CAREER: An Algorithm and System Co-Designed Framework for Graph Sampling and Random Walk on GPUs

图形分析是解决我们这个时代的重大挑战的关键技术之一，例如了解流行病的传播，设计超大规模集成电路和发现软件漏洞等。然而，随着图的大小继续增长，学习、挖掘和计算这样巨大的图变得无效、不切实际，并且可能是可怕的。幸运的是，Graph Sampling和Random Walk可以大大减少原始图的大小，同时仍然可以捕获下游图分析任务所需的属性。但是，一个全面的系统，可以执行图采样和随机游走在现实世界中的万亿边图在一个可接受的速度是缺席。这项研究开创了将各种图形采样和随机游走算法统一在一个用户友好的框架之后的努力，该框架可以利用世界一流的图形处理单元（GPU）计算设施，包括未来的exascale计算设施，以快速处理万亿边图形。该项目有助于实现美国增加科学和工程参与的国家目标，这对美国成功应对全球挑战，建立更强大，更多样化的劳动力队伍以及满足全球创新经济的需求至关重要。该项目产生了一个高性能的软件库，作为来自学术界，国家实验室和工业界的科学和工程从业人员的基础工具。通过有趣的投资和奖励教育计划，致力于帮助K-12，本科生，女性和代表性不足的少数民族（URM）人口在科学，技术，工程和数学（STEM）领域，该项目制定了一个全面的路线图，为下一代高性能图形分析专业工作者和研究人员做好准备。这个项目改造和创建核心课程，在研究生和本科水平的PI的家庭部门。为了使整个社会受益，该项目将项目数据，软件和出版物传播到更广泛的研究社区http://personal.stevens.edu/~hliu77/gsrw.html.The本研究的首要目标是使图形采样和随机游走快速，可扩展和用户友好。为此，这个职业建议倡导算法和系统共同设计的研究。首先，本研究介绍了新的更新和建设设计的各种主要的Monte Carlo方法的转移概率是必不可少的快速采样。其次，为了充分释放GPU的潜力，该项目将关键原语公式化为可以利用GPU上的通用和保留张量和光线跟踪核心的问题。第三，基于图采样和随机游走的异步处理特性，本研究利用远程直接内存访问（RDMA）辅助的任务和分区自适应调度机制，以减少可扩展的万亿边图采样和随机游走的数据传输。最后但并非最不重要的是，这项职业研究提供了一个以偏置为中心的框架，它为最终用户提供了表现力，不仅可以编程各种现有的GSRW算法，还可以编程未来的算法，并且通过隐藏上述先进的优化技术来实现简单性。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

Graph analytics is one of the key technologies to address the grand challenges of our time, such as understanding the spread of pandemics, designing extremely large-scale integrated circuits and uncovering software vulnerabilities among many others. However, as the size of the graph continues to grow, learning, mining and computing such gigantic graphs become ineffective, impractical, and potentially dire. Fortunately, Graph Sampling and Random Walk can dramatically reduce the size of the original graphs, while still capturing the desired properties for downstream graph analytics tasks. But a comprehensive system that can perform graph sampling and random walk on real-world trillion-edge graphs at an acceptable speed is absent. This research pioneers the effort of uniting various graph sampling and random walk algorithms behind a user-friendly framework that can take advantage of world-class Graphics Processing Unit (GPU) computing facilities, including the future exascale ones, to rapidly handle trillion-edge graphs. This project contributes to the U.S. national goal of increasing participation in science and engineering, which is crucial to America’s success in addressing global challenges, building a stronger and more diversified workforce, and meeting the needs of the global innovation economy. This project produces a high-performance software library that serves as a foundational tool for fellow science and engineering practitioners from academia, national laboratories and industry. With a commitment to helping K-12, undergraduate, female, and Underrepresented Minority (URM) populations in the Science, Technology, Engineering, and Mathematics (STEM) field through the interesting investment and rewarding education plan, this project lays out a comprehensive road map to prepare the next-generation high-performance graph analytics professional workers and researchers. This project revamps and creates core courses in both graduate and undergraduate levels for the PI's home department. To benefit the society at large, this project disseminates the project data, software, and publications to the broader research community at http://personal.stevens.edu/~hliu77/gsrw.html.The overarching goal of this research is to make graph sampling and random walk fast, scalable and user-friendly. Towards that end, this career proposal advocates algorithm and system co-designed researches. First, this research introduces novel update and construction designs for transition probability of various major Monte Carlo methods that are essential for fast sampling. Second, to fully unleash the potential of GPUs, this project formulates the key primitive into problems that can take advantage of general, and reserved tensor and ray tracing cores on GPUs. Third, based upon the asynchronous processing nature of graph sampling and random walk, this research exploits Remote Direct Memory Access (RDMA)-assisted task and partition adaptive scheduling mechanism to reduce the data transfers for scalable trillion-edge graph sampling and random walk. Last but not the least, this career research delivers a bias-centric framework, which offers end users expressiveness to program not only a variety of exiting GSRW algorithms but also future ones, and simplicity by hiding the aforementioned advanced optimization techniques.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.