面向HPC和AI领域的负载特征数据集
收藏资源简介：
为支撑HPC和AI领域负载特征分析与优化，本文构建含数据流图（DFG）、函数调用关系及程序结构的数据集。数据于2023-2025年在清华大学Intel Xeon测试机采集，含polybench测试集10个PDF格式DFG数据，及NPB-MPI 3.3版本8个程序的GML格式函数调用关系与程序结构数据。数据采集采用分层精准解析方案：DFG数据通过程序源码编译为LLVM IR后开展数据流分析，经图结构建模后以dot命令生成可视化PDF；函数调用关系与程序结构数据均通过解析程序二进制文件抽取核心信息，前者以函数为顶点、调用关系为边构建图结构，后者以循环、分支等结构为顶点、包含关系为边构建图结构，均存储为标准GML格式以保障数据通用性。数据时间精度对齐采集过程全流程节点，空间精度精准定位至程序函数及指令级结构，计算方式以静态程序分析与图结构建模为主。数据采集过程通过标准化测试集选取、统一格式存储及多步骤解析校验保障质量，确保数据的完整性与一致性。该数据集为HPC和AI领域负载特征建模、性能优化、编译优化等研究提供了精准的基础数据支撑，可助力相关算法的设计与验证，对推动HPC与AI融合场景下的系统优化、提升计算资源利用效率具有重要的潜在应用价值与研究意义。 数据集包含4个数据文件，DFG数据、函数调用关系数据、程序结构数据。（1）“DFG数据”文件夹下包含10个PDF文件，分别为polybench测试集中对应程序的Data-Flow Graph（DFG），共451KB。（2）“函数调用关系数据”文件夹下包含8个文本文件，分别为NPB-MPI（3.3版本）基准测试集中8个程序对应的函数数调用关系数据progam call graph（pcg），共543KB。（3）“程序结构数据”文件夹下包含8个文件夹，分别对应NPB-MPI（3.3版本）基准测试集中8个程序对应的所有函数结构信息，以标准GML格式的图数据结构存储， 共665KB。（4）“CFG数据与运行时采样数据”文件夹下下包含8个文件夹，分别对应NPB-MPI（3.3版本）基准测试集中8个程序对应的静态CFG数据（静态数据文件夹）和运行时性能采样数据（动态数据文件夹），另有数据集说明文件。
To support load characteristic analysis and optimization in the fields of High Performance Computing (HPC) and Artificial Intelligence (AI), this paper constructs a dataset containing Data-Flow Graphs (DFGs), function call relationships, and program structure data. The data was collected between 2023 and 2025 on an Intel Xeon test machine at Tsinghua University, including 10 PDF-formatted DFG datasets from the PolyBench test suite, and function call relationship and program structure data in GML format for 8 programs from the NPB-MPI 3.3 benchmark suite. Data was collected using a hierarchical precise parsing scheme: For DFG data, the program source code is compiled into LLVM IR, followed by data flow analysis, graph structure modeling, and generation of visual PDF files via the dot command; For function call relationship and program structure data, core information is extracted by parsing program binary files. The former constructs a graph structure with functions as vertices and call relationships as edges, while the latter constructs a graph structure with structures such as loops and branches as vertices and inclusion relationships as edges. Both are stored in standard GML format to ensure data universality. The temporal accuracy is aligned with all nodes throughout the data collection process, while the spatial accuracy precisely locates to program functions and instruction-level structures. The data generation method mainly adopts static program analysis and graph structure modeling. The data collection process ensures quality through standardized test suite selection, unified format storage, and multi-step parsing verification, guaranteeing data integrity and consistency. This dataset provides accurate basic data support for research such as load characteristic modeling, performance optimization, and compilation optimization in the HPC and AI fields, and can assist the design and verification of relevant algorithms. It has important potential application value and research significance for promoting system optimization in the integration scenario of HPC and AI, and improving the utilization efficiency of computing resources. The dataset comprises four data components: DFG data, function call relationship data, program structure data, and CFG data and runtime sampling data. 1. The "DFG Data" folder contains 10 PDF files, which are the Data-Flow Graphs (DFGs) corresponding to the programs in the PolyBench test suite, with a total size of 451 KB. 2. The "Function Call Relationship Data" folder contains 8 text files, which are the program call graph (PCG) data corresponding to 8 programs in the NPB-MPI 3.3 benchmark suite, with a total size of 543 KB. 3. The "Program Structure Data" folder contains 8 subfolders, corresponding to all function structure information of 8 programs in the NPB-MPI 3.3 benchmark suite, stored in standard GML-formatted graph data structures, with a total size of 665 KB. 4. The "CFG Data and Runtime Sampling Data" folder contains 8 subfolders, corresponding to the static Control Flow Graph (CFG) data (static data folder) and runtime performance sampling data (dynamic data folder) of 8 programs in the NPB-MPI 3.3 benchmark suite, plus a dataset description file.
