Performance comprehension of production software

在加拿大和美国，软件开发人员都是薪酬最高、最受欢迎的人才之一；仅加拿大就面临着18.2万名ICT员工的缺口。随着生产率的提高，短缺至少可以得到部分缓解。我们的项目将解决软件开发生产力中一个重要且经常被忽视的方面：性能调试和调优。最近的一项调查显示，92%的开发人员在前一年面临影响客户的性能问题；总而言之，约300名受访工程师花了130个人月的时间诊断和修复性能问题。受访者进一步承认，性能调整最困难的部分实际上并不是修复问题，而是收集和分析证据并找到其根本原因。*存在许多性能工具，但它们主要关注的是收集性能事件，如操作时间戳或硬件指标，而不是数据分析。因此，工程师们经常被大量数据淹没，但他们几乎不知道如何有效地利用这些数据。如今的工程师要处理大量不是他们编写的代码，这进一步加剧了性能调优：在不理解代码的情况下筛选性能数据几乎没有效率。*基于这些观察，我们项目的目标是将性能分析和程序理解统一起来。我们渴望设计方法、算法，并最终构建可视化丰富的工具，这些工具不仅使工程师能够有效地导航数十亿个性能事件，而且还使开发人员能够理解程序流、数据结构布局和数据访问模式，同时将它们与性能指标相关联。我们希望这些工具将帮助工程师同时了解程序是如何工作的，它是如何访问数据的，它是如何执行的，以及为什么它会这样执行。

Software developers are among the most highly paid and sought-after talent in both Canada and the US; Canada alone facing a gap of 182,000 ICT workers. Shortage can be at least partially relieved with improved productivity. Our project will address an important and often overlooked aspect of software development productivity: performance debuggingtuning. A recent survey revealed that 92% of developers faced customer-affecting performance issues during the preceding year; all in all, ~300 interviewed engineers spent 130 man-months diagnosing and fixing performance problems. The respondents further admitted that the most difficult part of performance tuning was not actually fixing the problem, but gathering and analyzing evidence and getting to its root cause. ******Many performance tools exist, but their primary focus is collecting performance events, such as operation timestamps or hardware metrics, and not on the analysis of the data. As a result, engineers are often swamped with swaths of data, but have little idea how effectively use it. Performance tuning is further aggravated by the fact that today's engineers deal with a lot of code they did not write: sifting through performance data without understanding the code is hardly efficient. ******Based on these observations, the goal of our project is to unite performance analysis and program comprehension. We aspire to design methodologies, algorithms, and ultimately build visually rich tools that will to engineers not only enable effectively navigating through billions of performance events, but also enable the developer to comprehend the flow of the program, the layout of data structures and data access patterns, while at the same time relating them to performance metrics. Our hope is that these tools will help the engineer to simultaneously understand how the program works, how it accesses data, how it performs, and why it performs the way it does.