Software Power Analysis And Optimization For Power-Aware Multicore Systems

摘要

Among all the factors in sustainable computing, power dissipation and energy consumption, arguably speaking, are fundamental aspects of modern computer systems. Different from performance metric, power dissipation is not easy to measure because hardware instrumentation is usually required. Yet as an indispensable component of a computer system, software becomes a major factor affecting power dissipation besides hardware energy-efficiency and power states. With detailed information on resource usage and power dissipation of an application/software, software developers will be able to leverage algorithms and implementations in order to produce power-efficient solutions. Hardware instrumentation, despite its accuracy, is costly and complicated to set up. A general solution to connect software with hardware along with detailed power and system information will improve the system overall efficiency.In this work, we design and implement a general solution to analyze and model software power dissipation. Based on the analysis, we propose a combined solution to optimize the energy efficiency of parallel workload. Starting from the hands-on power measurement method in detail, we provide a fine-grain power profile of two computer systems using hardware instrumentation.Being focusing on dynamic power dissipation analysis, we propose a two-level power model for power-aware multicore computer systems. Based on the model, we design and implement SPAN to relate power dissipation to the different portions of an application using the proposed power model. By using SPAN, developers can easily identify the sections of code consuming the most power in the program. Alternatively, to enable automatic source code instrumentation, we utilize compiler techniques to insert profiling code before and after each function in source code. The expected outcome includes an open source function level power profiling tool, Safari. Using the profiling tools,we propose a model to capture the relationship between concurrency (C), power (P) and execution time (T). By changing the system configuration for different parallel workload, we are able to achieve optimal/near optimal energy-efficient execution of a given workload on a specific platform.