DLIO Wins Best Paper Award (First Prize) at CCGrid'21
The paper "DLIO: A Data-Centric Benchmark for Scientific Deep Learning Applications" has received the Best Paper Award (First Prize) at the 2021 IEEE/ACM International Symposium on Cluster, Cloud, and Internet Computing (CCGrid'21). The work, authored by Hariharan Devarajan, Huihuo Zheng, Anthony Kougkas, Xian-He Sun, and Venkatram Vishwanath, introduces a benchmarking framework that exposes the I/O bottlenecks hidden in deep learning training pipelines.
The problem
Deep learning is used across scientific domains such as cosmology, particle physics, fusion, and astrophysics. Much work has gone into the computational performance of deep learning frameworks, but far less into their I/O, even though training relies on large and varied datasets and I/O is a significant bottleneck in large-scale distributed training.
DLIO
The paper studies the I/O behavior of several scientific deep learning workloads on the Theta supercomputer at the Argonne Leadership Computing Facility. From those profiles it builds DLIO, a benchmark suite that emulates the I/O behavior of modern scientific deep learning applications. Application developers and system architects can use it to find I/O bottlenecks and guide optimizations; the paper reports training times lowered by up to 6.7x.
The team
The authors are Hariharan Devarajan (then a PhD student in the SCS Lab, now at Lawrence Livermore National Laboratory), Huihuo Zheng and Venkatram Vishwanath of Argonne National Laboratory, Anthony Kougkas, and Xian-He Sun. The work grew out of a collaboration with the Argonne Leadership Computing Facility.
Continued development
The MLPerf Storage benchmark from MLCommons is built on DLIO, and development continues in the argonne-lcf/dlio_benchmark (opens in a new tab) repository.