Quality of Service Benefits of Flash in a Storage System

Open Access
Author:
Jain, Kanishk
Graduate Program:
Computer Science and Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
July 02, 2008
Committee Members:
  • Bhuvan Urgaonkar, Thesis Advisor
Keywords:
  • throughput
  • fairness
  • quality of service
  • hybrid store
  • flash
  • storage system
  • response time
  • random
  • sequential
Abstract:
Virtualized storage systems can be shared by applications (workloads) with substantially different resource requirements, workload patterns, and levels of importance. In such a system, fairness can be defined as a metric that estimates how well the application throughputs match their assigned level of importance (i.e. relative weights). Furthermore, clients paying for storage as a service may expect a certain minimum level of performance from the virtualized system. In general, for a storage system, there is a fundamental tradeoff between the goal of achieving fairness and the goal of maximizing overall performance. Thus in a virtualized storage system, an important quality of service (QoS) goal is to try to optimize fairness, while also ensuring that the overall performance stays above a minimum acceptable threshold. Storage performance is notoriously difficult to estimate in a storage system and varies significantly with workload characteristics. Due to the mechanical nature of a disk, a random workload achieves a lower throughput, and experiences a higher response time, than a sequential workload. As a result of this behavior, an algorithm that aims to achieve the aforesaid QoS goal, results in asymmetric fairness between a random and a sequential stream. Flash is a storage device that has the potential of making storage systems independent of the effects of randomness in workloads, and thus of making the performance of virtualized storage systems more predictable. The scope of this thesis is to investigate how flash can be used to augment virtualized storage systems, in order to improve fairness, while also achieving sufficient minimum performance.