Cooperative scans

Data mining, information retrieval and other application areas exhibit a query load with multiple concurrent queries touching a large fraction of a relation. This leads to individual query plans based on a table scan or large index scan. The implementation of this access path in most database systems is straightforward. The Scan operator issues next page requests to the buffer manager without concern for the system state. Conversely, the buffer manager is not aware of the work ahead and it focuses on keeping the most-recently-used pages in the buffer pool. This paper introduces cooperative scans -- a new algorithm, based on a better sharing of knowledge and responsibility between the Scan operator and the buffer manager, which significantly improves performance of concurrent scan queries. In this approach, queries share the buffer content, and progress of the scans is optimized by the buffer manager by minimizing the number of disk transfers in light of the total workload ahead. The experimental results are based on a simulation of the various disk-access scheduling policies, and implementation of the cooperative scans within PostgreSQL and MonetDB/X100. These real-life experiments show that with a little effort the performance of existing database systems on concurrent scan queries can be strongly improved