Efficient Sampling Startup for Sampled Processor Simulation

Modern architecture research relies heavily on detailed pipeline simulation. Simulating the full execution of an industry standard benchmark can take weeks to months. Statistical sampling and sample techniques like SimPoint that pick small sets of execution samples have been shown to provide accurate results while significantly reducing simulation time. The inefficiencies in sampling are (a) needing the correct memory image to execute the sample, and (b) needing a warm architecture state when simulating the sample. In this paper we examine efficient Sampling Startup techniques addressing two issues: how to represent the correct memory image during simulation, and how to deal with warmup. Representing the correct memory image ensures the memory values consumed during the sample's simulation are correct. Warmup techniques focus on reducing error due to the architecture state not being fully representative of the complete execution that proceeds the sample to be simulated. This paper presents several Sampling Startup techniques and compares them against previously proposed techniques. The end result is a practical sampled simulation methodology that provides accurate performance estimates of complete benchmark executions in the order of minutes.