9. Conclusions

This paper presents a new persistent data management layer that enhances the ability of clusters to support Internet services. This self-managing layer, called a distributed data structure (DDS), fills in an important gap in current cluster platforms by providing a data storage platform specifically tuned for services' workloads and for the cluster environment.

This paper focused on the design and implementation of a distributed hash table DDS, empirically demonstrating that it has many properties necessary for Internet services (incremental scaling of throughput and data capacity, fault tolerance and high availability, high concurrency, and consistency and durability of data). These properties were achieved by carefully designing the partitioning, replication, and recovery techniques in the hash table implementation to exploit features of cluster environments (such as a low-latency network with a lack of network partitions). By doing so, we have ``right-sized'' the DDS to the problem of persistent data management for Internet services.

The hash table DDS simplifies Internet service construction by decoupling service-specific logic from the complexities of persistent state management, and by allowing services to inherit the necessary service properties from the DDS rather than having to implement the properties themselves.

Acknowledgements

We are very grateful to Eric Anderson, Rob von Behren, Nikita Borisov, Mike Chen, Armando Fox, Jim Gray, Ramki Gummadi, Drew Roselli, Geoff Voelker, the anonymous referees, and our shepherd Bill Weihl for their very helpful suggestions that greatly improved the quality of this paper. We would also like to thank Eric Fraser, Phil Buonadonna, and Brent Chun for their help in giving us access to the Berkeley Millennium cluster for our performance benchmarks.