Cristiana Amza - Research Summary

Future research plans

• Scaling Dynamic Content Sites through Data Replication

  I have designed and implemented a replication technique using a novel scheduler design that provides scaling, strong data consistency and availability for dynamic content servers. The technique, called conflict-aware scheduling [USITS '03], improves scaling by using a lazy replication scheme with asynchronous updates, and by avoiding conflicts at each replica. At the same time, the user sees serializable executions.

• Novel Database Concurrency Control

  I have implemented a new concurrency control algorithm based on explicit versions for the database back-ends in a replicated dynamic content server. This algorithm reduces the overhead of consistency maintenance by optimizing the duration of conflicts in problem applications with frequent conflicts. Distributed versioning [MiddleWare '03] is a novel replication technique that uses this explicit versioning concurrency control in conjunction with a conflict-aware scheduler. Distributed versioning provides both the simplicity in application semantics and programming of eager replication, and performance close to the best achievable using lazy replication for most workloads studied.

• Performance Studies of Dynamic Content Servers

  I have performed an analysis of scale and performance in dynamic content web sites. In particular, in collaboration with other researchers at Rice, I have implemented three common dynamic content applications [WWC-5 '02]: e-commerce (TPC-W ), on-line bidding (modeled after eBay ), and bulletin board (modeled after Slashdot). Using these applications, I have studied the common bottlenecks and the effect of admission control and load balancing [TR '02], and transparent query caching [OSDI '02 submission] for a dynamic content server.

• Data Replication for Persistent Memory

  I have designed and implemented novel replication algorithms for applications that use in-memory persistent storage. The algorithms [DSN '00], provide both reliability and data availability while still maintaining very high transaction throughput. I investigated four possible designs in a primary-backup configuration, using a cluster of commodity servers connected by a write through system area network. I showed that logging approaches outperform mirroring approaches, because of better locality and in spite of communicating more data. I also showed that the logging versions scale well to small shared-memory multiprocessors.

• Software Distributed Shared Memory Systems

  My contributions in this area include a novel adaptive algorithm that switches between a single-writer and a multiple-writer protocol [HPCA '97]. This algorithm provides the benefits of lazy release consistency in software distributed shared memory systems (SDSM) while reducing its computation and memory overheads whenever the application pattern allows it. Later work explores using different consistency units in page-based SDSM [PPoPP '97], and an extension of my previous dynamic protocol adaptation in SDSM to include adaptation of the consistency unit to fit the application pattern [IEEE '99].

Back home