Evolution of massive parallel compute servers from a research object to a production tool

Abstract

Massive parallel systems built from microprocessors have caught up with conventional supercomputers in performance and are expected to far exceed the conventional supercomputers in the current decade. Programmer productivity and deliverable, scalable performance are important requirements that must be met before massively parallel systems can achieve broader acceptance for scientific computing. However this will not be enough to compete on that market and generate profits as a production tool: potential customers in advanced sciences keep lots of proprietary know-how in form of existing software with no chance of rewriting, unless to loose their competitive edge by missing the time-window to market their products. Except for projects starting from scratch, the solution to MPP may depend on the ability to deliver enough sustained performance to somewhat modified existing application software thus keeping the customers competitive in their own market.