Cache-Oblivious Query Processing

We propose a radical approach to relational query processing that aims at automatically and consistently achieving a good performance on any memory hierarchy. We believe this automaticity and stableness of performance is at times more desirable than some peak performance achieved through careful tuning, especially because both database systems and hardware platforms are becoming increasingly complex and diverse. Our approach is based on the cache-oblivious model, in which data structures and algorithms are aware of the existence of a multi-level memory hierarchy but do not assume any knowledge about the parameter values of the hierarchy, such as the number of levels in the hierarchy, the capacity and the block size of each level. Since traditional database systems are intrinsically aware of these parameters, e.g., the memory page size, our cache-oblivious approach requires us to rethink the query processing architecture and implementation. In this position paper, we present the architectural design of our cache-oblivious query processor, EaseDB, discuss the technical challenges and report our initial results.