RIKEN Advanced Institute for Computational Science

Menu
Menu
OVERVIEW 計算科学研究機構とは

Particle Simulator Research Team

Framework for the Simplified Development of High-performance Particle-based Simulation Programs

Systems modeled in numerical simulations can be classified as regular grid, irregular grid, and particle systems. As the computational environment evolves, ever larger simulations have become possible. Thus, particle-based methods which can handle complex and dynamically changing shapes without special treatment, have become more important. However, it is a challenge to develop efficient parallel programs for particle-based simulations on large parallel systems such as the K or post-K computers. We need to efficiently implement domain decomposition with near-ideal load balancing, exchange particles between domains, and improve interaction calculations of particles in different domains. On the other hand, all parallel particle simulation-programs require these functions.

Consequently, we are developing the Framework for Developing Particle Simulators (FDPS), which can be used to produce any kind of particle-based simulation. FDPS receives particle data structures and functions, evaluates interactions between particles, and generates efficient libraries for parallelization. Thus, users of FDPS can develop their own highly efficient parallel particle simulation-programs without spending large amounts of time, and so can concentrate on realizing their goals.

Recent Achievement

World’s first highly efficient general-purpose framework for particle-based simulations
By employing FDPS, we have succeeded in creating application programs for particle-based simulations not specifically optimized for large-scale parallel computing to run efficiently on top-ranked supercomputers such as the K computer. On the K computer, we achieved efficiency gains of around 50% for gravitational many-body simulations and for SPH simulations, at up to full-node execution. FDPS also supports GPGPU systems, and application programs can be developed both with C++ and the Fortran languages. We are currently working on a highly efficient implementation for heterogeneous many-core systems.

FDPS has already been used for various large-scale simulations, including the largest simulation ever of the giant impact hypothesis used to explain the moon’s formation, and the first full simulation of rings around asteroids. These simulations would have been impossible without FDPS.

Largest simulation of the giant impact hypothesis used to explain the moon’s formation

Connections

Team Leader Junichiro Makino

Team Leader
Junichiro Makino

Biography: Detail
Annual Report

FY2015 RIKEN AICS Annual Report
(PDF 246KB)
FY2014 RIKEN AICS Annual Report
(PDF 1.60MB)
FY2013 RIKEN AICS Annual Report
(PDF 833KB)
FY2012 RIKEN AICS Annual Report
(PDF 1.16MB)