CPL - Chalmers Publication Library
| Utbildning | Forskning | Styrkeområden | Om Chalmers | In English In English Ej inloggad.

Dynamic Load Balancing Based on Rectilinear Partitioning in Particle-in-Cell Plasma Simulation

I. Surmin ; A. Bashinov ; S. Bastrakov ; E. Efimenko ; Arkady Gonoskov (Institutionen för teknisk fysik, Kondenserade materiens teori) ; I. Meyerov
Lecture Notes in Computer Science. 13th International Conference on Parallel Computing Technologies, PaCT 2015, Petrozavodsk, Russian Federation, 31 August - 4 September 2015 (0302-9743). Vol. 9251 (2015), p. 107-119.
[Konferensbidrag, refereegranskat]

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.

Nyckelord: Load balancing, High performance computing, Plasma simulation, Particle-in-cell



Denna post skapades 2015-11-24. Senast ändrad 2015-12-08.
CPL Pubid: 226230

 

Läs direkt!


Länk till annan sajt (kan kräva inloggning)


Institutioner (Chalmers)

Institutionen för teknisk fysik, Kondenserade materiens teori (1900-2015)

Ämnesområden

Programvaruteknik

Chalmers infrastruktur