slide 37 of 37

Optimizing programs

• Don't use parallelism unless it helps...
  Watch the grain size and serialize where parallel grain size would be too small
• Try to minimize communication frequency
  • Keep multiple copies of read-only data
  • Clump/vectorize groups of small communications
  • Duplicate "cheap" computations

The next set of slides is available from http://yara.ecn.purdue.edu/~pplinux/ICPP96/