Future Technologies Colloquium Series

ABSTRACT:

Emerging many-core systems present unique challenges to system software. This talk presents methods to address two of these challenges, heterogeneity and layered parallelism, and show how these methods can improve the scalability of applications on both symmetric and asymmetric many-core systems. Heterogeneity and layered parallelism have been addressed so far in limited contexts, primarily through high-level guidelines to programmers. In this talk, I will present models -both empirical and analytical-, and system software support for synthesizing multiple forms of parallelism, and coping with the implications of asymmetry in core configurations and the memory hierarchy. More specifically, I will present a unified framework for modeling, synthesizing, and scheduling polymorphic parallelism on asymmetric multi-core processors, and demonstrate its effectiveness with results from a PS3 cluster and a Cell blade cluster. I will also present extensions of the Xen Hypervisor for optimally configuring cores in layered multi-core architectures to support parallel applications with diverse memory usage patterns.