Efficient pseudo-random number generators for biomolecular simulations on graphics processors

TL;DR

This paper compares two main GPU-based approaches for implementing pseudo-random number generators essential for biomolecular simulations, evaluating their performance and statistical quality through molecular dynamics tests.

Contribution

It introduces and benchmarks GPU implementations of RNGs using one-RNG-per-thread and one-RNG-for-all-threads approaches for biomolecular simulations.

Findings

GPU RNGs achieve significant speedups over CPU implementations.

The statistical quality of RNGs is validated through molecular dynamics simulations.

Performance profiling shows trade-offs between speed, memory, and randomness quality.

Abstract

Langevin Dynamics, Monte Carlo, and all-atom Molecular Dynamics simulations in implicit solvent, widely used to access the microscopic transitions in biomolecules, require a reliable source of random numbers. Here we present the two main approaches for implementation of random number generators (RNGs) on a GPU, which enable one to generate random numbers on the fly. In the one-RNG-per-thread approach, inherent in CPU-based calculations, one RNG produces a stream of random numbers in each thread of execution, whereas the one-RNG-for-all-threads approach builds on the ability of different threads to communicate, thus, sharing random seeds across the entire GPU device. We exemplify the use of these approaches through the development of Ran2, Hybrid Taus, and Lagged Fibonacci algorithms fully implemented on the GPU. As an application-based test of randomness, we carry out LD simulations of N independent harmonic oscillators coupled to a stochastic thermostat. This model allows us to assess statistical quality of random numbers by comparing the simulation output with the exact results that would be obtained with truly random numbers. We also profile the performance of these generators in terms of the computational time, memory usage, and the speedup factor (CPU/GPU time).