Dynamic Load Balancing Based on Rectilinear Partitioning in Particle-in-Cell Plasma Simulation
Paper in proceeding, 2015
This paper considers load balancing in Particle-in-Cell plasma simulation on cluster systems. We propose a dynamic load balancing scheme based on rectilinear partitioning and discuss implementation of efficient imbalance estimation and rebalancing. We analyze the impact of load balancing on performance and accuracy. On a test plasma heating problem dynamic load balancing yields nearly 2 times speedup and better scaling. On the real-world plasma target irradiation simulation load balancing allows to mitigate particle resampling and thus improve accuracy of the simulation without increasing the runtime. Balancing-related overhead in both cases are under 1.5% of total run time.
Load balancing
Particle-in-cell
Plasma simulation
High performance computing
Author
I. Surmin
Lobachevsky University
A. Bashinov
Lobachevsky University
Russian Academy of Sciences
S. Bastrakov
Lobachevsky University
E. Efimenko
Lobachevsky University
Russian Academy of Sciences
Arkady Gonoskov
Chalmers, Applied Physics, Condensed Matter Theory
I. Meyerov
Lobachevsky University
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
03029743 (ISSN) 16113349 (eISSN)
Vol. 9251 107-119978-3-319-21909-7 (ISBN)
Subject Categories
Software Engineering
DOI
10.1007/978-3-319-21909-7_12
ISBN
978-3-319-21909-7