Dynamic Load Balancing for Large Scale Particle Simulations

Beitrag bei einer Tagung
(Abstract zum Vortrag)


Details zur Publikation

Autorinnen und Autoren: Eibl S, Schornbaum F, Rüde U
Jahr der Veröffentlichung: 2018
Sprache: Englisch


Abstract

Current and future applications of rigid body dynamics involve an ever growing number of particles. Example
applications are in the pharmaceutical industries, chemical industries or additive manufacturing. Simulations of
billions of particles can only be performed by using specialized software. This software must use today’s largest
supercomputers efficiently. In these highly parallel environments special attention must be paid to the load balance
between the processes. Since the overall performance is dependent on the slowest process overloaded processes
will lead to a decrease of the performance. Typically rigid body dynamics algorithms are parallelized by partition-
ing the domain spatially into subdomains. These are subsequently assigned to processes. Each process evolves the
particles contained in its subdomain(s) in time. The domain partitioning must be chosen carefully. The workload
associated with each subdomain is related to the particles within. Deciding on the domain partitioning at the begin-
ning of a simulation and using it throughout the whole simulation works well for homogeneous setups where the
particle density does not fluctuate much. But typical rigid body dynamics simulations involve a huge number of
moving particles. This may change the workload of the subdomains constantly throughout the simulation. There-
fore a static domain partitioning that achieved a good load balance at the beginning of the simulation might not be
appropriate later on. To guarantee good performance over the whole simulation the software framework needs to
dynamically rebalance the workload. It has to update its domain partitioning and redistribute it at regular intervals.
This allows to avoid unbalanced workloads. Many popular software packages like LIGGGHTS, DEMOOP, ls1
mardyn already incorporate various load balancing strategies [1, 2, 3]. For fluid dynamics simulations good scaling
behaviour has already been reported for diffusion based load balancing algorithms [4, 5].
In this work, we study the applicability of load balancing algorithms based on space filling curves [6] and
diffusion based algorithms [7] for rigid body dynamics simulations. We investigate the scaling behaviour of these
algorithms on the state-of-the-art supercomputer Juqueen (TOP500 rank: 21) located at the Jlich Supercomputing
Centre. This supercomputer allows simulations with more than one million processes which we use to determine
the usability of these algorithms for future large scale simulations. Preliminary results for Hilbert space filling
curves presented in Fig. 1 show suboptimal scaling behaviour. The performance of the load balancing starts to
drop significantly above 8192 processes which prohibits scaling up to the full machine (1.8 million processes).
Due to inherent global gather operations needed for space filling curves performance decreases for a large number
of processes. We will also try to reproduce the results obtained for diffusive load balancing algorithms in fluid
dynamics [5] in rigid body dynamics.
References
[1] R. Berger, C. Kloss, A. Kohlmeyer, and S. Pirker, “Hybrid parallelization of the LIGGGHTS open-source DEM code,”
Powder Technology, vol. 278, pp. 234–247, July 2015.
[2] D. T. Cintra, R. B. Willmersdorf, P. R. M. Lyra, and W. W. M. Lira, “A hybrid parallel DEM approach with workload
balancing based on HSFC,” Engineering Computations, vol. 33, pp. 2264–2287, Nov. 2016.
[3] A. Heinecke, W. Eckhardt, M. Horsch, and H.-J. Bungartz, Supercomputing for Molecular Dynamics Simulations: Han-
dling Multi-Trillion Particles in Nanofluidics. Springer, 2015.
[4] F. Schornbaum and U. Rüde, “Massively parallel algorithms for the lattice boltzmann method on NonUniform grids,”
SIAM Journal on Scientific Computing, vol. 38, pp. C96–C126, Jan. 2016.
[5] F. Schornbaum and U. Rüde, “Extreme-scale block-structured adaptive mesh refinement,” arXiv preprint
arXiv:1704.06829, 2017.
[6] P. M. Campbell, K. D. Devine, J. E. Flaherty, L. G. Gervasio, and J. D. Teresco, “Dynamic octree load balancing using
space-filling curves,” Williams College Department of Computer Science, Tech. Rep. CS-03-01, 2003.
[7] G. Cybenko, “Dynamic load balancing for distributed memory multiprocessors,” Journal of Parallel and Distributed
Computing, vol. 7, pp. 279–301, Oct. 1989.


FAU-Autorinnen und Autoren / FAU-Herausgeberinnen und Herausgeber

Eibl, Sebastian
Lehrstuhl für Informatik 10 (Systemsimulation)
Rüde, Ulrich Prof. Dr.
Lehrstuhl für Informatik 10 (Systemsimulation)
Schornbaum, Florian
Lehrstuhl für Informatik 10 (Systemsimulation)


Zitierweisen

APA:
Eibl, S., Schornbaum, F., & Rüde, U. (2018). Dynamic Load Balancing for Large Scale Particle Simulations. Paper presentation at The 5th Joint International Conference on Multibody System Dynamics, Lisboa, PT.

MLA:
Eibl, Sebastian, Florian Schornbaum, and Ulrich Rüde. "Dynamic Load Balancing for Large Scale Particle Simulations." Presented at The 5th Joint International Conference on Multibody System Dynamics, Lisboa 2018.

BibTeX: 

Zuletzt aktualisiert 2019-24-04 um 06:21