OAC Core: SHF: SMALL: ICURE -- In-situ Analytics with Compressed or Summary Representations for Extreme-Scale Architectures

OAC 核心：SHF：SMALL：ICURE——针对超大规模架构的压缩或摘要表示的原位分析

• 批准号: 2034850
• 负责人: Gagan Agrawal
• 金额: $ 50万
• 依托单位: AUGUSTA UNIVERSITY RESEARCH INSTITUTE, INC.
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-07-01 至 2023-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2034850&HistoricalAwards=false
• 关键词: OAC; Core; SHF; SMALL; ICURE

Systems for High Performance Computing (HPC) have been providing rapidly increasing computing power. However, this growth has also led to systems where the memory and data movement bandwidth is relatively lower. This makes analyzing the data from scientific simulations very challenging. A paradigm called in-situ analytics has emerged in response. This project is further improving this paradigm, by using what can be referred to as homomorphic compressions. The idea of homomorphic compression is to compress the data in a way that queries can be directly executed on the compressed data (without need for decompression). This project is developing such compression methods, developing techniques to perform such compression efficiently on Graphic Processing Units (GPUs), techniques for query processing using such compressed representations, and finally, an overall system that will simplify development of in-situ analytics implementations. Overall, this project will be making analysis of data from simulations more effective on the upcoming systems for HPC. This project will seek to broaden participation in computing through direct participation in the project development teams by undergraduate and graduate students from under-represented groups. Systems for High Performance Computing (HPC) have been providing rapidly increasing computing power. However, this growth has also led to systems where the memory and data movement bandwidth is relatively lower. This makes analyzing the data from scientific simulations very challenging. A paradigm called in-situ analytics has emerged in response. This project is further improving this paradigm, by using what can be referred to as homomorphic compressions. The idea of homomorphic compression is to compress the data in a way that queries can be directly executed on the compressed data (without need for decompression). The resulting framework, ICURE, can facilitate in situ analytics on accelerators themselves, reduce overall memory requirements for the analytics, reduce total data movements costs, and even reduce the time cost of performing the analytics. Achieving the goals of ICURE involves many open challenges. The first is the choice of summarization structure and its constructions. This project experiments with two different summary or concise representations: bitmap indices and an integrated value index. The second issue is analyses methods using summary and compressed representations, where the focus is on the use of these representations for a variety of analyses tasks: computing aggregations, correlations, value-based joins, time-step selection, and interesting subregions analysis. The third issue is automating placement and quality. Driven by the consideration of providing the lowest interference between the simulation and analytics, this project automates decisions on placement of specific analytics operations and data within the node of HPC system. Similarly, automatic selection of sampling level driven by desired accuracy and overheads of the analyses is performed. This project will seek to broaden participation in computing through direct participation in the project development teams by undergraduate and graduate students from under-represented 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.

高性能计算 (HPC) 系统一直在提供快速增长的计算能力。然而，这种增长也导致系统内存和数据移动带宽相对较低。 这使得分析科学模拟的数据变得非常具有挑战性。 作为回应，一种称为原位分析的范式应运而生。 该项目通过使用所谓的同态压缩来进一步改进这一范例。 同态压缩的思想是以可以直接对压缩数据执行查询的方式来压缩数据（无需解压）。 该项目正在开发此类压缩方法、在图形处理单元 (GPU) 上高效执行此类压缩的技术、使用此类压缩表示进行查询处理的技术，以及最终将简化现场分析实现开发的整体系统。 总体而言，该项目将使即将推出的 HPC 系统上的模拟数据分析更加有效。 该项目将寻求通过来自代表性不足群体的本科生和研究生直接参与项目开发团队来扩大对计算的参与。 高性能计算 (HPC) 系统一直在提供快速增长的计算能力。然而，这种增长也导致系统内存和数据移动带宽相对较低。 这使得分析科学模拟的数据变得非常具有挑战性。 作为回应，一种称为原位分析的范式应运而生。 该项目通过使用所谓的同态压缩来进一步改进这一范例。 同态压缩的思想是以可以直接对压缩数据执行查询的方式来压缩数据（无需解压）。 由此产生的框架 ICURE 可以促进加速器本身的原位分析，减少分析的总体内存需求，减少总数据移动成本，甚至减少执行分析的时间成本。 实现 ICURE 的目标涉及许多开放的挑战。首先是摘要结构及其构造的选择。该项目尝试了两种不同的摘要或简洁表示：位图索引和综合值索引。第二个问题是使用摘要和压缩表示的分析方法，其中重点是将这些表示用于各种分析任务：计算聚合、相关性、基于值的连接、时间步长选择和有趣的子区域分析。 第三个问题是自动化布局和质量。出于在模拟和分析之间提供最低干扰的考虑，该项目自动决定在 HPC 系统节点内放置特定分析操作和数据。类似地，执行由所需精度和分析开销驱动的采样水平的自动选择。该项目将寻求通过来自代表性不足群体的本科生和研究生直接参与项目开发团队来扩大对计算的参与。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评估进行评估，被认为值得支持。影响审查标准。

项目成果

End-to-End LU Factorization of Large Matrices on GPUs

GPU 上大型矩阵的端到端 LU 分解

• 发表时间: 2023-02
• 期刊: Proceedings of PPoPP 2023
• 作者: Yang Xia; Peng Jiang
• 通讯作者: Peng Jiang

Scaling and Selecting GPU Methods for All Pairs Shortest Paths (APSP) Computations

为所有对最短路径 (APSP) 计算缩放和选择 GPU 方法

• DOI: 10.1109/ipdps53621.2022.00027
• 发表时间: 2022-05
• 期刊: Proceedings of the International Parallel and Distributed Processing Symposium
• 作者: Xia, Yang;Jiang, Peng;Agrawal, Gagan;Ramnath, Rajiv
• 通讯作者: Ramnath, Rajiv

GCD2: A Globally Optimizing Compiler for Mapping DNNs to Mobile DSPs

GCD2：用于将 DNN 映射到移动 DSP 的全局优化编译器

• DOI: 10.1109/micro56248.2022.00044
• 发表时间: 2022-10-01
• 期刊: 2022 55th IEEE/ACM International Symposium on Microarchitecture (MICRO)
• 作者: Wei Niu;Jiexiong Guan;Xipeng Shen;Yanzhi Wang;G. Agrawal;Bin Ren
• 通讯作者: Bin Ren

DNNFusion: accelerating deep neural networks execution with advanced operator fusion

DNNFusion：通过高级算子融合加速深度神经网络的执行

• DOI: 10.1145/3453483.3454083
• 发表时间: 2021-06
• 期刊: Proceedings of PLDI 2021
• 作者: Niu, Wei;Guan, Jiexiong;Wang, Yanzhi;Agrawal, Gagan;Ren, Bin
• 通讯作者: Ren, Bin

GPU Adaptive In-situ Parallel Analytics (GAP)

GPU 自适应原位并行分析 (GAP)

• 发表时间: 2022-10
• 期刊: Proceedings of the Conference on Parallel Architectures and Compilation Techniques
• 作者: Haoyuan Xing; Gagan Agrawal
• 通讯作者: Gagan Agrawal