A solution of the cache ping-pong problem in multiprocessor systems

Abstract

The cache ping-pong problem arises often in parallel processing systems where each processor has its own local cache and employs a copy-back protocol for the cache coherence. To solve the problem of large amounts of data moving back and forth between the caches in different processors, techniques associated with parallel compilers need to be developed. Based on the concept in [Fang, J. Z., Proc. International Conference on Parallel Processing, Aug. 1990, pp. II-271–II-275] regarding the relations between array element accesses and enclosed loop indices in nested parallel loops, we present an algorithm in this paper to reduce the unnecessary data movement between the caches for parallel loops with multiple array subscript expressions. By analyzing the array subscript expressions in the nested parallel loop constructs, the compilers can use the algorithm to prepare information at compile time and let the processor execute the corresponding iterations of parallel loops in terms of the data in its cache. It benefits the parallel programs in which parallel loops are enclosed by a sequential loop and have multiple different subscript expressions for the same array, whose elements are repeatedly used in the different iterations of the outermost sequential loop.