An Easy, Quantitative Method to Evaluate Replica Placement Policies in Distributed Storage Systems

摘要

In large scale distributed systems, data blocks are usually replicated and distributed across the storage nodes to achieve high availability and reliability. The placement of blocks and their replicas has a marked impact on the data reliability and system performance. Various replica placement policies have been proposed in the past decades, and some of them have been implemented in enterprise storage systems. However, it is difficult to verify that a selected placement scheme yield near-optimal performance or reliability in the real world due to several reasons. First, it is not feasible to build up a physical testbed with hundreds of storage nodes; and constructing a simulation model is also costly and hard to catch up with the fast-paced changes in industry. Second, previous studies usually carry out the analysis to address only one particular facet of schemes. Third, most algorithms are too complicated for programmers to adopt. Hence a synthetic and cost efficient method is needed. Here we use an analytic model to assess reliability and performance of replica placement policies by using a combination of quantitative metrics. We claim that this method is useful for both the analysis in the stage of architecture design and the parameter selection of existing systems. Four replica placement policies with deterministic declustering are under consideration.