Code Rate Adaptation in Cloud Storage Systems with Multiple Erasure Codes

Rui Zhu,Di Niu,Zongpeng Li

2016-01-01

Abstract:Erasure codes have been adopted for cloud storage systems. While achieving higher reliability at lower storage overhead as compared to replication, erasure codes usually incur high reading cost when recovering an unavailable block. Although local reconstruction code constructions have been proposed to reduce recovery cost, additional benefits can be achieved by adopting erasure codes with different code rates for data blocks with different popularity. In this paper, we study the problem of code rate selection and adaptation in cloud storage systems that adopt multiple erasure codes via online learning. Unlike offline optimization, which requires the knowledge or estimation of future demands, the online learning algorithms can make decisions only based on past observations and dynamically adapt to demand changes. To avoid solving a hard integer program, we perform a stochastic relaxation to the formulated online learning problem and solve it using a exponentiated gradient algorithm, resulting in sparse solutions. We show a regret bound of O( √ T ) of the proposed algorithm by showing that our algorithm is a special case of the FTRL online learning framework. Through trace-driven simulations based on real request traces from Windows Azure Storage, we show that our online algorithm performs close to the best fixed offline policy, and trades off between recovery cost during degraded reads and storage overhead.