Actually, highperformance servers and clusters of multiple processors are in wide use today. But it's difficult to program parallel systems.
Boston-based Interactive Supercomputing Inc. says its Star-P software overcomes this difficulty. "Users code algorithms and models on desktop computers using familiar mathematical software such as Matlab from MathWorks," says Interactive's Vice President Ilya Mirman. "Star-P parallelizes code on the fly, letting users scale applications across multiprocessor systems in real time."
So far, Star-P works on parallel servers such as SGI Altix and AMD Opteron-based clusters. Matlab was the initial application for Star-P, and interfaces to other desktop math programs are in the works. "Engineers will be able to load programs written in C or Fortran onto the server, call them through the Star-P client, and get supercomputer performance," he says.
Of course, parallel execution on multiprocessor computers is not new. "But writing software to run on parallel systems has been a challenge. Rewriting applications in languages such as C, Fortran, or MPI is time consuming a matter of years in some cases. What's more, algorithms and digital models are becoming more complex, dramatically outpacing the steady improvement in desktop computers' processing power," says Mirman.
Star-P reverses the trend, he adds, because it lets users work in familiar desktop software. "Users can write just enough of an application to start testing with real data. They can pause executions, change parameters, and restart them. These interactive workflows cut months off projects," he adds.