Efficient Compile-Time/ Run-Time Contraction of Fine Grain Data Parallel Codes

This research studies the contraction problem for data parallel languages. That is, we seek to efficiently execute code written for a fine grain virtual parallel machine on a much coarser grain actual parallel machine. This paper identifies issues involved in solving this problem. Three issues are then addressed in detail: efficiently implementing tasks to emulate virtual processors, efficiently scheduling these tasks, and reducing space overhead while maintaining data consistency among these tasks. In implementing tasks, saving a reduced register state at context switches is addressed. For scheduling tasks, heuristics are proposed that minimize scheduling cost and promote data locality. Minimizing space overhead using assumptions about the communication paradigm, run-time techniques and compile-time analysis is also discussed. Finally, experimental results are presented concerning one of the main issues discussed, scheduling. The results show that the proposed scheduling heuristics are both cost efficient and promote data locality for three different problems when compared to three other scheduling policies.