Data Distribution in Orca through Compiler Optimization

摘要

Orca is a language for writing parallel applications that run on distributed systems. Processes in Orca communicate through user-defined operations on shared data-objects. The communication model is therefore based on logically shared data, not on message passing. The performance of applications written in Orca depends strongly on how shared data-objects are distributed among the local physical memories of the different processors. The most important distribution technique Orca uses is replication of objects. Determining which objects to replicate and where to store nonreplicated objects is difficult, as it depends on how each process uses each object. The paper studies how compiler optimization can help to make these decisions. It describes the implementation of an optimizing Orca compiler. This compiler first computes the access patterns of all processes to all shared objects. Next, it uses these patterns to determine which objects should be replicated, as well as where to store nonreplicated objects.