Leveraging Locality to Boost the Update Performance of Network-Coding-Based Storage Systems

Network coding is becoming a very important technology to improve the reconstructing performance of distributed storage systems due to its high reliability and low redundancy. Update is a typical and important operation involved in the network-coding-based systems. However, the overhead incurred by the update operation has a significant impact on the system performance. Traditionally, the small updated data are aggregated as different big logs and then flushed to the back-end storage nodes, thus reducing the I/O overhead. We investigate a large volume of traces and find that the updated data normally contain very strong locality. Based on this observation, this paper proposes a Locality-based Update Scheme (LUS) which maintains the update related data in the cache longer to leverage the locality. The subsequent update related data will be cached, thus significantly decreasing the data accesses going to back-end storage nodes and the bandwidth required to transfer the data. Experimental results demonstrate that in contrast to a state-ofthe-art approach, LUS reduces the bandwidth consumption and update time up to 30% and 24.57%, respectively.