Differentiated Object Placement and Location for Self-Organizing Storage Clusters

Component additions or failures are common for large-scale storage clusters in production environments. To improve manageability, it is desirable to automatically integrate new resources into an existing system as one virtual volume and repair data redundancy damaged by component failures. A key enabling technique for these capabilities is a flexible and scalable object placement and location scheme that quickly adapts to the additions or departures of storage nodes. This paper presents an object placement and location protocol that differentiates between small and large data objects for better space utilization and reduced management overhead comparing with uniform strategies. In our protocol, small objects, which are typically in large quantities, are placed through a high-dimension consistent hashing scheme. Large objects, which are much fewer in practice, are placed through a usage-based policy, and their locations are tracked with Bloom filters. Our protocol is fully distributed and only relies on soft states. It is particularly suitable for large-scale storage clusters with several thousand nodes and millions of objects. We validate the effectiveness of proposed techniques through simulations and experiments.