Data Management Strategies for Relative Quality of Service in Virtualised Storage Systems

摘要

The amount of data managed by organisations continues to grow relentlessly.udDriven by the high costs of maintaining multiple local storage systems, thereudis a well established trend towards storage consolidation using multi-tier Virtualised Storage Systems (VSSs). At the same time, storage infrastructuresudare increasingly subject to stringent Quality of Service (QoS) demands.udWithin a VSS, it is challenging to match desired QoS with delivered QoS,udconsidering the latter can vary dramatically both across and within tiers.udManual efforts to achieve this match require extensive and ongoing humanudintervention. Automated efforts are based on workload analysis, which ignoresudthe business importance of infrequently accessed data.udThis thesis presents our design, implementation and evaluation of dataudmaintenance strategies in an enhanced version of the popular Linux Extendedud3 Filesystem which features support for the elegant specificationudof QoS metadata while maintaining compatibility with stock kernels. Usersudand applications specify QoS requirements using a chmod-like interface. Systemudadministrators are provided with a character device kernel interfaceudthat allows for profiling of the QoS delivered by the underlying storage. Weudpropose a novel score-based metric, together with associated visualisationudresources, to evaluate the degree of QoS matching achieved by any givenuddata layout. We also design and implement new inode and datablock allocationudand migration strategies which exploit this metric in seeking to matchudthe QoS attributes set by users and/or applications on files and directoriesudwith the QoS actually delivered by each of the filesystem’s block groups.udTo create realistic test filesystems we have included QoS metadata supportudin the Impressions benchmarking framework. The effectiveness of theudresulting data layout in terms of QoS matching is evaluated using a specialudkernel module that is capable of inspecting detailed filesystem data on-the-fly.udWe show that our implementations of the proposed inode and datablockudallocation strategies are capable of dramatically improving data placementudwith respect to QoS requirements when compared to the default allocators.