Skip to main content

5.1 Scaling Up Data Analysis Jupyter

Source repo: sdsc-summer-institute-2024 | Branch: main | Last synced: 2026-04-24 10:27:17.425 UTC

SDSC Summer Institute 2024​

Session 5.1 Scaling up Interactive Data Analysis in Jupyter Lab: From Laptop to HPC

Date: Thursday, August 7, 2024

Summary: In this session we will demonstrate scaling up data analysis to larger than memory (out-of-core) datasets and processing them in parallel on CPU and GPU nodes. In the hands-on exercise we will compare Pandas, Dask, Spark, cuDF, and Dask-cuDF dataframe libraries for handling large datasets. We also cover setting up reproducible and transferable software environments for data analysis.

Presented by: Peter Rose (pwrose @ucsd.edu)

Reading and Presentations:​

  • Lecture material:

  • Source Code/Examples:

    • Git Repository df-parallel: Comparison of Dataframe libraries for parallel processing of large tabular files

TASK 1: Launch Jupyter Lab on Expanse using a CONDA environment​

  1. Open a Terminal Window ("expanse Shell Access") through the Expanse Portal Training Account

  2. Clone the Git repository df-parallel in your home directory

git clone https://github.com/sbl-sdsc/df-parallel.git
  1. Change into the df-parallel directory
cd df-parallel

  1. Launch Jupyter Lab using the Galyleo script

    This script will generate a URL for your Jupyter Lab session.

galyleo launch --account ${SI24_ACCOUNT} --reservation ${SI24_RES_GPU} --qos ${SI24_QOS_GPU} --partition gpu-shared --cpus 10 --memory 92 --gpus 1 --time-limit 01:00:00 --conda-env df-parallel-gpu --conda-yml environment-gpu.yml --mamba
  1. Open a new tab in your web browser and paste the Jupyter Lab URL. It may take a few minutes to launch your session.

You should see the Satellite Reserver Proxy Service page launch in your browser.

TASK 2: Benchmark Dataframe Libraries using a csv Input File​

For this task you will compare the runtime for a simple data analysis using 5 dataframe libraries.

  1. Go to the Jupyter Lab session

    Navigate to the df-parallel/notebooks directory.

  2. Copy data files

    Run the 1-FetchLocalData.ipynb notebook to copy two data sets of gene information (gene_info.tsv, gene_info.parquet) to the scratch disk on the GPU node.

  3. Run the Dataframe notebooks with a csv input file

    Run the following Dataframe notebooks and write down the runtime shown at the bottom of each notebook.

2-PandasDataframe.ipynb
3-DaskDataframe.ipynb
4-SparkDataframe.ipynb
5-CudaDataframe.ipynb
6-DaskCudaDataframe.ipynb

To get exact timings, run the notebook with the >> (Run All) button!

TASK 3: Benchmark Dataframe Libraries using a Parquet Input File​

In the following notebooks, change the file format to parquet: file_format = "parquet" and run them again. Write down the runtime shown at the bottom of each notebook. How does this compare with using csv files?

2-PandasDataframe.ipynb
5-CudaDataframe.ipynb

To get exact timings, run the notebook with the >> (Run All) button!

TASK 4: Measure Parallel Efficiency​

In this task you will measure and plot the parallel efficiency of using a Dask dataframe with 1, 2, 4, and 8 cores. In the notebook below, select the file format parquet and the dataframe library Dask from the menu. Analyze the parallel efficiency plot and suggest how many cores would be ideal for this task.

7-ParallelEfficiency.ipynb

At the end of the session, don't forget to shutdown the process.

File -> Shutdown