SHF: Small: Latency Tolerance Aware Runtime Optimization for General-Purpose GPU Architectures

计算机几乎涉及我们现代生活的方方面面——从基因测序到物理模拟，再到为互联网提供动力，从实时图像和语音识别到预测股市趋势。高性能加速器（如图形处理器）的可编程性进步使得大量不同的通用算法能够在gpu上享受性能加速。然而，由于独特的延迟容忍GPU设计特性通常没有考虑到，最先进的解决方案导致次优性能改进。提出的延迟容忍感知运行时优化框架（LATTE）可以帮助计算行业实现其高性能和高能效的愿景。大型和多样化输入数据集的重要通用算法的性能加速对所有研究领域和社会的进步具有深远的影响。研究议程由侧重于异构计算的教育议程补充。本项目提出的LATTE框架旨在探索和提出架构解决方案，并创建系统支持，以加速gpu上重要通用应用程序的执行。提出的LATTE运行时优化框架围绕着识别和设计完全感知延迟容差的优化技术，以最大限度地提高gpgpu的性能。分析了直接将以cpu为中心的优化应用于GPU架构的缺陷，并利用这些缺陷来激发主动管理高度时间复用的计算和内存资源的需求，同时考虑到GPU的临界延迟容忍特性。研究结果可作为未来性能优化研究的基础，避免盲目复制，为GPGPU加速做出重要的科研贡献。

Computing touches almost all aspects of our modern day lives -- from gene sequencing to physics simulations to powering the Internet, from real-time image and voice recognition to predicting stock market trends. The programmability advancement of high-performance accelerators, such as graphics processors, has enabled a large, diverse set of general-purpose algorithms to enjoy performance acceleration on GPUs. However, because the unique latency tolerance GPU design feature is often not taken into account, the state-of-the-art solutions lead to sub-optimal performance improvement. The proposed Latency Tolerance Aware runtime optimization framework (LATTE) can help the computing industry realize its high performance and high energy efficiency vision. Performance acceleration of important general-purpose algorithms with large and diverse input data sets has a profound impact on the advancement of all research domains and on society. The research agenda is complemented by an education agenda focusing on heterogeneous computing.The LATTE framework proposed in this project aims to explore and propose architectural solutions and to create system supports to accelerate the execution of important general-purpose applications on GPUs. The proposed LATTE runtime optimization framework revolves around identifying and designing optimization techniques that are fully latency tolerance aware in order to maximize performance improvement for GPGPUs. The pitfalls for directly adapting CPU-centric optimization to GPU architectures are analyzed and used to motivate the need to proactively manage the highly time-multiplexed computation and memory resources considering the critical latency tolerance characteristic of GPUs. The findings can serve as foundations for future performance optimization research to avoid blind replication, leading to important scientific research contributions for GPGPU acceleration.