Efficient execution of fine-grain parallelism on a tightly-coupled multiprocessor

Abstract

In multiprocessor systems, the overheads caused by inter-processor synchronization and communication continue to be impediments to the efficient execution of parallel programs. Reduction of these types of overhead is necessary in systems that focus on large-scale and fine-grain parallelism. This paper proposes a Fine-Grain Multi-Processor (FGMP) based on a shared-memory/shared-bus architecture, which can efficiently handle fine-grain concurrency in parallel. New strategies for management of hardware resources in the system are discussed, and two innovative hardware mechanisms are proposed that work well for fine-grain parallelism with the above strategies: Elastic Barrier (a light synchronization mechanism), which is derived from a generalization of a barrier-type mechanism, and an Inter-Cache Snoop Control Mechanism that switches snoop-protocols dynamically to reduce the overhead associated with shared data handling. After introducing the FGMP system, which incorporates the above strategies and mechanisms, the paper closes with a discussion of the FGMP's characteristics and efficiency.