Efficient monitoring of large scale infrastructure as a service clouds
Abstract
Cloud computing has had a transformative effect upon distributed systems
research. It has been one of the precursors of supposed big data revolution
and has amplified the scale of software, networks, data and deployments.
Monitoring tools have not, however, kept pace with these developments.
Scale is central to cloud computing but it is not its chiefly defining property.
Elasticity, the ability of a cloud deployment to rapidly and regularly change
in scale and composition, is what differentiates cloud computing from
alternative paradigms of computation. Older tools originating from cluster,
grid and enterprise computing predominantly lack designs which allow
them to tolerate huge scale and rapid elasticity. This has led to the
development of monitoring as a service tools; third party tools which
abstract the intricacies of the monitoring process from the end user. These
tools rely upon an economy of scale in order to deploy large numbers of
VMs or servers which monitor multiple users’ infrastructure. These tools
have restricted functionality and trust critical operations to third parties,
which often lack reliable SLAs and which often charge significant costs. We
therefore contend that an alternative is necessary.
This thesis investigates the domain of cloud monitoring and proposes
Varanus, a new cloud monitoring tool, which eschews conventional
architectures in order to outperform current tools in a cloud setting. We
compare a number of aspects of performance including monitoring latency,
resource usage and elasticity tolerance. Through investigation of current
monitoring approaches in conjunction with a thorough examination of
cloud computing we derive a design for a new tool which leverages peer
to peer and autonomic computing in order to build a tool well suited to
the requirements of cloud computing. Through a detailed evaluation we
demonstrate how this tool withstands the effects of scale and elasticity
which impair current tools and how it employs a novel architecture which
reduces fiscal costs. We demonstrate that Varanus maintains a low, near 1
second monitoring latency, regardless of both scale and elasticity and does
so without imparting significant computational costs. We conclude that this
design embodied by this tool represents a successful alternative to current
conventional and monitoring as a service tools.
Type
Thesis, PhD Doctor of Philosophy
Collections
Items in the St Andrews Research Repository are protected by copyright, with all rights reserved, unless otherwise indicated.