Date: October 29th, 2010
PhD Thesis Proposal
ENHANCING STORAGE PERFORMANCE IN VIRTUALIZED ENVIRONMENTS: A PRO-ACTIVE APPROACH
Abstract :
Virtualization of computer resources is widely recognized
as an important revolution in computing industry over the last decade.
Virtualization technology provides some unique opportunity for better resource
utilization and more effective server and application consolidation. While CPU,
memory and network virtualization have been studied in depth in the research
community, the impact of virtualization on storage systems has not been analyzed
in depth and understood completely. We argue that certain I/O functions that
were traditionally the responsibility of file systems become sub-optimal in the
virtual machine context because a physical resource is typically shared across
multiple operating domains. In this dissertation, we present a pro-active
approach to storage performance enhancement in virtualized cloud computing
environments.
This dissertation research makes three unique contributions. First, we analyze
the effect of system virtualization on storage performance through a variety of
experiments. We identify the set of storage system parameters that are critical
to storage performance in virtualized cloud environments. We examine how each
individual virtual machine contends for the physical storage by running
different kind of workload combinations under different resource allocations in
the individual virtual machines. We make interesting observations from these
experiments that shed more light on the problems to be addressed in order to
achieve good storage performance in a virtualized computing environment. The
information divide in the virtual machine world is an instance of a more general
issue of functionality placement between the lower I/O layers and the file
system. To enable better co-operation in functionality between the layers of the
I/O stack, the second contribution we make in this dissertation research is to
introduce the notion of Pro-active Disks, which represents a novel mode of
interaction between storage systems and higher layers of the storage stack such
as file systems. Instead of viewing the storage system as a passive entity in
the storage stack that simply responds to higher level requests, we propose a
model where the disk can pro-actively and opportunistically initiate requests on
the higher layers. We show that the proactive capability of initiating requests
on the higher storage layer offers a unique advantage: It enables exploiting
internal information specific to the storage system to optimize higher level
policies without the higher layer needing to know the details of the
storage-level information. Finally, we develop an evaluation framework
consisting of a block-level I/O trace replay tool. A salient feature of this
development is being 'context-aware' in the sense that the trace replay can
adjust itself according to the system in which it is replayed, instead of
plainly following the original trace characteristics.
In this dissertation proposal exam, I will give an overview of our proactive
approach to storage performance optimization, focusing on addressing the
following two questions: (I) how does the system virtualization impact on the
performance of storage systems, and (ii) what opportunities do the proactive
disk approach offer in terms of storage management and storage performance
optimization. We end the talk by providing a summary of our ongoing research
that extends the architecture of Pro-active Disks in virtualized environments
and leverage additional opportunities that are specific to virtualized systems.
We also plan to modify our context aware I/O trace replay framework to account
for extra layering delays introduced by virtualization, in order to use it to
evaluate the concept of Pro-active Disks in virtualized environments.
Committee Members:
Dr. Ling Liu
(Professor, School of CS, Georgia Tech) [Advisor]
Dr. Calton Pu (Professor, School
of CS, Georgia Tech)
Dr. Umakishore Ramachandran
(Professor, School of CS, Georgia Tech)
Dr. Shamkant Navathe (Professor,
School of CS, Georgia Tech)
Talk Slides:
Report: