CDS&E: An Effective Thermal Simulation Methodology for GPGPUs Enabled by Data-Driven Model Reduction

近年来，由于科学、工程和统计计算的需要，对通用图形处理单元(GPGPU)的需求迅速增长。与此同时，GPGPU也迅速成为全球数据中心的重要组成部分。由于最近社交网络、电影流媒体、在线购物、大数据、物联网等的爆炸性增长，数据中心的数量正在急剧增长。由于每个GPGPU中运行着数百或数千个内核，严重的加热是一个严峻的挑战，除非采用有效的冷却，否则会显著降低GPGPU的性能、可靠性和能效。然而，数据中心的有效冷却需要巨大的能源消耗。为了缓解这些问题，需要对GPGPU运行进行有效的热管理和热感知任务调度，但这需要一个准确的仿真工具，能够以合理的空间分辨率提供高效的动态热预测。目前，缺乏高效率、高精度、分辨率合理的热模拟工具。这项工作旨在开发一种基于简化学习算法的高效模拟方法，该算法能够以高分辨率预测GPGPU中的准确动态温度分布。随着这种新方法在GPGPU中的实施，有效的热管理和任务调度将成为可能，并将提高GPGPU的性能和可靠性。这也将提高在冷却、计算和流媒体方面的能源节约，并将地球的环境压力降至最低。该项目还将有助于跨学科的劳动力培训，并使学生为GPGPU计算中新出现的供热问题做好准备。与拟议工作有关的研究将被纳入由私人投资促进机构教授的几门课程。课程项目将由从事拟议工作的博士和本科生开发。这将为本科生和研究生提供课本和讲座之外的有用的学习体验。PIS还将扩大和整合几项正在进行的活动，以扩大代表不足的群体在STEM中的参与，例如通过联合PIS的NSF REU网站。将特别努力招募和指导来自PI大学附近的印第安人保留地的美洲原住民加入STEM活动并在STEM中追求他们的职业生涯。该项目的目标是开发一种多块模拟方法，通过简化的学习算法来高效、准确地预测GPGPU的动态热分布。为了减少仿真空间，从而减少计算时间，同时保持精确的热解，将GPGPU的区域结构投影到由简化学习方法获得的一组基函数描述的函数空间上。然而，这种投影学习过程需要收集整个GPGPU的大量热数据，并且在计算上是困难的。因此，应用域分解将GPGPU域划分为数百个较小的通用构建块。这种构建块方法使得能够更有效地训练基函数以开发多块热模型。与直接数值模拟相比，这种方法将半导体芯片热模拟的计算时间减少了几个数量级。目前，GPGPU的热模拟依赖于高效紧凑的电阻-电容(RC)热模型，该模型分辨率低，热分布不准确。预计所开发的热模拟模型将比紧凑型RC模型更高效。此外，多块方法具有高效并行计算的天然优势。该项目将在GPGPU的数百个核心中实施开发的多块模型，以执行并行GPGPU计算，这将进一步加快GPGPU的热模拟。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Demands for general purpose graphics processing units (GPGPUs) in recent years have increased rapidly due to the needs for scientific, engineering and statistical computing. Meanwhile, GPGPUs are also quickly becoming an essential part of data centers around the globe. The number of data centers are growing drastically due to the recent explosion of social networking, movie streaming, online shopping, big data, internet of things, etc. With hundreds or thousands of cores running in each GPGPU, severe heating is a serious challenge which can significantly degrade GPGPU performance, reliability and energy efficiency unless effective cooling is employed. However, effective cooling of data centers requires enormous expenditure of energy. To ease all these problems, effective thermal management and thermal-aware task scheduling for GPGPU operation are needed, which however requires an accurate simulation tool that is able to offer efficient dynamic thermal prediction with a reasonable spatial resolution. Currently, there is a lack of thermal simulation tools that offer high efficiency and accuracy with a reasonable resolution. The proposed work aims to develop an efficient simulation methodology based on a reduced learning algorithm that is capable of predicting accurate dynamic temperature distributions with a high resolution in GPGPUs. With this novel approach implemented in GPGPUs, effective thermal management and task scheduling will become possible and will improve GPGPU performance and reliability. This will also improve energy savings in cooling, computing and streaming and minimize the earth’s environmental stress. This project will also contribute to interdisciplinary workforce training and prepare students for the emerging challenge of heating problems in GPGPU computing. Research related to the proposed work will be integrated into several courses taught by the PIs. Course projects will be developed by the Ph.D. and undergraduate students working on the proposed work. This will offer undergraduate and graduate students a useful learning experience beyond the textbooks and lectures. The PIs will also expand and integrate several ongoing activities to broaden participation of underrepresented groups in STEM, e.g. through the Co-PI's NSF REU site. A special effort will be made to recruit and mentor Native Americans from an Indian Reservation near the PI’s university to join STEM activities and to pursue their careers in STEM.The goal of this project is to develop a multi-block simulation methodology for efficient, accurate prediction of dynamic thermal profiles of GPGPUs derived from a reduced learning algorithm. To reduce simulation space and thus the computational time while maintaining accurate thermal solution, the domain structure of a GPGPU is projected onto a functional space described by a set of basis functions obtained from the reduced learning method. This projection learning process however requires collection of massive amounts of thermal data for the entire GPGPU and is computationally prohibitive. Domain decomposition is therefore applied to partition the GPGPU domain into hundreds of smaller generic building blocks. This building-block approach enables more efficient training of the basis functions to develop the multi-block thermal model. This methodology offers a reduction in the computational time by several orders of magnitude for thermal simulation of semiconductor chips, compared with the direct numerical simulation. Currently, thermal simulations of GPGPUs rely on the efficient compact resistance-capacitance (RC) thermal model that provides poor resolution and inaccurate thermal profiles. It is expected that the developed thermal simulation model will be even more efficient than the compact RC model. Also, the multi-block approach possesses a natural advantage of effective parallel computing. This project will implement the developed multi-block model in hundreds of cores in a GPGPU to perform parallel GPGPU computing that will further speed up the thermal simulation of GPGPUs.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.