To scale with increasing number of clients and I/O requests, the design of a storage subsystem should not contain bottleneck components or single points of failure. Traditionally, a large storage system is built using a distributed file server (e.g. AFS). In such servers, the collection of objects is partitioned across a set of independent nodes, with each node managing objects in its partition. This approach has two shortcomings. First, when the accesses across partitions are highly skewed, some nodes become hot-spots or bottlenecks. Second, each node becomes a single point of failure for the objects in its partition. To address these shortcomings, data is dynamically moved or replicated across multiple nodes. However, given its large space overhead, replication is not a scalable solution.

We have investigated the design of clustered servers that consist of a group of cooperating nodes connected by a scalable interconnect. We have designed data layout policies that balance load across nodes and that increase data availability. We instantiated our design in a research prototype.

Representative Publications:

1. R. Tewari, R. Mukherjee, D.M. Dias, and H.M. Vin, Design and Performance Tradeoffs in Clustered Video Servers, In Proceedings of the IEEE International Conference on Multimedia Computing and Systems 1996 (ICMCS'96), Tokyo, Japan, Pages 144-150, May 1996 [ Abstract | Paper ]
   
   
   

2. R. Tewari, D.M. Dias, R. Mukherjee, and H.M. Vin, High Availability in Clustered Multimedia Servers, In Proceedings of the IEEE International Conference on Data Engineering, New Orleans, Pages 345-354, February 1996 [ Abstract | Paper ]