Abstract: Data reliability are drawn much concern in large-scale storage systems built from thousands of storage devices, which highly  depends on many inter-dependent system parameters, such as the replica placement policies, number of stored objects and so on. Previous work has discussed the impacts of these system parameters on reliability roughly and separately, and seldom provided their optimal values, nor mentioned their optimal combination. This paper presents a new object-based-repairing analytic model. Based on analyzing this model in three popular replica placement policies, it figures out the individual optimal values of these parameters at the beginning, and then works out their optimal combination. Compared with the existing models, this model is easier to solve while reaching more integrative and practical conclusions. These conclusions can directly and effectively instruct the designers to build more reliable storage systems.

Key words: data reliability, reliability model, large scale data center

摘要： 在大规模存储系统中，数据的可靠性越来越受到人们的关注。已有的研究分析了在系统规模已知的条件下，某些系统参数，如副本分布策略、存储对象数目等，对可靠性的粗略影响，但较少提及它们的最优值或者最优组合。提出了一种基于对象粒度恢复的可靠性新模型；基于该模型，在分析三种主流的副本分布策略的基础上，分别计算出了各个系统参数的独立最优值及其组合最优值。与已有模型相比，该模型更易于求解，且获得了更加综合实用的最优值，这些最优参数值能直接有效地指导系统设计者构建更可靠的大规模存储系统。

关键词: 数据可靠性, 可靠性模型, 大规模数据中心