Conclusions

• Problem: the marker scheme can use a lot of memory
• Proposed: identify sets of goals which are a ``backtracking family''
  • Little additional run-time machinery needed
  • Less memory usage
  • Possibility of better speed-ups
  • Does not interfere with flexible scheduling and general backtracking

• Problem: slow task startup for fine-grain computations
• Proposed: distribute task startup among processors
  • Source-to-source transformation can support faster task-startup
  • Dynamic unfolding yields good results for lists (even with the associated overhead)

• Other interesting optimizations proposed

• Further work
  • Formal characterization of backtracking families
  • Find out suitable analysis framework and run-time conditions
  • Devise low-level support for indexed task creation