Asynchronous Task Scheduling of the Fast Multipole Method Using Various Runtime Systems

Abstract

In this paper, we explore data-driven execution of the adaptive fast multipole method by asynchronously scheduling available computational tasks using Cilk, C++11 standard thread and future libraries, the High Performance ParalleX (HPX-5) library, and OpenMP tasks. By comparing these implementations using various input data sets, this paper examines the runtime system's capability to spawn new task, the capacity of the tasks that can be managed, the performance impact between eager and lazy thread creation for new task, and the effectiveness of the task scheduler and its ability to recognize the critical path of the underlying algorithm.