Simulation methodology for decision support workloads

Abstract

The impact of any new architectural proposal must be evaluated under realistic working conditions. This class of analysis requires trustworthy simulation tools and representative workloads that allow us to know the real effectiveness of the improvement. In this work, we propose a methodology that enables the use of an important family of transactional workloads, such as decision support system workloads, in a full system simulator. In contrast to numerical applications, with this type of workload it is not possible to scale down the problem size in order to reduce the computational requirements of the simulation. We will show the stationary behaviour of the workload and how it can be employed to reduce computational requirements without significant loss. Taking into account this fact, we will show how simulating only 3% of the benchmark the maximum error in the main system performance metrics is approximately 5%.