Scalable local reconstruction code design for hot data reads in cloud storage systems

Since demand for data is significantly heterogeneous in cloud storage systems (CSSs), there is traffic congestion in nodes storing hot data. In erasure-coded CSSs, traffic congestion can be alleviated by degraded reads sacrificing the bandwidth of surviving nodes. Local reconstruction codes (LRCs) reduce the bandwidth consumption of degraded reads, but cannot provide skewed throughput gain for the hot data. In this paper, we propose a scalable local reconstruction code (SLRC) that relies on LRCs but is more flexible in improving the throughput of a specific data block. First, we develop the local maximum throughput (LMT) to measure the maximum throughput of the hot data blocks by analyzing the actual read arrival rate of LRCs. Further, we elaborate on the structure of SLRC and analyze their performance metrics, which include storage overhead, reconstruction cost, and LMT. To select the appropriate code, we present the minimum reconstruction cost, minimum storage overhead, and minimum penalty algorithms. Finally, we implement extensive experiments on several typical SLRCs on the Hadoop distributed file system. Higher LMT and lower bandwidth consumption can be provided by SLRCs for hot data block degraded reads in CSSs compared with RS codes and LRCs.