CAREER: Input-Centric Program Behavior Analysis and Adaptation

通过分析和预测程序动态行为，程序行为分析为程序转换和资源管理提供了基础支持。它的有效性对计算效率的最大化至关重要。本研究提出将项目投入这一迄今几乎被忽视的维度纳入项目行为分析的重点，培养一种新的范式，即以投入为中心的项目行为分析与适应。这种以输入为中心的模式将为增强软件和硬件之间的匹配创造许多新的机会，从而显著提高现代计算的性能和能效。所提出的以输入为中心的程序行为分析和自适应技术由三个部分组成。前两个部分，程序输入表征和输入行为建模，解决了程序输入的复杂性，提取了重要特征，并识别了特征输入特征与程序行为之间的相关性。第三个组件，以输入为中心的适应，利用前两个组件创造的新机会，使动态优化具有前瞻性和整体性，但不会失去对输入和环境变化的适应性。总之，这三个组成部分使可进化的编程系统比以前更可行。在这样的系统中，输入-行为模型体现了中央知识库，它在程序生产运行中逐渐增长。随着知识库变得越来越大，行为预测变得越来越准确，从而促进更好的软硬件匹配，并使程序和运行时系统的性能越来越好。由于程序行为分析在软硬件匹配中的基本作用，本研究有助于为推进软件执行堆栈（编译器、虚拟机、操作系统等）各层的优化铺平道路。

By analyzing and predicting program dynamic behaviors, program behavior analysis offers the fundamental support for program transformations and resource management. Its effectiveness is crucial for the maximization of computing efficiency. This research proposes to include program inputs---a so far virtually ignored dimension---into the focus of program behavior analysis, cultivating a new paradigm, namely input-centric program behavior analysis and adaptation. This input-centric paradigm will create many new opportunities for enhancing the matching between software and hardware, hence significantly improving the performance and power efficiency in modern computing.The proposed technique, input-centric program behavior analysis and adaptation, consists of three components. The first two components, program input characterization and input-behavior modeling, resolve the complexities of program inputs, extract important features, and recognize the correlations between characterized input features and program behaviors. The third component, input-centric adaptation, capitalizes on the novel opportunities that the first two components create, making dynamic optimizations proactive and holistic, but without losing the adaptivity to inputs and environmental changes. Together, the three components make evolvable programming systems more feasible than before. In such a system, the input-behavior models embody the central knowledge base, which grows incrementally across program production runs. As the knowledge base becomes larger, behavior prediction becomes more accurate, stimulating better software-hardware matching and making the program and runtime systems perform increasingly better. Because of the fundamental role of program behavior analysis in software-hardware matching, this research helps pave the way for advancing the optimizations in various layers in the software execution stack (compilers, virtual machines, OS, etc.).