Gavin J. Pringle and David S. Henty
EPCC, JCMB, Kings Buildings, The University of Edinburgh, Mayfield Road,
Edinburgh EH9 3JZ, Scotland.
The Portland Group's High Performance Fortran compiler, PGHPF, is a popular HPF compiler and is available for many platforms including the Cray T3D and T3E. Although programming in HPF is relatively straightforward, users often observe poor performance both in terms of single processor speed and scaling behaviour.
In this talk we will consider a real HPF application, a code currently running in full production on the 512-processor Cray T3D at EPCC, which simulates 3D non-linear, resistive Magnetohydrodynamics. This work was done in support of the MHD Consortium (led by Dr. Alan Hood) and was funded by the UK's High Performance Computing Initiative.
The code contains two routines which have been considered to parallelise poorly due to the method of parallelisation by PGHPF. These routines involve an "all-to-all" broadcast and referencing values of neighbouring points on a regular grid. Various styles of HPF programming are considered which reveal aspects of the PGHPF compiler's "character".
Furthermore, on the T3D, there was support for an alternative data parallel model via Cray's Craft compiler. As well as being non-standard, Craft was much more restrictive than HPF. However, Craft codes could often achieve performance superior to their HPF equivalents due to better scaling. On the T3E, Craft is supported as part of PGHPF. We report on the relative merits of HPF/Craft on the T3E and compare the results with those previously obtained on the T3D.