Easy and expressive LLC contention model

Abstract

Last-Level-Cache (LLC) contention modeling is important to ensure good application performance on modern HPC platforms. Existing LLC contention models either rely on Performance Counters (PC), or, Cache Access Pattern (CAP) based data. The former models are easy-to-develop but inexpressive as their modeling power is limited to the contention levels observed during model construction, while the latter are expressive but not easy-to-develop as they require an in-depth understanding of hardware mechanisms that are typically complex, evolving, and protected information. In this paper we propose a hybrid approach for LLC contention modeling and thereby create models that are both easy-to-develop, and expressive. In order to evaluate our approach we construct the LLC contention models for two modern processors (Intel Sandy-bridge and Ivy-bridge) and nine SPEC CPU2006 benchmarks. Our models are easy-to-develop as they are automatically constructed without requiring the complex and protected information on hardware mechanisms. The overall errors in the LLC miss ratio prediction are 5.74% and 7.27%, respectively. The models are also expressive and remain applicable for Exterior-Points, i.e., LLC contention values greater than those observed during model construction. The errors for Exterior-Point prediction are 7.74% and 10.41%, respectively.