Particle-in-cell (PIC) plasma simulation techniques consist in following the continuous phase-space evolution of a simulation-particle population, while taking into account the mutual interaction between each pair of simulation particles by means of an electromagnetic field, computed only at the points of a discrete spatial grid and then interpolated at each particle position. A parallelization strategy is discussed, which consists in distributing statically the particle population among the processors, while assigning the whole domain to each processor. Partial contributions to the total pressure at the grid points, which is required to update the electromagnetic fields, are then communicated among processors and summed together. Different from standard domain-decomposition techniques, the amount of difficulties in translating the serial version of the code into the parallel one is very contained and load balancing is automatically ensured; on the opposite side, the physical-space resolution is limited by the RAM resources of the single processor, and the intrinsic scalability of the spatial resolution with the number of processors is lost. Such strategy is then particularly suited for target architectures with a moderate number of nodes. Results obtained with a HPF parallel version of a hybrid MHD-gyrokinetic PIC code running on an IBM Scalable Power system (SP2) are presented; both the number of processors nprocs (ranging from 2 to 8) and the average number of particle per cell NPPC≡Nparticle/Ncell(ranging from 2 to 1024) have been varied. Efficiency (≡speed-up/nprocs) reaches its ideal value (= 1) at larger values of NPPC the greater the number of processors is: increasing the number of processors the average number of particles per cell assigned to each processor decreases, and the importance of the grid-related calculation tends to overcome the particle-related one. In fact, when plotted against NPPC/nprocs, the efficiency values approximatively fall on a `universal' curve.
|Pages (from-to)||696 -|
|Journal||Computer Physics Communications|
|Publication status||Published - 1999|
All Science Journal Classification (ASJC) codes
- Hardware and Architecture
- Physics and Astronomy(all)