Parallel Pearls 

     from

     Practical Parallel Programming
     by Barr E. Bauer, 1992 


 - Parallel-safe loops can be run serially in reverse
   index order and give correct results.   page 49

 - Do I/O outside parallel regions.   page 87

 - Avoid indirect indexing from within parallel
   regions.   page 95

 - Do precision-sensitive reductions serially.   page 98

 - Avoid convoluting code just to introduce 
   parallelization.   page 99

 - Proper cache management significantly contributes
   to performance.   page 104

 - Avoid requiring specific numbers of threads
   in the code.   page 113

 - Run-time bugs in programs containing local
   common blocks should be checked carefully for
   coding bugs that modify block members.  page 136

 - Local code executes on ALL threads.  page 155

 - Check pointers carefully for dependences.   page 195

 - Consolidate adjacent loops, where appropriate.   page 199

 - Isolate single-variable dependences in independent
   blocks.  page 206

 - Keep critical blocks out of pfor blocks.   page 208

 - Liberally comment directives used globally.  page 258

 - Incorrect assertions result in incorrect programs.  page 259

 - Use roundoff=2 for all code if roundoff error is 
   not important.   page 271

 - Let the clock determine the optimum mode of 
   execution for a loop nest.   page 298

 - Convert pointer arithmetic to array form within
   for loops before PCA analysis.   page 332

 - Declare parallel-safe functions in a header file using
   +pragma no side effects.   page 348

 - Let the PCA do the unrolling.   page 370

 - Fuse adjacent loops, when appropriate.   page 373

 - Parallel program debugging is far more difficult than
   serial program debugging.   page 419



Other "Parallel Pearls" that I like from a different source:

Origin 2000 Quick Reference: Multiprocessor Tuning

 - Step One: Tune Single Processor Performance
             See other side. 

 - Step Two: Parallelize Code

 - Step Three: Identify Bottlenecks

 - Step Four: Fix False Sharing
  
 - Step Five: Tune for Data Placement

Origin 2000 Quick Reference: Single-Processor Tuning (other side)

 - Step One: Get the Right Answer

 - Step Two: Use Existing Tuned Code

 - Step Three: Find Out Where to Tune

 - Step Four: Get the Compiler to Do the Work

 - Step Five: Tune Cache Performance