A reliability assessment of series-stacked servers with server-to-bus differential power processing

Abstract

As demand for data centers grows with the expansion of cloud computing and storage, so too does the power consumption. In the search for more efficient and reliable power delivery to the servers that make up the data centers, researchers and industry have been investigating various power delivery options. Recently, series-stacked architectures have been proposed to address the efficiency limitation of conventional power delivery architectures in data centers by processing only the difference in power between a group of servers. This paper addresses the reliability aspect of the series-stacked architectures, and compares it to the conventional power delivery architecture in which the full server power needs to be processed. Through the calculation of a cluster level down-time in both power delivery architectures, the series-stacked architecture is shown to be a competitive alternative to the conventional architecture because of reduced converter failure rate.