Why there won't be apps: The problem with MPPs

Abstract

Gordon Bell gbellQmojave In spite of many years of research, massively funded, massively parallel (AKA “scalable”) computers aren’t yet successful. Nor are they likely to be unless they undergo a massive transformation to leverage developments in the mainstream computer and communications industries. T h e latest threat comes from standard workstations and fast, low-latency networks based on ATM. Like MPPs, these networks offer size scalability (from fewer to more processors), but they also offer geizeration scalability (from previous to future generations) and space scalability (from multiple nodes in a box, to computers in multiple rooms, buildings, or geographic regions). Furthermore, these networks offer a critical capability that MPPs sorely lack: application compatibility with workstations and multiprocessor servers. T h e meager existence to date of special-purpose MPPs stems from four factors: stanford.edu