Optimizations for Energy-Aware, High-Performance and Reliable Distributed Storage Systems

With the decreasing cost and wide-spread use of commodity hard drives, it has become possible to create very large-scale storage systems with less expense. However, as we approach exabyte-scale storage systems, maintaining important features such as energy-efficiency, performance, reliability and usability became increasingly difficult. Despite the decreasing cost of storage systems, the energy consumption of these systems still needs to be addressed in order to retain cost-effectiveness. Any improvements in a storage system can be outweighed by high energy costs. On the other hand, large-scale storage systems can benefit more from the object storage features for improved performance and usability. One area of concern is metadata performance bottleneck of applications reading large directories or creating a large number of files. Similarly, computation on big data where data needs to be transferred between compute and storage clusters adversely affects I/O performance. As the storage systems become more complex and larger, transferring data between remote compute and storage tiers becomes impractical. Furthermore, storage systems implement reliability typically at the file system or client level. This approach might not always be practical in terms of performance. Lastly, object storage features are usually tailored to specific use cases that makes it harder to use them in various contexts. In this thesis, we are presenting several approaches to enhance energy-efficiency, performance, reliability and usability of large-scale storage systems. To begin with, we improve the energy-efficiency of storage systems by moving I/O load to a subset of the storage nodes with energy-aware node allocation methods and turn off the unused nodes, while preserving load balance on demand. To address the metadata performance issue associated with large creates and directory reads, we represent directories with object storage collections and implement lazy creation of objects. Similarly, in-situ computation on large-scale data is enabled by using object storage features to integrate a computational framework with the existing object storage layer to eliminate the need to transfer data between compute and storage silos for better performance. We then present parity-based redundancy using object storage features to achieve reliability with less performance impact. Finally, unified storage brings together the object storage features to meet the needs of distinct use cases; such as cloud storage, big data or high-performance computing to alleviate the unnecessary fragmentation of storage resources. We evaluate each proposed approach thoroughly and validate their effectiveness in terms of improving energy-efficiency, performance, reliability and usability of a large-scale storage system