A Strategy for Parallel Implementations of Stochastic Lagrangian Simulation

Abstract

International audienceIn this paper, we present some investigations on the parallelization of stochastic Lagrangian simulations. The challenge is the proper management of the random numbers. We review two different object-oriented strategies: to draw the random numbers on the fly within each MPI's process or to use a different random number generator for each simulated path. We show the benefits of the second technique which is implemented in the PALMTREE software developed by the Project-team Sage of Inria Rennes. The efficiency of PALMTREE is demonstrated on two classical examples

Optimizing parallel particle tracking in Brownian motion using machine learning

“…Lagranginan tracking is an essential technique in numerous research areas, some of which employ Monte-Carlo-based models. In order to speed up the simulation, many of those research studies such as Roberti et al (2005), Lenôtre (2016), Larson and Nasstrom (2002), Charles et al (2008), Breuer et al (2006) and Beaudoin et al (2007) addressed the parallelization of particle tracking algorithms. In most cases, the authors suggest two strategies of decomposition: (i) Domain Decomposition, in which the space integration domain is partitioned into smaller volumes, and each volume (subdomain) is addressed to a different processor as demonstrated by Roberti et al (2005), and (ii) Decomposition Per Particle (DPP), where each processor carries out a certain number of particles throughout their lifetime.…”

“…The second decomposition approach does not involve any additional communication burden, as stated above, but if particles are characterized by highly variable computation time, that still brings unequally balanced processor utilization. Lenôtre (2016), Roberti et al (2005 and Larson and Nasstrom (2002) introduce a parallel approach based on decomposition per particle (DPP) to avoid any additional communication required in the domain decomposition. To the best of our knowledge, none of them considers the problem of reduced parallel efficiency.…”

“…Generating a trajectory of a single particle is independent from generating other particles trajectories, which at the first glance makes this task trivial to parallelize by a simple static decomposition. By using Decomposition Per Particle (DPP), each processor is responsible for modeling trajectories of approximately equal number of particles, as demonstrated by Lenôtre (2016), Roberti et al (2005) and Larson and Nasstrom (2002). When solving problems in which the particle terminates its motion by decay or deposition, this approach shows the downside in reduced parallel efficiency caused by variability of computation time necessary to generate trajectories of individual particles.…”

How Much Energy Can Green HPC Cloud Users Save?

“…3) Palmtree Workflow: Palmtree [16] is a library for the parallelization of Monte Carlo methods where the challenge is the proper management of the random numbers. The workflow structure is composed of 2 parallel tasks and its execution is CPU-intensive only.…”