Abstract

Recent benchmark suite releases such as Parsec specifically utilize the tightly coupled cores available in chip-multiprocessors to allow the use of newer, high performance, models of parallelization. However, these techniques introduce additional irregularity and complexity to data sharing and are entirely dependent on efficient communication performance between processors. This paper thoroughly examines the crucial communication and sharing behaviour of these future applications on an accurate full-system model.

The infrastructure used allows both accurate and comprehensive program analysis, employing a full Linux OS running on a simulated 32-core x86 machine. Experiments use full program runs, with communication classified at both core and thread granularities. Migratory, read-only and producer-consumer sharing patterns are observed and their behaviour characterized. The temporal and spatial characteristics of communication are presented for the full collection of Splash-2 and Parsec benchmarks. Our results aim to support the design of future communication systems for CMPs, encompassing coherence protocols, network-on-chip and thread mapping.