Price theory based power management for heterogeneous multi-cores

Reviewer: Cristiana Bolchini

Runtime resource management is a hot topic given the increasing interest in, and adoption of, system architectures constituted by a number of computing resources. Examples of such resources include multicore and many-core systems used in scenarios where the workload cannot always be characterized in advance, meaning that design-time approaches only offer a suboptimal solution. In this scenario, a few solutions have been proposed in the literature (the authors mention some of them). These solutions are generally aimed at dynamically optimizing resource allocation to the applications or tasks to be executed, with respect to a selected figure of merit (such as power consumption). The aspect that attracted my attention is the use of a strategy taken from a different context, price theory, to drive the optimization policy. In fact, within the performance/power optimization scenario, control theory or artificial intelligence (AI)-driven solutions are typically adopted, while only a few works exploit economic theory or welfare economics to tackle the problem, as the authors discuss toward the end of the paper in the related work section. Starting from the idea of borrowing the resource management strategy from economic theory, the authors present the many details of their proposal, which is aimed at optimizing power consumption for heterogeneous multicore systems and constituted by computing resources characterized by different performance and power profiles. To be more specific, the authors select the big.LITTLE multicore produced by ARM, comprising two different types of cores. The paper initially introduces all of the models (such as architecture, application, and power), setting the background for the introduction of the proposed framework devoted to power management, which is presented in section 3. This section is the core of the work, and the authors introduce all of the elements of the framework, including the dynamics and mechanisms at the basis of the management strategy. To help the reader follow the discussion, a running example is adopted, allowing for the exemplification of complex chip dynamics. As anticipated, an overview of the existing approaches aimed at optimizing power consumption is presented, leading to a discussion of experimental results (section 5) and a comparison of the solution to the one available on the adopted big.LITTLE architecture. More precisely, the section presents the adopted setup for the experimental campaigns (such as the selected workload) and introduces “the HL scheduler released by Linaro in Linux kernel release 3.8.” The analysis is carried out by comparing performance and power consumption for the proposed and alternative solutions. The results show interesting savings without severely affecting performance. The approach to scalability is also discussed since the complexity of the management strategy (at both design time and runtime) may constitute a limitation as the size of the workload and/or the number of resources increases. Conference proceedings papers usually have a limited number of available pages; this is not true for the 19th International Conference on Architectural Support for Programming Languages and Operating Systems. As a consequence, the authors were able to describe in depth all aspects of their proposal in depth, allowing the reader to get a precise idea of the solution, which is well supported by the reported experimental campaigns. Thinking about an extension of the work for a journal publication, it would be nice to see a comparison with other solutions (such as control theory-based ones) coming from the research community (instead of the “normal” HL scheduler). Online Computing Reviews Service