A Study on Achieving Performance Isolation and Improvement in an SSD-based Shared Storage System

Abstract

Storing data in shared storage systems saves both hardware cost and management cost Conventional shared storage systems use magnetic hard drives (HDDs) as the main storage media In recent years NAND ?ash-memory-based solid-state drives (SSDs) have been increasingly used in shared storage systems Typically storage users specify the target service-level objectives (SLO) for their virtual storage and the system administrators need to provide performance isolation among the virtual storage to avoid SLO violations To achieve performance isolation accurate I/O cost accounting is crucial However the interferences among the competing virtual storage devices could lead to inaccurate I/O cost accounting and damage the performance isolation Since an SSD has very different features compared to a HDD the interference sources could be also different We have found that the main interference sources in the SSD-based shared system are queuing delay interference and the garbage collection interference In this thesis we propose a framework called VSSD to eliminate the interferences and then to achieve performance isolation A novel I/O scheduler called FACO and a flash translation layer called ViSA are adopted to address the two interferences respectively in VSSD On the other hand since the performance of a virtual storage device is usually highly related to the size of the provisioned space and it is difficult for users to estimate the actual required space the space of the virtual storage devices may not be well-utilized To address this we design a run-time storage lending mechanism called RALIS to allow a virtual storage (i e borrower) to temporarily use the space belonging to other virtual storage that do not need the space in that moment (i e lenders) to boost the performance RALIS dynamically determines the role (i e lender or borrower) of the virtual storage To minimize the possible impact to the lenders a block returning policy is used to make the borrowers to return the borrowed blocks before the lenders needs them Multiple sets of workloads were used to simulate the shared storage environment The results show that the VSSD can eliminate the interferences and achieve performance isolation In addition RALIS can boost the performance of the borrowers effectively The storage lifetime can also be extended significantly (35 9% to 85 6%) by RALIS

Date of Award	2016 Feb 4
Original language	English
Supervisor	Da-Wei Chang (Supervisor)