Enabling Compositionality for Multicore Timing Analysis

Abstract

Timing compositionality is assumed by almost all multicore timing analyses. In this paper, we show that compositional timing analysis can be incorrect even for simple microarchitectures with in-order execution. We then introduce three approaches to enable sound compositional analysis: two based on analysis and one based on a hardware modification. In the experimental evaluation we explore the strengths and weaknesses of these three approaches. One of the two analysis-based approaches provides an attractive trade-off between analysis cost and precision, enabling sound compositional timing analysis even for microarchitectures with out-of-order execution.