Network Aware Resource Management in Disaggregated Data Centers

Abstract

COMPUTING power has become a fundamental prerequisite to modern life and is available in abundance to companies, researchers, and end-users. For a given price, practically unlimited computing power is available to anyone, and cloud computing removes any geographic barriers from access to computing infrastructure. However, irrespective of how easy and flexible access to computing power has become, data centres (DCs), the physical location that hosts the computing infrastructure itself, are based on the same stale architectural principals since their creation.

A DC is an extensive set of interconnected servers. Each server consists of a fixed set of computing resources, i.e. central processing unit (CPU), random access memory (RAM), networking interfaces and storage, all of which are connected to the motherboard available only to applications that are executed on that server. Even though DCs have repeatedly proven their ability to serve advanced applications and millions of users, the concept of disaggregated resources will revolutionize the paradigm. Disaggregation provides flexibility in resource management, improving overall hardware utilization and reducing energy footprints.The disaggregation concept was recently introduced to the research community, may be a "tabula rassa" for the researchers and based on the specific requirements of it to operate, its an opportunity to reinvent some of the core features of the data centres.

In this thesis, we propose a modular disaggregated data centre network architecture and a management software specifically designed for disaggregated computing. This management software is responsible to allocate both computing and network resources. We conduct an energy study based on the proposed management software and discuss the benefits that
disaggregation will bring to operators. We also analyse the factors that may affect the overall energy consumption and the resource utilization. Finally, we showcased that resource allocation algorithms specifically designed for disaggregated data centres require 85% less power than algorithms that are currently implemented in modern data centres.

Date of Award	12 May 2020
Original language	English
Awarding Institution	University of Bristol
Supervisor	Reza Nejabati (Supervisor) & Dimitra Simeonidou (Supervisor)