Data placement in parallel object-based systems

Parallel object-oriented database management systems provide both the functionality and the high performance required by many application domains. In parallel systems based on the shared-nothing architecture, the database is horizontally declustered across multiple nodes, enabling the system to employ multiple processors and disks to speed up the execution time of navigational queries and improve system throughput. This dissertation quantifies the tradeoffs associated with alternative object placement techniques for parallel Object-Oriented Database Management Systems (OODBMS) executing navigational queries. Two approaches for distributing objects in parallel OODBMSs are described. The first technique, CLASP, focuses on assigning objects to the nodes in order to take advantage of intra-query parallelism and minimize the response times of queries. The second strategy, MINT, assigns objects to the nodes in order to take advantage of inter-query parallelism. MINT attempts to minimize the overall processing overheads of each query while distributing the overall system workload evenly across the nodes in order to maximize the throughput of the system. This dissertation also investigates the effects of utilizing multiple copies of objects on overall system performance. In particular, the effects of utilizing a second copy of each object for query processing are evaluated. Several "availability strategies" have been developed that advocate storing a backup copy of the data in order to keep all data available in the presence of hardware failures. This redundant data can be utilized by the system for query processing when no failures have occurred in order to reduce network overheads. Several algorithms for assigning a backup copy of each object to the nodes are proposed. In order to evaluate the proposed object placement techniques, a simulator was developed to model a parallel OODBMS based on the shared-nothing architecture. Using this simulator, a trace-driven simulation study was executed to evaluate these object placement techniques. The object placement strategies were evaluated with respect to several factors, including: (1) network overheads, (2) I/O overheads, (3) skew in access to the disks, (4) query response time, and (5) system throughput.