Optimizing Skeletal Stream Processing for Divide and Conquer

Algorithmic skeletons intend to simplify parallel programming by providing recurring forms of program structure as predefined components. We present a new distributed task parallel skeleton for a very general class of divide and conquer algorithms for MIMD machines with distributed memory. Our approach combines skeletal internal task parallelism with stream parallelism. This approach is compared to alternative topologies for a task parallel divide and conquer skeleton with respect to their aptitude of solving streams of divide and conquer problems. Based on experimental results for matrix chain multiplication problems, we show that our new approach enables a better processor load and memory utilization of the engaged solvers, and reduces communication costs.