A comparison of the performance of a fluid code on the C90 and the iPSC/860

The performance of nonlinear runs of a fluid code on the NERSC C90 and the ORNL Intel iPSC/860 is compared. This version of the KITE code is an initial value, three-dimensional, nonlinear, five-field code used for plasma turbulence studies. Because of the Fourier harmonic expansions in both sines and cosines, it has ten equations. The runs have from 385 to 513 radial grid points and 455 to 539 Fourier harmonics. The main components of the code are convolutions of Fourier harmonics and block tridiagonal matrix solves. Both parallel codes are optimized for the computer they use. With the collaboration of NERSC, microtasking calls have been added to the KITE serial code on the C90. The loop that contains calls to the block tridiagonal matrix solver is performed in parallel over the number of Fourier harmonics. The convolution routines also contain microtasking commands to compute the do loops in parallel over the number of Fourier harmonics. The authors are using an average of six concurrent CPUs for runs of thousands of timesteps in batch and many of the subroutines called every timestep are not microtasked yet. The KITE parallel code is running on 128 processors of the iPSC/860. The massively parallel version of this code allocates a radial region for each node. In this way, convolutions at a fixed radius are performed in parallel, and communication is limited to boundary values for each radial region. A second parallelization takes place for the block-tridiagonal solves. The Fourier components are distributed in nodes. In each node, the block-tridiagonal matrix is inverted for each of the allocated Fourier components. The authors compare CPU times, constraints imposed by memory size of the machines, difficulties in storing resulting data, and wallclock time needed to complete nonlinear runs