SPX: Collaborative Research: Pinpointing and Resolving Scalability Culprits Hidden in Different Components of the Whole System Stack

现代计算机利用多核或多核处理器来加速并行应用程序。不幸的是，由于一些隐藏的可伸缩性问题，这些应用程序的加速通常远非理想。以前的研究主要集中在应用程序代码上，以识别可伸缩性瓶颈，而忽略了应用程序代码与许多外部组件（包括内存分配器、第三方运行时库和操作系统）交互的事实。因此，理解和修复可伸缩性问题不应局限于应用程序代码，而应考虑整个软件堆栈。该项目的新奇之处在于，它可以精确地找出隐藏在整个堆栈不同组件中的可伸缩性问题，并自动修复可伸缩性瓶颈。该项目的影响是显著提高了运行在多核处理器上的应用程序的性能，从而加速了科学发现和节能。该项目旨在从用户空间系统地查明和解决整个软件堆栈的所有组件中潜在的软件争用。由于多核和多核硬件的普遍使用，所提出的方法是紧迫的。此外，根据Amdahl定律，任何组件中潜在的少量争用都可能极大地限制这些现代硬件的加速潜力。该研究计划设计低开销的分析器来获取系统调用、内存分配器行为和组件之间所有交互事件的运行时信息，以及分析器来自动查明可伸缩性瓶颈的根本原因。通过运行时优化器，该研究旨在修复已确定的可伸缩性问题，而无需程序员的干预。该项目有可能大大减少软件优化的人工工作，并提高现代硬件上并行应用程序的性能。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Modern computers leverage multi-core or many-core processors to accelerate parallel applications. Unfortunately, the speedup of these applications is typically far from ideal, due to some hidden scalability issues. Previous research mainly focuses on application code to identify scalability bottlenecks, neglecting the fact that the application code interacts with numerous external components, including memory allocator, third-party runtime libraries, and the operating system. Understanding and fixing scalability problems should hence go beyond application code and consider the whole software stack. The project's novelties are to pinpoint scalability culprits hidden in different components of the whole stack and automatically fix the scalability bottlenecks. The project's impacts are significantly improved performance for applications running on multi-core processors and thus accelerated scientific discoveries and energy saving.This project aims to systematically pinpoint and resolve latent software contention in all components of the whole software stack from user space. The proposed approaches are urgent due to the pervasive use of multi-core and many-core hardware. Also, according to Amdahl's law, a small degree of latent contention in any of the components may substantially limit the speedup potential on these modern hardware. The research plans to design low-overhead profilers to obtain runtime information for system calls, memory allocator behaviors, and all interacting events between components, as well as analyzers to automatically pinpoint the root causes of scalability bottlenecks. Through a runtime optimizer, the research aims to fix the identified scalability issues without intervention from the programmer. The project has potential to dramatically reduce manual effort for software optimization and improve performance for parallel applications on modern hardware.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.