6.1.2.2 MPP systems

MOLPRO will in principle run on distributed-memory multiprocessor systems, including workstation clusters, under the control of the Global Arrays parallel toolkit. In practice, many parts of the code are not parallelized in this way, and therefore run with replicated work. Additionally, some of those parts which have been parallelized rely on fast inter-node communications, and can be very inefficient across ordinary networks. Therefore, at present, some caution is needed to avoid waste of resources in a multiuser environment.

A programming model has been adopted in which disk I/O is simulated within the global memory of all processors. This avoids potential bottlenecks arising from many processors simultaneously trying to do I/O, taking advantage of the large total memory in a typical MPP machine in order to store all two-electron integrals and other quantities in memory. By default, all unnamed MOLPRO files are implemented in this way. The amount of memory available for such activities is reserved in advance, and can be adjusted by setting the environment variable MOLPRO_GA_MEMORY, which specifies a number of 8-byte words to be used on each processor. This memory is distinct from, and not available to, the normal workspace specified in the memory input command.