MPI and OpenMP Paradigms on Cluster of SMP 在SMP集群MPI和OpenMP范式.pptVIP

10/10/2002 Yun (Helen) He, GHC2002 08/07/2002 NERSC, LBNL Effective Methods in Reducing Communication Overheads in Solving PDE Problems on Distributed-Memory Computer Architectures Chris Ding and Yun (Helen) He Lawrence Berkeley National Laboratory Outline Introduction Traditional Method Ghost Cell Expansion (GCE) Method GCE Algorithm Diagonal Communication Elimination (DCE) Technique Analysis of GCE Method Message Volume Communication Time Memory Usage and Computational Cost Performance Test Problem Test Results Conclusions Background On a distributed system, each processor holds a problem subdomain, which includes one or several boundary layers (usually called ghost cells). Ghost cells contain the most recent values of the neighboring processors. They must be updated at each time step. And it is done so in traditional approach. The message sizes exchanged between processors are usually very small. To speedup the communication, one idea is to combine messages for different time steps and exchange the bigger messages less frequently to reduce the communication latency. Traditional Field Update Ghost Cell Expansion (GCE) Method Algorithm for GCE Method Diagonal Communication Elimination (DCE) Technique A regular grid in 2D decomposition has 4 immediate neighbors (left and right), and 4 second nearest neighbors (diagonal). A regular grid in 3D decomposition has 6 immediate neighbors (horizontal and vertical), 12 second nearest neighbors (planar corner), and 8 third nearest neighbors (cubic corner). Diagonal Communication Elimination (DCE) Technique (cont’d) Message Volume Communication Time Theoretical Speedup Memory Usage and Computational Cost Test Problem 2D Laplacian Equation with Dirichlet boundary conditions: Performance Analysis Test 1: Global size 3200x3200, P=16, L=1 Test 2: Global size 3200x3200, P=16, L=2 Test 3: Global size 6400x6400, P=