Comparing reliabilities of centralized and distributed switching architectures for reconfigurable 2D arrays

Abstract

Whether used as main processing engines or as special-purpose adjuncts, processor arrays are capable of boosting performance for a variety of computation-intensive applications. For large processor arrays, needed to achieve the required performance level in the age of big data, processor malfunctions, resulting in loss of computational capabilities, form a primary concern. There is no shortage of alternative reconfiguration architectures and associated algorithms for building robust processor arrays. However, a commensurately extensive body of knowledge about the reliability modeling aspects of such arrays is lacking. We study differences between 2D arrays with centralized and distributed switching, pointing out the advantages of the latter in terms of reliability, regularity, modularity, and VLSI realizability. Notions of reliability inversion (modeling uncertainties that might lead us to choose a less-reliable system over one with higher reliability) and modelability (system property that makes the derivation of tight reliability bounds possible, thus making reliability inversion much less likely) follow as important byproducts of our study.