.1

APAR status

• Closed as program error.

Error description

• A runtime performance degradation is seen when using the
  XLSMPOPTS binding option as follows:
  
  =====Compile Line:
  xlf_r -O -q64 -qsmp=omp matmul.f -o matmul.exe
  
  =====Testcase:
  $ cat matmul.f
       program matmul1
        implicit none
        real*8, allocatable, dimension(:,:,:) :: a,b,c
        integer, parameter :: n = 4000
        integer i
  
        allocate(a(n,n,0:1),b(n,n,0:1),c(n,n,0:1))
        a = 1.d0
        b = 1.d0
  !$OMP parallel do private(i)
        do i = 0,1
         c(:,:,i) = matmul(a(:,:,i),b(:,:,i))
        enddo
        deallocate(a,b,c)
        end
  $
  
  ====Runtime output:
  #!/bin/ksh
  set -x
  export XLFRTEOPTS=intrinthds=1
  export XLSMPOPTS=parthds=2
  time ./matmul.exe
  export XLSMPOPTS=parthds=2:bindlist=proc=0,2
  time ./matmul.exe
  
  The following are from a Power7 booted in SMT-2 mode (3.56GHz)
  
  $ ./run
  + export XLFRTEOPTS=intrinthds=1
  + export XLSMPOPTS=parthds=2
  + ./matmul.exe
  
  real    0m27.04s
  user    0m45.03s
  sys     0m0.00s
  + export XLSMPOPTS=parthds=2:bindlist=proc=0,2
  + ./matmul.exe
  
  
  
  real    1m6.88s
  user    0m57.23s
  sys     0m0.00s
  

Local fix

• N/A
  

Problem summary

• USERS AFFECTED:
  Users using the bind and bindlist options of XLSMPOPTS are
  potentially affected by this issue.
  
  PROBLEM DESCRIPTION:
  When using the bind or bindlist options of XLSMPOPTS, the
  compiler does not respect it and binds all threads to one CPU
  causing a performance degradation during runtime.
  

Problem conclusion

• The issue was in the binding code within the compiler that binds
  threads to different CPUs. This has been fixed.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  XL SMP AIX

• Fixed component ID

  5725C7401

Applicable component levels

• R310 PSY

     UP