Methodologies for Performance Assessment for Parallel Systems

1The recent experience has shown that parallel systems are no longer experimental and useful systems can be built (e.g. the Connection Machine) to solve real problems (e.g. linear algebra, partial differential equations). It is also no longer the question of whether a Tera FLOPS system can be built, rather it is how can such powerful systems be used effectively. In many cases only a small fraction of system's potential power is utilized. The efficient use of these systems is a function of a number of factors, but primarily depends on exploting the application's inherent parallelism, choosing a compatible architecture, and optimally mapping the application functions into the processors. Our innovation is to develop mehtodologies (and eventually a set of software tools and an environment) which a user can employ to perform above and other tasks in support of designing an application and evaluating its performance on a parallel system. Phase I of this research will be devoted to assessing the feasibility of synthesizing the application and the architecture from a set of kernels. Phase II and III will be devoted to the actual development of methodologies and the software tools.