Scaling up smoothed particle hydrodynamics simulations with GPUs

Astrophysical simulations require significant computational resources. Most high-performance supercomputing clusters nowadays include a complement of graphics processing units (GPUs), as these have been shown to be highly efficient in terms of computation for the amount of energy consumed. Using only the CPUs (central processing unit) of a supercomputer leave the GPUs underutilized, wasting their potential. We plan to write algorithms that make efficient use of GPUs for astrophysical smoothed particle hydrodynamics simulations. Working in an astrophysics context introduces unique challenges due to the vast range of length and time scales involved. One strategy is to use individual timesteps, where portions of the simulation can evolve on faster timescales than other regions, which can afford orders of magnitude savings in computational expense. We focus on the problem of writing GPU efficient code for individual timesteps, with the goal of creating a hybrid CPU / GPU astrophysical simulation software.

Faculty Supervisor:

Terrence Tricco

Student:

Partner:

Kobe University

Discipline:

Computer science

Sector:

Technology

University:

Memorial University of Newfoundland

Program: