||CMES: Computer Modeling in Engineering & Sciences, Vol. 20, No. 1, pp. 1-10, 2007
||Full length paper in PDF format. Size = 229,043 bytes
||parallel-in-time, parallel efficiency, Navier-Stokes, finite volume.
||In this paper, we discuss the efficiency and speed-up of parallel-in-time calculations of the unsteady incompressible Navier-Stokes equations in a PC-cluster. The parallel-in-time method is based on the alternate use of coarse global sequential solvers with fine local parallel ones in an iterative predictor-corrector fashion. Therefore, the efficiency of parallel calculations is strongly dependent on the number of iterations required for convergence. The one-dimensional scalar transport equation and the two-dimensional incompressible unsteady form of the Navier-Stokes equations were used to conduct numerical experiments to derive some conclusions concerning the accuracy and convergence of the iterative method. A simple performance model is proposed to estimate the efficiency of the parallel calculations as a function of the most relevant parameters that contribute to the computing time required to perform a parallel-in-time calculation. Among them, we have analyzed the influence of the number of processors, the number of iterations in the parallel-in-time algorithm and the influence of the coarse to fine time-grid step size ratio. The good agreement between the obtained parallel efficiency and the values estimated by the proposed performance model allows to conclude that parallel-in-time efficiency is quantitatively different from the parallel efficiency of the space domain decomposition, regarding the number of processors available for a fixed problem dimension. A significant speed-up is possible when the temporal scale of the problem is large and enough processors are available.