Shivaram Venkataraman

What they work on

1. Characterize performance variability across many types of accelerator hardware, where identical applications run at different speeds on chips of the same type, and identify which applications are most affected (Fortuna).
2. Design cluster scheduling mechanisms that use measured accelerator variability and knowledge of each application to improve resource utilization in accelerator-rich clusters.
3. Build Diamond, a web-service interface for HPC deep learning that automates container setup, distributed-training scaling, hyperparameter tuning and model sharing for domain scientists.
4. Develop Diamond performance optimizations, including planning for long training jobs, performance-aware model placement, cross-cluster training and data management.
5. Create software that makes training machine learning models on large graph datasets faster and easier across many machines, including efficient handling of graphs that grow with new data.
6. Reduce the cost of making predictions (inference) with machine learning models trained on large graph datasets.

More about the research

Why it matters

Better use of accelerator hardware could let data centers and supercomputers serve more users with the same equipment. Diamond could let scientists in fields outside computer science apply deep learning without HPC expertise. Faster, cheaper graph learning tools could benefit data scientists working on social networks, drug discovery and search. Open-source releases are intended to support adoption by other clusters and by industry.

Known grants

From NIH and NSF research grants. Labs often have other funding we can't see.

Closest labs

The school

University of Wisconsin-Madison: 91 labs on PI Pages, $1.9B a year on research, 854 research doctorates a year.

More about University of Wisconsin-Madison

Reach out

A start for your email. Fill in the brackets.