Date on Master's Thesis/Doctoral Dissertation

8-2024

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Altiparmak, Nihat

Committee Co-Chair (if applicable)

Gentili, Monica

Committee Member

Gentili, Monica

Committee Member

Lauf, Adrian

Committee Member

Nasraoui, Olfa

Committee Member

Zhang, Harry

Author's Keywords

energy efficiency; storage IO

Abstract

Our modern society has become ever more connected and reliant on ever larger quantities of data, which must be collected and processed over large geographic areas. The data centers that store, compute, and share this data have grown to such a scale that many data centers use as much electricity as a city. Priorities in research have shifted from not only seeking greater performance, but also towards greater energy efficiency. The control and management of computer systems has an impact on power consumption beyond the needs of individual components. In this dissertation, we examine techniques for improved performance and energy efficiency of data centers. First we investigate high performance parallel storage, where storage and computation is spread across multiple nodes in a network, and present techniques for virtual machine placement and replica selection with energy-aware improved performance. Second, we analyze operating system design concepts in the storage stack for individual systems, specifically request IO submission, IO scheduling, and IO completion mechanisms. We find the surprising result that polling based completion can provide greater energy efficiency than interrupt completion, despite its greater CPU utilization. We show how IO scheduling has become less beneficial to performance, especially with low latency devices. Lastly, we consider conditions that favor various IO interfaces to aid application developers concerned with both performance and energy efficiency.

Download

Share

COinS