Local Speculative Evaluation for Distributed Graph Reduction

Abstract

In a parallel graph reduction system, speculative evaluation can increase parallelism by performing potentially useful computations before they are known to be necessary. Speculative computations may be coded explicitly in a program, or they may be scheduled implicitly by the reduction system as idle processors become available. A general approach to both kinds of speculation incurs a great deal of overhead that may outweigh the benefits of speculative evaluation for fine-grain speculative tasks.

The basic principle of local speculation is to permanently bind all implicit speculative computations to the sparking processor. Should all local mandatory tasks become blocked, local speculation offers a low- cost alternative to task migration. Restricting speculation to the local processor simplifies the problems of speculative task management, and opens the door for fine-grain speculative tasks.

Though there are fewer opportunities for local speculation than for more general speculation, local speculation can often make use of the same idle processor time that would normally trigger task migration. For distributed graph reduction systems, local speculation may prove to be a worthwhile, low-cost alternative to potentially expensive task migration.