This paper presents HeteroMates, a solution that uses heterogeneous processors to extend the dynamic power/performance range of client devices. By using a mix of different processors, HeteroMates offers both high performance and reduced power consumption. The solution uses core groups as the abstraction that groups a small number of heterogeneous cores to form a single execution unit. Group heterogeneity is exposed as multiple heterogeneity (H) states, an interface similar to the P-state interface already used for frequency scaling. An H-state controller governs H-state transitions based on dynamic policies maximizing performance or minimizing power consumption, while a ‘core switcher’ transparently migrates tasks to the appropriate core, i.e., the one matching the chosen H-state. Experimental evaluations use real-world client applications and a unique experimental testbed comprised of heterogeneous cores and a shared uncore component. Results show that core groups can provide significant performance improvements while also lowering energy consumption for a diverse set of applications when compared to homogeneous processor configurations. Also demonstrated is the importance of ‘uncore’ power in total SoC power consumption and the need for uncore power scalability when seeking to extend a platform’s dynamic power range.
- Georgia Institue of Technlogoy: Vishal Gupta,
- Intel Labs: Paul Brett, David Koufaty, Dheeraj Reddy, Scott Hahn, Karsten Schwan
- Intel Corporation: Ganapati Srinivasa
International Green Computing Conference (IGCC12) http://www.green-conf.org [PDF(/igcc12.pdf)