Implementation of OpenMP Work-Sharing on the Cell Broadband Engine Architecture

Abstract

The Cell Broadband Engine (CBE) is a single-chip multiprocessor composed of one Power Processor Element (PPE) core and multiple Synergistic Processor Element (SPE) cores. Due to the heterogeneous processor type of the CBE, multi-core programming for the CBE is very difficult which can be alleviated by OpenMP. In this paper, we introduce some issues in implementing OpenMP work-sharing constructs on the CBE. To address overheads in creating and managing work-sharing control blocks across different processor types, we have come up with a method for aggressively reusing the work-sharing control block by precisely checking the reusability of it. This scheme allows us better performance over naive implementation of work-sharing constructs, when evaluated with modified EPCC benchmark.