The Communication Performance of the Cray T3D and its Effect on Iterative Solvers

On many distributed memory systems, such as workstation clusters or the Intel iPSC/860, the multigrid algorithm suffers from having extensive communication requirements and, in general, it is not very competitive in comparison to the conjugate gradient algorithm. This is in contrast to the sequential problem whereby the multigrid algorithm is very effective in reducing the global residual, particularly for very large linear systems of equations. These two algorithms are now compared on the Cray T3D for solving very large systems of linear equations (resulted from grids of the order 2563 cells). The communication performance of the Cray T3D is first measured by the standard ping-pong tests and also by practical communication tasks that are found frequently in CPD calculations. It is found that the Cray T3D has a low latency (= 6 ps) and a high bandwidth interprocessor communication (120 MB/s) when the low-level intrinsic communication routines are used. As a result, the multigrid algorithm is found to be very competitive when compared with the conjugate gradient algorithm for solving the very large linear systems arising from the Direct Numerical Simulation of turbulent Combustion (DNSC). Results are contrasted by those on the Intel iPSC/%O.