Architecture Sensitive Database Design: Examples from the Columbia Group.
Kenneth Andrew Ross, John Cieslewicz, Jun Rao, Jingren Zhou · 2005
In this article, we discuss how different aspects of modern CPU architecture can be effectively used by database systems. While we shall try to cover most important aspects of modern architectures, we shall not aim to provide an exhaustive survey of the literature. Rather, we will use examples from our own recent work to illustrate the principles involved. While there is related work on using other components of modern machines such as graphics cards and intelligent disk controllers for database work, we limit our discussion here to just a single CPU and the memory hierarchy (excluding disks). In recent years the primary focus of database performance research has fundamentally shifted from disks to main memory. Systems with a very large main memory allow a database’s working set to fit in RAM. At the same time processors have continued to double in performance roughly every eighteen months in line with “Moore’s Law, ” while memory speed improvements have been marginal. The CPU now spends many hundreds of cycles waiting for data to arrive from main memory, and database workloads are often memory-bound. Because of this bottleneck, database research at Columbia and elsewhere has focused on effectively using the cache hierarchy to improve database performance. Changes in processor technology have also influenced database research. Modern processors feature a very long pipeline, which helps to provide better performance. It also results in a higher branch misprediction penalty,