Staggering and blocking: a technique to optimize the parallel programs for large scale parallel processing systems

We analyze the thrashing due to false sharing of data in various array access patterns and coexisting phenomenon of thrashings due to true and false sharing of data, introduce a concept of access interval for parallel loops, and propose a method called blocking and staggering to eliminate the false sharing thrashing at compile time. Combined with the blocking and staggering, we present an idea of array expansion for multidimensional arrays with the purpose of simplifying the discussion of the false sharing thrashing in complex access patterns and making it possible to block outer loops to eliminate the thrashing. Finally, experimental results are given.<
>