Automated performance optimization based on post-mortem parallel performance analysis

事后并行性能分析允许识别并行程序执行中的性能瓶颈。尽管最近在性能分析自动化方面取得了进展，但是仍然需要大量的人力来利用分析结果并优化程序。本文将研究用于静态和动态性能优化的新自动化技术，利用APT小组开发的程序执行分析工具和库。这些工具使用人工神经网络来自动识别性能异常及其可能的原因，但仍然需要专业程序员进行干预以提高性能。这是一项耗时且容易出错的任务，本项目旨在实现自动化。该项目的目的是研究替代启发式方法进行动态程序优化的可能性，特别是针对非统一内存访问体系结构系统以及大规模分布式内存系统中的工作调度和数据放置问题。通过更强大的机器学习技术，可以自动适应新的执行环境或动态发展的执行条件（例如，节点故障需要重新平衡负载，动态应用程序行为，如在网格细化中创建运行时不平衡等）。曼彻斯特大学之前的工作表明，自动检测此类事件的发生是可能的，并且在某种程度上，可以确定性能下降的原因。这提供了一个令人兴奋的机会，可以在这项研究的基础上进行构建，并将最后缺失的部分自动化。优化器输入数据的主要来源将是compperf，本文将研究如何缓解已识别的瓶颈。它可以包括重新编译应用程序代码，修改程序二进制或改变并行执行中的调度和数据位置。在第一阶段，将选择一组基准来建立执行基线。它将包括公开可用的应用程序，但也可以使用能够更好地显示所应用优化的有效性的应用程序进行扩展。来自欧盟项目EuroEXA（754337）的大规模应用，如天气和气候建模，物理和生命科学模拟。将使用三个主要指标来评估优化器：吞吐量、延迟和功率效率。人工神经网络的使用有可能成为自动化性能优化过程的合适工具，因为它们需要最少的假设，并且能够基于不完整的数据进行有效的推广。一种可能的训练设置是使用执行跟踪和初始应用程序信息作为输入，并将参考优化作为目标。最后，本文旨在表明，这些方法可以用于推广不同的架构，从现成的英特尔平台到EuroEXA超级计算机原型。

Post-mortem parallel performance analysis allows to identify performance bottlenecks in the parallel program execution. Despite recent advances in performance analysis automation, a significant human effort is still required to take advantage of the analysis results and optimize programs. The thesis will investigate new automation techniques for static and dynamic performance optimization, leveraging the program execution analysis tools and libraries developed in the APT group. These tools employ artificial neural networks to automatically identify performance anomalies and their likely causes, but still require an expert programmer to intervene in order to improve performance. This is a time-consuming, error-prone task which this project aims to automate.The aim of the project is to investigate the possibility of replacing heuristic approaches for dynamic program optimization, in particular targeting the problems of work scheduling and data placement in non-uniform memory access architecture systems as well as in large scale distributed memory systems, with more robust machine learning techniques that can automatically adapt to either new execution environments or to dynamically evolving execution conditions (e.g. node failure requiring to re-balance the load, dynamic application behaviour such as in mesh refinement creating run-time imbalance, etc.). Previous work at The University of Manchester has shown that it is possible to automatically detect the occurrence of such events and, to some extent, to identify the causes of performance degradation. This presents an exciting opportunity to build on top of this research and to automate the last missing piece.The main source of the input data to the optimizer will be ComPerf and this thesis will investigate how identified bottlenecks can be mitigated. It can either involve recompiling the application code, modification of the program binary or changing the schedule and data placement in the parallel execution. In a first stage, a set of benchmarks will be selected to establish the execution baseline. It will include publicly available applications, but also can be extended with applications that better shows effectiveness of the applied optimizations. Large-scale applications from the EU project EuroEXA (754337), such as weather and climate modeling, physics and life science simulation. Three leading metrics will be used to asses the optimizer: throughput, latency and power efficiency.The use of Artificial Neural Networks has a potential to be a suitable tool to automate the performance optimization process as they require minimum assumptions and are able to effectively generalize based on incomplete data. One of the possible training settings is to use execution traces and initial application information as an input and a reference optimization as the target. Finally, this thesis aims to show that such approaches can be used to generalize over different architectures, from off-the-shelf Intel platforms to the EuroEXA supercomputer prototype.