A Hybrid Framework Bridging Locality Analysis and Cache-Aware Scheduling for CMPs

Industry is rapidly moving towards the adoption of Chip Multi-Processors (CMPs). The sharing of memory hierarchy becomes deeper and heterogeneous. Without a good understanding of the sharing, most current systems schedule processes in a contention-oblivious way, causing systems severely underutilized with sub-optimal throughput and cache thrashing. In this report, we propose a three-stage framework to analyze shared-cache locality. It is based on inclusive locality model and unifies online and offline adaptive analysis. Unlike previous methods, inclusive locality model addresses all factors of cache contention at the same time. The goal is to produce a comprehensive understanding of the relations between program characteristics and run-time behavior in shared-cache systems, meanwhile developing a scalable adaptive contention-aware scheduling system. The preliminary experiments demonstrate the potential benefits of contention-aware scheduling on CMPs, and the promise of accurate runtime locality measurement, a critical component of the locality analysis framework.