Awesome

HPCA'22 Paper: AI-Enabling Workloads on Large-Scale GPU-Accelerated System: Characterization, Opportunities, and Implications

Production high-performance computing (HPC) systems are adopting and integrating GPUs into their design to accommodate artificial intelligence (AI), machine learning, and data visualization workloads. To aid with the design and operations of new and existing GPU-based large-scale systems, we provide a detailed characterization of system operations, job characteristics, user behavior, and trends on a contemporary GPU-accelerated production HPC system. Our insights indicate that the pre-mature phases in modern AI workflow take up significant GPU hours while underutilizing GPUs, which opens up the opportunity for a multi-tier system. Finally, we provide various potential recommendations and areas for future investment for system architects, operators, and users.

Dependencies

We have used Python version 3.7 and Jupyter Notebook version 6.2. The required packages can be installed using

pip install -r requirements.txt

Download data

The SuperCloud data is available for download from the Amazon Open Data Registry via the following bucket:

s3://mit-supercloud-dataset/2022-hpca/

There are three files available:

• dcgm.csv : This file contains the GPU utilization information for each GPU assigned to each job, including power consumption, resource usage.
• scheduler_data.csv: This file contains the scheduler-level information about each job id, including start/end/wait time, user id, job type.
• nvidia_smi.csv: This file contains the time-series information captured by nvidia-smi, for about 2000 jobs. Data is recorded every 100ms.

To download, please first install the Amazon Web Service Command Line Interface AWS CLI here

Once AWS CLI is installed, run the following command to list the availble files:

aws s3 ls s3://mit-supercloud-dataset/2022-hpca/ --no-sign-request

The nvidia_smi.csv file is 42GB. For demonstration purpose, you just need to download dcgm.csv and scheduler_data.csv to be able to run the provided notebook. Run the following command to store the data at the Downloads directory:

mkdir ~/Downloads/hpca22_supercloud
aws s3 cp s3://mit-supercloud-dataset/2022-hpca/dcgm.csv ~/Downloads/hpca22_supercloud --no-sign-request
aws s3 cp s3://mit-supercloud-dataset/2022-hpca/scheduler_data.csv ~/Downloads/hpca22_supercloud --no-sign-request

Demonstration

We have provided a jupyter notebook file to demonstrate how to use the data to generate some of the figures in the paper. You can find the notebook here

Usage

Please feel free to email me any questions about the dataset. My email: li.baol@northeastern.edu

If you are interested in using the dataset, please cite this paper.

@INPROCEEDINGS{9773216,
  author={Li, Baolin and Arora, Rohin and Samsi, Siddharth and Patel, Tirthak and Arcand, William and Bestor, David and Byun, Chansup and Roy, Rohan Basu and Bergeron, Bill and Holodnak, John and Houle, Michael and Hubbell, Matthew and Jones, Michael and Kepner, Jeremy and Klein, Anna and Michaleas, Peter and McDonald, Joseph and Milechin, Lauren and Mullen, Julie and Prout, Andrew and Price, Benjamin and Reuther, Albert and Rosa, Antonio and Weiss, Matthew and Yee, Charles and Edelman, Daniel and Vanterpool, Allan and Cheng, Anson and Gadepally, Vijay and Tiwari, Devesh},
  booktitle={2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA)}, 
  title={AI-Enabling Workloads on Large-Scale GPU-Accelerated System: Characterization, Opportunities, and Implications}, 
  year={2022},
  volume={},
  number={},
  pages={1224-1237},
  doi={10.1109/HPCA53966.2022.00093}}