Search

Massively parallel (ML) computing is seen by engineers and scientists as a tool useful for reaching this goal. Unfortunately, for those who simply wish to use the tool, exploiting parallelism in physical problems brings with it a new set of challenges. Issues that were not important for vector supercomputer implementation can seriously impact a simulation's performance on a ML machine. These include decomposing the physical problem into naturally parallel parts, load balancing the computation across multiple processors, efficiently communicating data between processors, and fast movement of data in and out of the machine. It is these issues-which the engineer/scientist views as tangential to the task of actually solving the problem at hand, that have created the perception that parallel computers are a specialty breed and have slowed their acceptance by industry. As large parallel machines become more generally available to the traditional vector supercomputer user community, and programmers become accustomed to thinking about their problems from the perspective of parallelism, creating new parallel simulations will become commonplace and thus easier. A second component of the solution is auxiliary tools supplied by vendors and third parties.