ITR: Cache-Resident Databases

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

Abstract Databases are at the very heart of the information economy. Database performance is a driving factor that dictates what is possible through the use of information technology. While database management systems have evolved since they were invented several decades ago, their current design is unfortunately antiquated given the state-of-the-art computer memory hierarchies of today (and even more so tomorrow). This project seeks to alleviate this problem. While processor speeds double every year, memory access speeds follow a much shallower improvement curve. To bridge this speed gap, small, fast memories called caches are used to hold frequently accessed data and instructions close to the processor. When executing database workloads, accesses often miss in the (fast) cache and access the (slow) memory, thereby reducing performance significantly. Hardware approaches are typically limited by access time constraints and by applicability to a wide range of workloads. To keep the hardware design feasible, caches typically use simplistic data placement and replacement schemes, and are oblivious to the memory access behavior of the application. Cache-conscious software methods are, on the contrary, extremely promising. The proposed algorithms collect data statistics in order to correctly group data with similar usage patterns and optimize cache utilization. By carefully observing behavior, data is prefetched into the cache before it is used. Preliminary results demonstrate that these techniques (i) minimize the number of misses in the cache and (ii) significantly reduce the incurred penalties.