PSTSWM T3E-900 Algorithm Comparison

Performance Studies using

PSTSWM

SGI/Cray Research T3E-900 Protocol Performance

(distributed LT experiment I-B )

Date/Person: May 11, 1998 / P. Worley
Platform: T3E-900 at National Energy Research Scientific Computing Center (mcurie.nersc.gov):
   532 450-MHz DEC Alpha EV5 RISC processors
Environment: UNICOS/mk 2.0.2.28
mpt 1.2.0.2
Cray CF90 Version 3.0.2.1
Code Version: 6.3
Compilation Options: -dp -Oscalar3 -Wl,-Dstreams=on
Math Library: libsci
Communication Library: MPI
SHMEM
Number of Timesteps: 12
Results:

Distributed LT (1) (mpi) / Row Major Processor Ordering
Algorithm Comparison
  T42L2     T42L2     T42L16    T85L4     T85L4     T85L32 
  P=2x32     P=4x16     P=8x8     P=2x32     P=4x16     P=8x8  
  optimal algorithm   halfsum  halfsum  halfsum  ringpipe  halfsum  halfsum 
  (allreduce-min)/min     0.337    0.093    0.066    0.453    0.192    0.077 
  (generic-min)/min     0.416    0.192    0.022    0.105    0.050    0.015 

Distributed LT (1) (shmem) / Row Major Processor Ordering
Algorithm Comparison
  T42L2     T42L2     T42L16    T85L4     T85L4     T85L32 
  P=2x32     P=4x16     P=8x8     P=2x32     P=4x16     P=8x8  
  optimal algorithm   halfsum  halfsum  halfsum  ringpipe  ringpipe  halfsum 
  (generic-min)/min     0.414    0.238    0.035    0.336    0.187    0.021 

Distributed LT (1) (combined) / Row Major Processor Ordering
Communication Library Comparisons
  T42L2     T42L2     T42L16    T85L4     T85L4     T85L32 
  P=2x32     P=4x16     P=8x8     P=2x32     P=4x16     P=8x8  
  optimal library   shmem  shmem  shmem  shmem  shmem  shmem 
  (allreduce-min)/min     1.020    0.726    0.328    0.944    0.518    0.159 
  (mpi-min)/min     0.512    0.580    0.246    0.338    0.273    0.076 

Patrick H. Worley / ( worleyph@ornl.gov)
Last Modified Monday, 15-Jul-2002 10:24:38 EDT.