Towards Automated Performance Tuning for Complex Workloads

Abstract

In this paper we explore the problem of automatically adjusting DBMS multiprogramming levels and memory allocations in order to achieve a set of per-class response time goals for a complex multiclass workload. We start by describing the phenomena that make this a very challenging problem, the foremost of which is the interdependence between classes that results from their competition for shared resources. We then describe M&M, a feedback-based algorithm for simultaneously determining the MPL and memory settings for each class independently, and we evaluate the algorithm's effectiveness using a detailed simulation model. We show that our algorithm can successfully achieve response times that are within a few percent of the goals for mixed workloads consisting of short transactions and longer-running ad hoc join queries. 1 Introduction As database management systems continue to increase in function and to expand into new application areas, the diversity of database workloads is increas...