Date: Mon, 25 Nov 1996 22:41:56 GMT Server: NCSA/1.5 Content-type: text/html Last-modified: Fri, 29 Sep 1995 15:00:09 GMT Content-length: 3612 Art Bernstein

Art Bernstein

Professor, Department of Computer Science

Ph.D. 1962, Columbia University


My interest centers on concurrent and distributed algorithms, database systems, and transaction processing.

I am a member of the High Performance Transaction Processing Group

-------

Traditional algorithms for controlling concurrent access to shared data, by requiring serializable execution, may unduly restrict concurrency, and hence performance, in situations where data is accessed heavily or distributed or where transactions run for long periods of time. We have explored a variety of techniques for dealing with these problems, including the use of optimistic algorithms, multi-version data, and replication.

Recently we developed algorithms for improving concurrency by allowing limited violations of database integrity constraints. Synchronization can be relaxed if integrity constraints are not strictly enforced. By controlling synchronization, perhaps in a state-dependent way, the extent of the violation can be controlled. Violations can be corrected through appropriate compensation.

We are currently investigating the extent to which the semantics of transactions (expressed as proofs in a formal system) can be exploited to improve performance. We use semantics in several ways: to define a new correctness criterion for concurrent (non-serializable) transaction execution, to decompose transactions into smaller units so that locks can be released early, and to design a new concurrency control that guarantees correct execution when these units are interleaved. We are taking two approaches to transaction decomposition. In the first we decompose a transaction into a sequence of steps. Steps are atomic and isolated and release all conventional locks when they complete. A new concurrency control and lock mode is required to implement this approach. The second approach guarantees correctness using only conventional locks, but in a non-two-phase fashion. The algorithms are being implemented and a test bed constructed to evaluate the ideas.

We are also interested in the use of transaction semantics to better understand problems associated with federated databases and compensation.

Selected Recent Publications



Return to the department home page
If you have problems with this page, please send E-mail to:
gerstl@cs.sunysb.edu