Year: 2016
Type: dissertation
Abstract: Storage systems rely on maintenance tasks, such as backup and defragmentation, to provide a multitude of guarantees that range from data availability and security, to good performance. These tasks access large amounts of data and can significantly impact foreground applications. The work presented as part of this thesis outlines the challenges of storage maintenance and provides mechanisms that lower its impact, allowing it to be performed more efficiently. Maintenance operations act as insurance for storage systems. To avoid the prohibitive cost incurred in their absence, storage operators pay the price of performing them periodically. In the case of reliability, for example, we find that the error characteristics of hard disk drives in Google's data centers are diverse across models and operating environments. Due to this unpredictability, the use of periodic maintenance becomes unavoidable. Furthermore, recent research efforts on disaster prediction depend on metrics collected through periodic maintenance operations. Unfortunately, periodic storage maintenance requires data to be accessed multiple times. As a case in point, we find that customers of Symantec and Veritas schedule backups multiple times a week. This often results in excess maintenance work that fails to complete within the scheduled downtime periods. At this point, foreground workload impact becomes unavoidable in order to maintain the same reliability guarantees. To lessen this impact, maintenance I/O can be scheduled only when storage devices are otherwise idle. To achieve this, idleness must be predicted accurately. Fortunately, our extensive study of disk workloads reveals characteristics that allow idleness to be predicted accurately by simple models. Given the number of maintenance tasks available today, idle time can prove insufficient. For such scenarios we present Duet, a framework that provides notifications to tasks about relevant changes in the system cache. Using this mechanism, tasks can reduce their I/O by prioritizing processing of cached data. Data can be cached either due to other maintenance tasks requesting it previously, or due to overlapping foreground I/O activity. In other words, tasks running concurrently can implicitly collaborate, collectively reducing the I/O required for maintenance and, by extension, the idle time needed to execute it.
Source: TSpace (University of Toronto)
Author: Georgios Amvrosiadis
Institution: University of Toronto
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Keywords: data backup, defragmentation, disaster prediction
Open Access status: green
OpenAlex ID: W2620346426
