Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 41: Statistical Physics: General
DY 41.9: Vortrag
Donnerstag, 21. März 2024, 11:45–12:00, BH-N 128
Population annealing on massively-parallel and distributed compute hardware — •Denis Gessert1,2, Wolfhard Janke1, and Martin Weigel3 — 1Institut für Theoretische Physik, Universität Leipzig, 04081 Leipzig — 2Centre for Fluid and Complex Systems, Coventry University, Coventry CV1 5FB, UK — 3Institut für Physik, Technische Universität Chemnitz, 09107 Chemnitz
Population annealing (PA) is a Monte Carlo (MC) method well suited for problems with a rough free energy landscape such as glassy systems. A PA simulation starts by equilibrating R replicas at an easy-to-sample high temperature. Akin to simulated annealing (SA) of a single system, the replicas are collectively cooled down to an otherwise hard-to-sample low temperature. Unlike in SA, before MC updates at the next lower temperature a selection process is carried out in PA, which makes the algorithm less prone to trapping in metastable states. The population size required to reach equilibrium strongly depends on the “difficulty” of the studied system, with some spin-glass simulations requiring populations of several million replicas. Despite the immediate technical challenge, this opens up an opportunity of achieving a level of parallelism that grows with the difficulty of the task.
Here, we present a simple replica-redistribution protocol for a distributed compute architecture that significantly reduces network traffic as compared to previous approaches, thus improving performance. For small instances, our protocol is only slightly worse than the optimal redistribution protocol found by brute-force. Finally, in some cases a form of speculative execution can be used to hide the network latency.
Keywords: Monte Carlo simulations; Population Annealing; HPC; Parallel Computing