SHF: Medium: Spectral Profiling: Understanding Software Performance without Code Instrumentation

动态分析（例如概要分析程序执行）被广泛使用，因为它们可以度量软件系统运行时行为的各个方面，并且在软件工程中具有广泛的应用。这些分析通常是通过向软件添加探针来执行的，这会增加空间/时间开销，是侵入性的，并且会对软件行为产生负面影响。为了解决这些问题，我们提出了一种新颖的方法，可以通过利用计算机在执行代码时产生的电磁发射来准确且非侵入性地分析软件行为。我们的方法可以通过简单地在系统旁边放置一个设备来收集关于软件系统的运行时信息。因此，它不仅可以对以前不可能的各种软件系统（例如，嵌入式系统）进行分析，而且还可以在更传统的环境中进行动态分析。该项目将结合各种机器学习和静态分析技术，为不同的代码模式构建可能的电磁签名，研究哪种代码粒度提供最准确的电磁发射与代码匹配，并探索在运行时执行此匹配的自适应和分层技术。这项研究本质上是跨学科的，有望在软件工程、编程语言、计算机体系结构和电磁学等几个综合领域开辟新的领域并产生更广泛的影响。与之前的电磁发射分析工作不同，我们的方法将收集足够细粒度的运行时信息，以测量短序列语句的执行情况，如果不是单个指令的话。这将使我们能够将我们的方法应用到几个软件工程任务中。事实上，如果成功，这项研究将提供一个坚实的概念基础，其他研究人员将能够利用它，并研究一组特定的技术和工具，这些技术和工具建立在这个基础上，以支持诸如零开销性能测量、调试和异常检测等任务。

Dynamic analyses such as profiling program execution are widely used because they can measure various aspects of the runtime behavior of a software system and have a wide range of applications in software engineering. These analyses are typically carried out by adding probes to the software, which imposes space/time overhead, is intrusive, and can negatively affect software behavior. To address these issues, we propose a novel approach that allows for analyzing software behavior accurately and non-intrusively by leveraging the electromagnetic emissions produced by a computer as it executes code. Our approach can collect runtime information about a software system by simply placing a device next to the system. It can thus not only enable profiling for a variety of software systems for which this was previously impossible (e.g., embedded systems), but also benefit dynamic analyses in more traditional contexts.This project will combine various machine learning and static analysis techniques to build likely electromagnetic signatures for different code patterns, investigate which code granularity provides the most accurate matching of electromagnetic emissions to code, and explore adaptive and hierarchical techniques for performing this matching at runtime. This research is inherently interdisciplinary and promises to break new ground and have broader impact in several combined areas, including software engineering, programming languages, computer architecture, and electromagnetics. Unlike previous work on electromagnetic emissions analysis, our approach will collect runtime information that is fine-grained enough to measure the execution of short sequences of statements, if not individual instructions. This will let us apply our approach to several software engineering tasks. In fact, if successful, this research will both provide a solid conceptual foundation, which other researchers will be able to leverage, and investigate a set of specific techniques and tools that build on this foundation to support tasks such as zero-overhead performance measurement, debugging, and anomaly detection.