Self-distributing computation

Abstract

In this thesis, I propose a new model for distributing computational work in a parallel or distributed system. This model relies on exposing the topology and performance characteristics of the underlying architecture to the application. Responsibility for task distribution is divided between a run-time system, which determines when tasks should be distributed or consolidated, and the application, which specifies to the runtime system its first-choice distribution based on a representation of the current state of the underlying architecture. Discussing my experience in implementing this model as a Java-based simulator, I argue for the advantages of this approach as they relate to performance on changing architectures and ease of programming.