IMPLEMENTATION OF A METHOD FOR SOLVING THE 2D HEAT CONDUCTION EQUATION ON THE HYBRID ARCHITECTURE (CPU + GPU)
V. O. Anisov, E. V. Vaziev, D. A. Ushakov VANT. Ser.: Mat. Mod. Fiz. Proc 2021. Вып.1. С. 39-52.
The paper presents a way of organizing the computation process on a hybrid-architecture computer: a general-purpose processor (CPU) and a graphics accelerator (GPU). A numerical method for solving the 2D heat conduction equation was implemented on the architecture with two-level parallelism. An unstructured grid of arbitrarily shaped triangular and quadrangular cells was used. The heat conduction equation solution is reduced to solving a system of linear equations using the conjugate gradient method. The paper analyses the iterative part of the method. Paralleling on a distributed memory is performed by decomposing the original grid into subdomains, with each subdomain being calculated using its own MPI process either on CPU (MPI+OpenMP), or on GPU (MPI+CUDA). The authors mean that the hybrid compute process is the concurrent use of the both options. A simple formula has been obtained to find decomposition parameters, which allows achieving a maximum hybrid computing performance. This formula is based on the assessment of the memory subsystem capabilities for the devices in use, because the memory bandwidth is the major speed-limiting factor for the method presented. In simulation experiments on one node, the hybrid compute mode efficiency was 75-89%. Keywords: hybrid computing, heat conduction, conjugate gradient method.
|