Seminar presented as part of the University of Adelaide Computer Science Department Honours Program Series on Thursday 30th October, 1997.
Networks of workstations are a popular choice for use as parallel computing systems. They use 'off the shelf' products and are therefore cheaper than dedicated supercomputers. In addition, high speed general purpose networks and powerful workstation processors are closing the gap in terms of performance between networks of workstations and dedicated supercomputers. Currently, the major programming model for networks of workstations is the message passing model. However, programmers who are used to the traditional shared memory model find the message passing model difficult to adjust to. Cid attempts to provide a low entry cost into distributed parallel programming by extending C with MIMD threads, global objects and a 'join variable' mechanism which can be used to wait for the completion of asynchronous actions. These extensions are provided via a preprocessor which converts Cid code into C code and a runtime library which has been written in standard C code. Therefore Cid programs may be compiled using standard C compilers.
My talk will assess Cid by comparison with MPI, a definition of a library of message passing routines which may be linked to C code by a standard C compiler. Several applications will be presented which have been implemented in both Cid and MPI. These applications will be used to assess each model on the basis of both expressiveness of parallel concepts and performance.
PostScript version of the slides