Bounding Communication Energy Overhead in Parallel Networks with Power-Delay Scalability

We propose the energy scalability and power-delay product scalability as new metrics for calculating and estimating boundaries on the energy overhead of parallel processing systems. Because both properties consider network topology and the algorithm to be implemented, they are shown to be effective in predicting the energy scalability of a multiprocessor in the most general case. Power-delay scalability is then shown to provide upper bound and lower bound estimates when analyzing the energy scalability of a multiprocessor while significantly reducing the computational cost of directly calculating the energy scalability. Energy scalability is an effective metric for predicting the energy overhead cost when using a parallel processor versus a single processor. Power-delay scalability is a means for estimating energy scalability on a system when exact calculation may be difficult due to variation in power levels which have become common with dynamic voltage frequency scaling.