SDN Enabled Network Efficient Data Regeneration for Distributed Storage Systems

Abstract

Distributed Storage Systems (DSSs) have seen increasing levels of deployment in data centers and in cloud storage networks. DSS provides efficient and cost-effective ways to store large amount of data. To ensure reliability and resilience to failures, DSS employ mirroring and coding schemes at the block and file level. While mirroring techniques provide an efficient way to recover lost data, they do not utilize disk space efficiently, resulting in large overheads in terms of data storage. Coding techniques on the other hand provide a better way to recover data as they reduce the amount of storage space required for data recovery purposes. However, the current recovery process for coded data is not efficient due to the need to transfer large amounts of data to regenerate the data lost as a result of a failure. This results in significant delays and excessive network traffic resulting in a major performance bottleneck.

In this thesis, we propose a new architecture for efficient data regeneration in distribution storage systems. A key idea of our architecture is to enable network switches to perform network coding operations, i.e., combine packets they receive over incoming links and forward the resulting packet towards the destination and do this in a principled manner. Another key element of our framework is a transport-layer reverse multicast protocol that takes advantage of network coding to minimize the rebuild time required to transmit the data by allowing more efficient utilization of network bandwidth.

The new architecture is supported using the principles of Software Defined Networking (SDN) and making extensions where required in a principled manner. To enable the switches to perform network coding operations, we propose an extension of packet processing pipeline in the dataplane of a software switch. Our testbed experiments show that the proposed architecture results in modest performance gains.