Skip to Main Content
In cluster-based storage systems, the metadata server cluster must be able to adaptively distribute responsibility for metadata to maintain high system performance and long-term load balance, due to workload skew and metadata servers' heterogeneity. In this paper, we describe a simple and adaptive metadata load management scheme, called self-balancing uniform (SBU) randomization, to efficiently and continually adapt the metadata distribution to current demands in heterogeneous metadata server cluster. We implement our system within a discrete event driven simulation environment, along with two other systems, simple randomization (SR) and performance aware distribution (PAD) to serve as points of comparison, and evaluate the performance of our SBU algorithms against SR and PAD algorithms by both a trace workload and a synthetic workload. Simulation results verify that our SBU algorithm achieves load self-balance, provides consistent response latencies and resource utilization. Simulation results also indicate that SR cannot cope with skew and heterogeneity and PAD requires a larger shared state to achieve optimal performance.