Collaborative Research: PPoSS: LARGE: General-Purpose Scalable Technologies for Fundamental Graph Problems

该项目旨在加速大型分布式机器上的大型图问题的执行，例如数据中心中的机器。所考虑的图计算出现在计算生物学问题(例如，物种如何进化)、社会网络分析问题和软件系统的验证(例如，如何证明软件是正确的)中。这些问题有许多基本子计算的共同点，本项目将加快这一点。研究人员将确定执行这些子计算的新方法，这些方法更有效率，并将构思能够更快执行这些计算的新计算硬件。我们的社会将受益，因为这项工作将使我们能够更快、更少地解决这些问题的更大版本。此外，该项目还包括一个教育项目，将向高中、大学本科生和研究生教授计算机科学，重点是来自弱势背景的学生。所考虑的图问题的挑战既源于所使用的算法的复杂性，也源于许多图问题的大量计算和存储需求。为了应对这些挑战，该项目进行了一项雄心勃勃的跨层工作，基于三个相互依赖的主要目标：用于图形问题的新算法，用于有效执行这些问题的核心软件框架，以及为这些问题提供加速的异类硬件。第一个重点集中在所考虑的应用领域的几个高回报算法方向：静态和动态设置中的图形聚类；在保留重要信息的同时构建图形；以及机器学习(ML)技术的应用。在所有这些方向上，该项目都使用近似值。在第二个推力中，我们开发了一个灵活的编程层，可以为数据中心规模的平台生成高效的代码。该项目介绍了一个图形编程框架，其中包含一种用于图形的新型领域特定语言(DSL)、用于图形处理的高性能数值库(使用可扩展的稀疏方法)，以及一个具有两个使用机器学习(ML)技术的中间表示的智能编译器。在第三个推力中，该项目使用新型硬件加速器加快了图形应用程序在大型分布式机器上的执行速度。该加速器采用高级指令集体系结构(ISA)，其中的指令可在瓦片上执行稀疏矩阵运算。智能自动调谐器软件帮助生成代码并将其映射到各种加速器和通用引擎。调查人员是伊利诺伊大学厄巴纳-香槟分校、麻省理工学院和印第安纳大学的十名教授，他们在几个不同的领域拥有专业知识。这项工作将与工业研究团体密切合作。这一奖项反映了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。

This project seeks to accelerate the execution of large graph problems on large, distributed machines, such as those found in datacenters. The graph computations considered appear in computational biology problems (for example, how species evolved), social network analysis problems, and verification of software systems (for example, how to prove that software is correct). These problems have many basic sub-computations in common, which this project will accelerate. The investigators will identify new ways to perform these sub-computations that are more efficient and will conceive new computing hardware that can execute them faster. Our society will benefit because this work will enable solving bigger versions of these problems faster and with less energy consumption. In addition, the project includes an education program that will teach computer science to high-school, college undergraduate and graduate students, with an emphasis on students from disadvantaged backgrounds. The challenges of the graph problems considered stem from both the complexity of the algorithms used and the large compute and storage requirements of many graph problems. To address these challenges, this projects pursues an ambitious, cross-layer effort based on three interdependent main thrusts: new algorithms for graph problems, a core software framework for the efficient execution of these problems, and heterogeneous hardware to provide acceleration to these problems. The first thrust focuses on a few high-payoff algorithmic directions for the application domains considered: graph clustering in both static and dynamic settings; graph construction while preserving important information; and the application of machine learning (ML) techniques. In all these directions, the project uses approximations. In the second thrust, we develop a flexible programming layer that generates efficient code for a datacenter-scale platform. The project introduces a graph programming framework with a novel Domain-Specific Language (DSL) for graphs, high-performance numerical libraries for graph processing with scalable sparse methods, and a smart compiler with two intermediate representations that uses machine learning (ML) techniques. In the third thrust, the project speeds up the execution of graph applications in a large, distributed machine with a novel hardware accelerator. The accelerator features a high-level Instruction Set Architecture (ISA) with instructions that perform sparse matrix operations on tiles. A smart auto-tuner software helps generate and map code to various accelerators and general-purpose engines. The investigators are ten professors at the University of Illinois Urbana-Champaign, MIT, and Indiana University, with expertise in several distinct areas. The work will be done in close collaboration with industrial research groups.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.