Reducing Delay of Flexible Download in Coded Distributed Storage System
Qiqi Shuai, Victor O. K. Li · 2016
Download delay is a crucial performance metric in distributed storage systems as it greatly impacts user experience. Recently, plenty of research has pointed out that coding can reduce delay compared with replication. However, almost all previous studies only focus on the case in which all of the users require the same size of files and hence must download all files of a codeword and ignore the case in which some users only need some of the files in the codeword. Moreover, they do not consider the advantage of codes with original data nodes, such as systematic codes. In this paper, based on a more general and practical case in which download requests may desire different sizes of files and hence only some of the files of a codeword in a systematic (n, k) MDS-coded storage system, we propose the compound read method, characterize its mean download delay in low arrival rate scenario and derive upper and lower bounds on its mean download delay in high arrival rate scenario. We also compare the delay performance of compound and k-access reads and propose a scheme C & K to dynamically take advantage of them according to users' required size of files to reduce the mean download delay. In addition, with real service time traces from Amazon S3, we conduct trace-driven simulations to verify our theoretical analyses and the effectiveness of the C & K scheme.