Efficient Synchronization for Stencil Computations Using Dynamic Task Graphs
Zubair Wadood Bhatti, Roel Wuyts, Pascal Costanza, Davy Preuveneers, Yolande Berbers · Procedia Computer Science · 2013
Executing stencil computations constitutes a significant portion of execution time for many numerical simulations running on high performance computing systems. Most parallel implementations of these stencil operations suffer from a substantial synchronization overhead. Furthermore, with the rapidly increasing number of cores these synchronization costs keep rising. This paper presents a novel approach for reducing the synchronization overhead of stencil computations by leveraging dynamic task graphs to avoid global barriers and minimizing spin-waiting, and exploiting basic properties of stencil operations to optimize the execution and memory management. Our experiments show a reduction in synchronization overhead by at least a factor four when compared to state-of-the-art stencil compilers like Pochoir and Patus.