CInC: Workload Characterization In Context of Resource Contention

Increased contention for hardware resources is a consequence of more, distinct workloads running on the same system; this is especially true for a shared cache. However, shared cache analysis is dominated by methods that center on how a workload behaves when run alone, or in isolation. Though evaluating a workload in isolation is important, it is increasingly irresponsible to not evaluate a workload under contention as a standard part of the analysis.Characterizing in Context, or CInC is both a position paper on modern performance analysis and a framework of contention-forward performance analysis. This work provides evidence that it is no longer responsible to make decisions using a workload’s isolation behavior. Additionally, we provide a way to talk about the contention that is formal and clear. Further, we provide two methods for cache sensitivity analysis built on top of the CInC framework: one simplifies multi-capacity-curve analysis via a Condensed Representation Model and the other distills capacity curves into Curve Analysis Features. We also expand curve analysis metrics by expressing the rich information available in a workload’s contention behavior through a novel measurement of stability under contention. The tools enable a contention-forward characterization of a subset of SPEC 2017 rate workloads run on a real system. Additionally, we present a case study exploring a CInC-based SVM classifier which we apply towards co-scheduling (accuracy 88% vs 78% with a clustering tool).