ALCF Announces the DLIO Benchmark
The Argonne Leadership Computing Facility has announced DLIO, the deep learning I/O benchmark developed by researchers at ALCF and the Scalable Computing Software (SCS) Lab at Illinois Tech. The benchmark is maintained in ALCF's GitHub organization (argonne-lcf/dlio_benchmark).
What DLIO Does
DLIO is designed to:
- Identify I/O bottlenecks in deep learning training pipelines
- Optimize data loading strategies for large-scale training jobs
- Evaluate storage system performance under realistic AI workloads
- Inform infrastructure decisions for future systems
Collaborative Development
DLIO was developed by SCS Lab researchers Hariharan Devarajan, Anthony Kougkas, and Xian-He Sun with ALCF computer scientists Huihuo Zheng and Venkatram Vishwanath. The benchmark was built from I/O profiles of eight scientific deep learning applications on ALCF's Theta supercomputer, and the CCGrid 2021 paper describing it received the conference's best paper award. Devarajan began the work as an ALCF intern in 2020.
Why I/O Matters for AI
As deep learning models grow, data loading takes a larger share of training time. DLIO helps researchers and system administrators find and remove these bottlenecks.