Memory reference metrics and instruction trace sampling

Trace-driven simulation models have been widely used to accurately estimate performance of proposed microarchitectures. Accuracy of the estimated performance greatly depends on the instruction traces as well as the model itself. Although long instruction traces are preferred, e.g. to fill a large cache, a trace could often be redundantly long; it might merely represent a small, tight loop. In this paper we try to select representative samples from a given trace consisting of multiple samples. We introduce three new metrics for memory references: the Access Scatter Function, the Block Execution Interval, and the Local Execution Intervals. These new metrics and basic block sizes, branch distances, execution intervals, and the Stack Growth Function are used to characterize each sample. Then, samples are clustered to select representative samples. Analysis of two real instruction traces of batch and database workloads in a large mainframe is shown as an example.