• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

Parallelism is key to high performance relational database systems. Since there are several parallel architectures suitable for database systems, a few interesting problems arise, mostly from an emphasis on the differences among the architectures. Specifically, in the literature, differences rather than similarities between the architectures are pointed out, and the specific details of a particular architecture, crucial to high performance, are generally ignored. In this thesis we have attempted to remedy this situation by emphasizing the similarities and a deeper understanding of two popular parallel architectures, shared nothing and shared memory, from a database perspective. We show that there is complementarity and similarity in the two architectures by showing that software shared-memory support can be used to improve performance on shared-nothing hardware and by showing that shared-nothing software can run on shared-memory hardware with performance comparable to that of "native " algorithms. We also show that by understanding the architectural details and tradeoffs, we can design algorithms that have superior performance. We illustrate this via examples of hash join algorithms on shared-memory hardware that exploit cache memories, hash aggregation algorithms on sharednothing hardware that tradeoff communication for memory consumption, and hash aggregation algorithm on shared-memory hardware that tradeoff computation for reduced latch conflicts. All these algorithms show performance superior to the previously known algorithms. ii

Citations