Accelerating Parallel Monte Carlo Simulations for Statistical Physics: Portability on Many-Core Processors
SessionResearch Posters Display
Event Type
Posters
Research Posters
TP
XO / EX
TimeThursday, 18 November 20218:30am - 5pm CST
LocationSecond Floor Atrium
DescriptionMonte Carlo (MC) simulations are important tools for studying thermodynamics and materials properties at finite temperature. Scaling up to larger systems allows for simulations comparable to experimental length scale. In this work, we focus on improving the weak scaling performance to simulate large magnetic spin systems up to millions of atoms. We parallelize the computation of physical observables within a single random walker using checkerboard domain decomposition implemented with OpenMP and compare the performance of the same code base on different many-core processors. We analyze the effects of thread affinity and hyper-threading on the runtime performance and achieve 14x/12x/6x speedups for Arm/Intel/AMD architectures respectively for the largest system we report in this work (about 4.2 millions spins). This study is helpful in informing end-users on the choice of the best multi-core architecture suited to their needs for similar pattern of parallel applications.
Archive view
