Abstract
High performance computing boasts a steadily increasing number of adherents, for excellent reasons, traditional and innovative. However, the promises for which scientists and engineers have been drawn to the extreme of the spectrum are more elusive than the simple extrapolation of decades of exponential improvement suggests. Programming today's systems for high performance is already problematic for many scientists and the learning curve will inexorably steepen. Expectations should be brought in line with the increasing difficulty of pushing the high performance frontier, and, for once, the modeling, algorithms, and software advances required to use the emerging hardware effectively should not be left until the hardware begins to arrive.