Thread migration in distributed memory multicomputers

Abstract

While the paradigm offered by SMP designs is a relatively clean one, programming paradigms offered on distributed memory platforms rarely offer the same simplicity of comprehension, and ease of use. The most common paradigms for programming distributed memory computers offer either distributed shared memory, or a complex and error prone message passing library. We have implemented a runtime system for distributed memory platforms. NOMAD which offers not only the transparency of data location provided by distributed shared memory systems, but also transparency of processing location by a highly optimised and lightweight thread migration mechanism. We argue that thread migration is in many cases actually a more efficient computation strategy than the data migration in conventional distributed shared memory systems.