Improving memory performance for indirect accesses on SIMD computers

SIMD machines operate more efficiently on a wider range of problems when they have the ability to access memory with both global and local addresses. Recent work has made possible the use of caches for global addresses. This paper examines techniques for employing caches to improve memory accesses with local addresses. Specifically, we examine the improvement from utilizing a cluster-based indirect-only cache. This is a simple extension of the direct-only cache, and yet it improves indirect memory references significantly