Load Balancing

• Balance1 (source):
  inner loops get smaller ("triangular matrix")
  unbalanced distribution of workload
  
• Balance2 (source):
  SCHEDULE(STATIC, blocksize)
  partition index range in blocks of fixed length (blocksize)
  blocks are dealt out to threads in round robin way
  
• Balance3 (source):
  SCHEDULE(DYNAMIC, blocksize)
  partition index range in blocks of fixed length blocksize (as above)
  when a thread completes a block, it gets the next one
  important for loops with strongly varying times per iteration

  • (even better load balancing in our example)