Integrating complete-system and user-level performance/power simulators: the SimWattch approach

Most applications driving the advancements in microarchitecture and memory system research have a non-negligible interaction with the operating system. Yet, most architectural investigations are based on user-level simulators in which operating system activity is not modelled. This has motivated us to design SimWattch, a microarchitectural modeling infrastructure. SimWattch is based on Simics – a system-level simulation tool – and Wattch (SimpleScalar extended with power modeling) – a flexible user-level simulation tool. As a result, it can analyze performance and power dissipation in microarchitectures at the cycle level for complex workloads running on commodity operating systems. In this paper, we present the design issues we had to resolve in integrating a system-level with a user-level simulator. We then use SimWattch to identify the type of errors a user-level simulator typically does when predicting performance and power dissipation while omitting operating system activity. Our results, which are based on applications such as SPEC95, SPEC JVM98, and TPC-B, show that the error can be large for commercial workloads such as TPC-B, in which IPC rate is overestimated by more than 20% and total energy used is underestimated by more than 100%. Although errors are lower for other benchmarks, they are still surprisingly high. This paper demonstrates 1) that O/S activity must be modelled for accurate microarchitecture evaluations 2) that modelling O/S activity is feasible and its cost is reasonable, and 3) that complex and realistic microarchitecture evaluations including O/S activity can be done in reasonable simulation times today.