Deep Learning I/O Benchmark

Deep Learning I/O (DLIO) Benchmark is a benchmark suite aiming at emulating the I/O pattern and behavior of deep learning applications. The benchmark is delivered as an executable that can be configured for various deep learning workloads. It uses a modular design to incorporate different data loaders, data formats, dataset organizations, and use training configuration parameters similar to the actual deep learning applications. It is able to represent the I/O process of a broad spectrum of deep leanrning applications.

The main features of DLIO include:

• Easy-to-use configuration through YAML files which represent the I/O process of different deep learing applications.

• Easy-to-use data generator capable to generate synthetic datasets of different formats, different data organizations and layouts.

• Full transparency over emulation of I/O access with logging and profiling at different levels with DFTracer.

• Supporting emulating both sequential training and distributed data parallel training.

GitHub repo: https://github.com/argonne-lcf/dlio_benchmark.

---

Overview

• Introduction
• High-level Design

Getting Started

• Installation
• DLIO Configuration
• How to create a DLIO configuration YAML file
• Running DLIO
• Examples

Custom data loader and reader plugins

• Creating a Data Loader Plugin
• Creating a Custom Data Reader
• Creating a Checkpointing Plugin

Tested systems and Known issues

• Tested systems
• Instructions for running DLIO Benchmark on Lassen@LLNL
• Limitations and future works

How to contribute

• Contributing Guide

Resources

• Resources

Acknowdgments

• Acknowledgments

Appendix

• Analysis on JPEG data generator
• Profiling

Legal

• Copyright
• License

Indices and tables

• Index

• Module Index

• Search Page