Mapping numerical software onto distributed memory parallel systems
The aim of this thesis is to further the use of parallel computers, in particular distributed memory systems, by proving strategies for parallelisation and developing the core component of tools to aid scalar software porting. The ported code must not only efficiently exploit available parallel processing speed and distributed memory, but also enable existing users of the scalar code to use the parallel version with identical inputs and allow maintenance to be performed by the scalar code author in conjunction with the parallel code. The data partition strategy has been used to parallelise an in-house solidification modelling code where all requirements for the parallel software were successfully met. To confirm the success of this parallelisation strategy, a much sterner test was used, parallelising the HARWELL-FLOW3D fluid flow package. The performance results of the parallel version clearly vindicate the conclusions of the first example. Speedup efficiencies of around 80 percent have been achieved on fifty processors for sizable models. In both these tests, the alterations to the code were fairly minor, maintaining the structure and style of the original scalar code which can easily be recognised by its original author. The alterations made to these codes indicated the potential for parallelising tools since the alterations were fairly minor and usually mechanical in nature. The current generation of parallelising compilers rely heavily on heuristic guidance in parallel code generation and other decisions that may be better made by a human. As a result, the code they produce will almost certainly be inferior to manually produced code. Also, in order not to sacrifice parallel code quality when using tools, the scalar code analysis to identify inherent parallelism in a application code, as used in parallelising compilers, has been extended to eliminate dependencies conservatively assumed, since these dependencies can greatly inhibit parallelisation. Extra information has been extracted both from control flow and from processing symbolic information. The tests devised to utilise this information enable the non-existence of a significant number of previously assumed dependencies to be proved. In some cases, the number of true dependencies has been more than halved. The dependence graph produced is of sufficient quality to greatly aid the parallelisation, with user interaction and interpretation, parallelism detection and code transformation validity being less inhibited by assumed dependencies. The use of tools rather than the black box approach removes the handicaps associated with using heuristic methods, if any relevant heuristic methods exist.