Load Stall Minimization

负载失速最小化

• 批准号: 536287-2018
• 负责人: Kent, Kenneth
• 金额: $ 3.08万
• 依托单位: University of New Brunswick
• 依托单位国家: 加拿大
• 项目类别: Collaborative Research and Development Grants
• 财政年份: 2021
• 资助国家: 加拿大
• 起止时间: 2021-01-01 至 2022-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=720119
• 关键词: Load; Stall; Minimization

The genesis of the idea for this project was an observation made while performing the analysis of profiles from Java benchmarks being run on a Java virtual machine. We observed a number of instances where there were methods in the top 10-20 of the profile who had 80% or so of their execution attributed to a small number of instructions and that when we viewed those instructions the execution time appeared to be due to load stalls. Drawing from this the first goal in this project is to identify sources of stalls in a selection of Java workloads and try to understand the semantics of the program's behaviour at the point of the stall so we could classify the stalls. The goal is to understand where stalls are happening and try to identify patterns with a view to informing techniques to reduce the stalls. Possible approaches to reducing the stalls fall in two broad families of techniques: 1) instruction scheduling and prefetch insertion to mitigate the stalls and; 2) improve object placement in the garbage collector. While both approaches are possible in combination, the benefit of each and the blended performance is unknown.

这个项目的想法的起源是在对运行在Java虚拟机上的Java基准测试执行配置文件分析时所做的观察。我们观察到了许多实例，在配置文件的前10-20个方法中，80%左右的执行归因于少量指令，并且当我们查看这些指令时，执行时间似乎是由于加载停滞。由此得出结论，该项目的第一个目标是在选定的Java工作负载中识别停滞的来源，并尝试理解停滞时程序行为的语义，以便我们可以对停滞进行分类。目标是了解摊位在哪里发生，并试图确定模式，以期告知减少摊位的技术。减少停顿的可能方法分为两大类技术：1)指令调度和预取插入以减轻停顿；2)改进垃圾收集器中的对象放置。虽然这两种方法可以组合在一起，但每种方法和混合性能的好处是未知的。