Data-driven Resource Allocation in Virtualized Environments
Abstract
Modern advances in virtualization technologies have revolutionized the way we build and manage computer systems. Virtualization technologies, however, adversely impact the predictability of system performance. This introduces several challenges in balancing performance and resource utilization. In this dissertation, we explore and address performance challenges introduced by virtualization in two application scenarios: network functions virtualization and distributed network emulation. First, we investigate the performance of virtualized network functions (VNFs) and propose a framework, NFV-VITAL, that characterizes performance impacts of hardware and software options and determines the optimal configuration for initial deployment of a VNF. Then we propose a system, Elastic resource flexing for Network functions VIrtualization (ENVI), to make accurate online scaling decisions based on evolving neural network classifiers. ENVI trains initial neural networks using experimental data sets collected during an offline stage and continues to update them using a window-based rewinding mechanism during online operation to capture emerging workload patterns. Second, we study the experiment fidelity problem in a distributed network emulation cluster comprising heterogeneous physical machines. We quantify the traffic processing capability of the physical machines and design an algorithm, Waterfall, that uses this information, together with the experimental topology, to determine an efficient mapping of the network experiment that preserves experiment fidelity.
Degree
Ph.D.
Advisors
Fahmy, Purdue University.
Subject Area
Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.