2016 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) (2016)
Chicago, IL, USA
May 23, 2016 to May 27, 2016
Achieving high performance has traditionally been the primary goal of large-scale system design. However, future systems will place an increasing emphasis on energy efficiency. Coarse tools are available today to reduce the energy consumption of system components that are not on the performance critical path, but we expect future systems to make use of fine-grained power scaling and gating features to make the most efficient use of tightly constrained power budgets. The tools to effectively utilize these mechanisms, however, remain largely undeveloped. Dynamic Power Steering is one such methodology for dynamically routing power across an imbalanced system to resources where it can make the most impact. With Dynamic Power Steering, a heuristic algorithm assigns processor cores to available p-states (i.e., power states) in order to optimize energy consumption while maintaining performance levels. In this work, we present a modeling methodology that quantifies the impact Dynamic Power Steering will have on both performance and overall energy usage. This modeling methodology is applicable to both current and future large-scale systems capable of providing fine-grained power scaling. Using synthetic workloads that can be tuned to capture a wide variety of application behaviors, we validate the resulting models and pinpoint areas in which further performance and energy efficiency gains are possible.
Distributed processing, Conferences, Government
K. J. Barker and D. J. Kerbyson, "Modeling the Performance and Energy Impact of Dynamic Power Steering," 2016 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), Chicago, IL, USA, 2016, pp. 1380-1389.