I³MS - Pitsianis Seminar
Prof. Nikos Pitsianis, Ph.D. - Dynamic Parallelism of Adaptive FMM
School of Electrical and Computer Engineering, Aristotle University, Thessaloniki and
Computer Science and Electrical and Computer Engineering, Duke University
In this talk I will introduce an adaptive scheduling strategy for the parallel execution of the fast multipole method (FMM) family of algorithms on a shared memory computing environment. With its linear complexity, the FMM is at the core of many iterative methods for a large class of structured matrices that represent pairwise interactions. The parallel FMM on irregular domain geometry and on sparse and non-uniformly distributed sampling has been a challenging tasks. The scheduling we developed for parallel FMM is adaptive to variations in the algorithm (symmetric or non-symmetric, scale variant and invariant kernels), change in data distribution (dense/uniform or sparse-nonuniform sampling locations) as well as specifics and dynamics in the computing system (relative non-uniform memory hierarchy access costs). I will present a recursive description of the FMM dependence graph, a graph-theoretic analysis of our scheduling method and experiment results with parallel FMM on multicore systems.
The talk is based on joint work with Xiaobai Sun at Duke University, Jingfang Huang at University of North Carolina, in Chapel Hill NC and Bo Zhang at Indiana University in Bloomington, IN, USA.