Donald Bren School of Information and Computer Sciences

Percolation Scheduling, a technique for compile-time code parallelization, has proven very successful for exploiting fine-grain irregular parallelism in ordinary programs. Currently, this technology is targeted only to VLIW (Very Long Instruction Word) machines, which have the advantages of 'free' synchronization and communication. Shared memory multi-processors can simulate the execution characteristics of VLIW machines with the use of static barriers. Preliminary results show that Percolation Scheduling can be used with good results on this type of architecture by increasing the granularity from operation level to source statement level, removing any redundant synchronization, and providing an efficient implementation of multi-way jumps.