Scheduling Policies for Heterogeneous, Approximate Computing Systems

Energy consumption is a primary concern for modern computer systems. Conservative approaches, such as DVFS, which have been used in the past to optimize the performance / power tradeoff have reached their limits. Heterogeneity is a promising approach: devices with different characteristics, each performance- and energy-efficient for specific computational patterns are combined in the same system. Approximate computing is another more disruptive solution: many applications can tolerate controlled quality loss in exchange to significant improvement of performance and energy footprint. In this paper we introduce three scheduling policies that exploit heterogeneity, one of them combining it with approximate computing. These policies can selectively optimize performance, energy consumption, or the tradeoff between energy consumption and quality of results. They monitor the execution of tasks at runtime in order to identify the appropriate mapping of tasks to devices, as well as to control the degree of approximation. Our experimental evaluation indicates that all three policies closely match the effectiveness of the optimal configuration, selected by an "oracle".