Dynamic Computational Load Balancing During Simulations of Instability Flows

Abstract

In this work we describe the algorithm of dynamic computational load balancing during modeling of instability flows over dynamically refined mesh using three dimensional eulerian gas dynamic program. Balancing of computational load between MPI-fragments is executed along lines (set of MPI-fragments that are situated along preferential direction) independently and may be carried out in three regimes: accounting number of cells, accounting real calculation time, and automatic (or composite) regime. We use adaptation with reservation to minimize overheads for adaptation and balancing: we construct a sequence of adaptation, balancing and calculation stages during fixed count of steps – "Adaptation — Balancing — Calculation" cycle (A–B–C cycle). The optimal count of steps in A–B–C cycle is determined by minimization of task calculation time that depends on count of cells. For testing of realized algorithm we considered a spherical implosion of a light material by a dense shell (Youngs’ problem). A source of instabilities was the initial harmonic perturbation on the contact boundary. The calculations are made using second-level dynamically adaptive mesh refinement in the region of instability growth. We analyze the calculation time of the problem using different regimes of load balancing algorithm. The result of the research proves the efficiency of using adaptive mesh refinement with reservation within A–B–C cycle. The automatic regime of load balancing is the most stable for three-dimensional modelling.