Date of Award

Spring 2015

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical and Computer Engineering

First Advisor

Mithuna S. Thottethodi

Committee Chair

Mithuna S. Thottethodi

Committee Member 1

Samuel P. Midkiff

Committee Member 2

T. N. Vijaykumar

Committee Member 3

Vijay S. Pai


Data-set sizes are growing. New techniques are emerging to organize and analyze these data-sets. There is a key access pattern emerging with these new techniques, large sequential file accesses. The trend toward bigger files exists to help amortize the cost of data accesses from the storage layer, as many workloads are recognized to be I/O bound. The storage layer is widely recognized as the slowest layer in the system. This work focuses on the tradeoff one can make with that storage capacity to improve system performance. ^ Capacity can be leveraged for improved availability or improved performance. This tradeoff is key in the storage layer, as this allows for data loss prevention and bandwidth aggregation. Typically these tradeoffs do not allow much choice with regard to capacity use. This work will leverage replication as the enabling mechanism to improve the capacity-performance tradeoff in the storage tier, while still providing for availability. ^ This capacity-performance tradeoff can be made at both the local and distributed file system level. I propose two techniques that allow for an improved tradeoff of capacity. The local file system can be employed on scale-out or scale-up infrastructures to improve performance. The distributed file system is targeted at distributed frameworks, such as MapReduce, to improve the cluster performance. The local file system design is MorphStore, and the distributed file system is BoostDFS. ^ MorphStore is a file system that significantly improves performance when accessing large files by using two innovations. MorphStore combines (a) load-adaptive I/O access scheduling to dynamically optimize throughput (aggregation), and (b) utility-xiii driven replication to best use capacity for performance. Additionally, adaptive-access scheduling can be utilized to optimize scheduling of requests (for throughput) on systems with a large number of storage devices. Replication is utilized to make available high utility files and then optimize throughput of these high utility files based on system load. ^ BoostDFS is a distributed file system that allows a better capacity-performance tradeoff via inter-node file replication. BoostDFS is built on the observation that distributed file systems currently inter-node replication for availability, but provide no mechanism to further improve performance. Replication for availability provides diminishing returns on performance, this is due to saturation of locality. BoostDFS exploits the common by improving I/O performance of these local tasks. This is done via intra-node replication by leveraging MorphStore as the local file system. This technique allows for capacity to be traded for availability as well as performance, with a small capacity overhead under constant availability. ^ Both MorphStore and BoostDFS utilize replication. Replication allows for both bandwidth aggregation and availability, This work primarily focuses on the performance utility of replication, but does not sacrifice availability in the process. These techniques provide an improved capacity-performance tradeoff while allowing the desired level of availability.