Data center disaggregation: when and how much?

Abstract

Disaggregation has been proposed by both the industry and academia as an approach to reduce resource fragmentation in cloud data centers. With disaggregation the entire rack is viewed as a single machine in which all resources are remotely accessible from anywhere in the rack. Because a large number of resources may be used remotely, the average access time to resources tends to be elongated compared to a non-disaggregated rack organization. This elongation represents a technology penalty of disaggregation and impacts application performance. This paper presents a model to capture basic tradeoffs between virtual machine (VM) lifetime elongation and improved memory mutualization in a rack with disaggregation. Our results show that disaggregation has the potential to significantly improve the maximum VM launch throughput provided the technology penalty remains moderate. Additionally, our model provides guidance for the number of servers to combine through disaggregation.