Inherent Parallelism

To run a software application on a large number of parallel processors, N, and expect to obtain speed increases near N, each processor must be used effectively. Furthermore, given that the multiplier on speed, M, is sufficiently close to N, then one may want to add more processors while maintaining the value of M, implying a linear increase in speed with number of processors. Anyone who has conducted scientific experiments along these lines will attest to the difficulty of achieving success as N gets large. Except in special “embarrassingly parallel ” cases, the amount of time and resource invested in hardware and software to achieve these goals has been disproportionate to the (lack of) success. It is not unusual to find M less than 1 as N gets large. In most unsuccessful cases, failure appears to be due to a lack of understanding of the problem. Given that the problem is understood, there has been a clear lack of software tools and techniques available to solve it, until now, see [1] and [2]