PandasBench
收藏资源简介:
PandasBench是一个用于评估Pandas API优化技术的基准测试套件。它包含从Kaggle随机选择的102个真实世界的Jupyter笔记本,总共包含3,721个代码单元格。该数据集旨在评估Pandas API在真实世界代码中的覆盖范围,并为优化技术提供一个高质量的评估基准。数据集的创建过程中,研究者们从Kaggle收集了笔记本,修复了它们,清除了与Pandas API无关的代码,并使它们适合用于基准测试。PandasBench还引入了一种非均匀的数据缩放方案,以更好地模拟真实世界的工作负载。该数据集广泛应用于评估各种Pandas API优化技术,如Modin、Dask、Koalas和Dias,旨在解决现有基准测试不足以评估Pandas API的问题。
PandasBench is a benchmark suite for evaluating Pandas API optimization techniques. It includes 102 real-world Jupyter notebooks randomly selected from Kaggle, with a total of 3,721 code cells. This dataset is designed to assess the coverage of the Pandas API in real-world code and provide a high-quality evaluation benchmark for optimization technologies. During its creation, researchers collected notebooks from Kaggle, repaired them, removed code unrelated to the Pandas API, and adapted them for benchmarking purposes. PandasBench also introduces a non-uniform data scaling scheme to better simulate real-world workloads. This dataset is widely applied to evaluate various Pandas API optimization techniques such as Modin, Dask, Koalas and Dias, aiming to address the issue that existing benchmarks are insufficient for evaluating the Pandas API.
PandasBench数据集概述
数据集简介
- 名称:PandasBench
- 性质:首个针对Pandas API的系统性基准测试
- 内容:包含真实世界的Kaggle笔记本集合,用于评估
pandas及其替代方案(如Modin、Dask和Koalas)的性能
数据集结构
runner/:运行PandasBench的基础设施explore/:用于探索PandasBench的代码notebooks/:真实世界的Kaggle笔记本集合,按原始用户名和内核slug保存scripts/:设置基准测试时使用的实用工具plots/:生成PandasBench论文中图表的代码
设置与运行
Docker镜像
-
Full镜像:包含运行PandasBench所需的所有依赖和原始数据 bash docker run -it --name pandasbench_c_full --rm baziotis/pandasbench:1.0.0-full bash
-
Lite镜像:仅包含Python环境,需手动下载数据和代码
手动设置
- 运行
./scripts/setup_everything.sh(假设Ubuntu环境)
执行步骤
-
准备:缩放输入数据 bash python3 run_all.py --scale_input_and_exit
或 bash python3 run_all.py --scale_input
-
运行基准测试
-
使用
pandas: bash python3 run_all.py -
使用替代Pandas API实现: bash python3 run_all.py --lib <pandas alternative>
-
使用Pandas API补充: bash python3 run_all.py --add_on <add on>
-
扩展PandasBench
- 在
setup_import_lib.py中添加新的Pandas API实现或补充
相关出版物
- 标题:PandasBench: A Benchmark for the Pandas API
- 作者:Alex Broihier, Stefanos Baziotis, Daniel Kang, Charith Mendis
- 出处:Arxiv, 2025
- 链接:https://arxiv.org/abs/2506.02345

- 1PandasBench: A Benchmark for the Pandas API伊利诺伊大学(UIUC) · 2025年



