Abhinav Bhatele
University of Maryland, MD·Computer Science·Early-career
Supercomputers and large data centers run software that is split across thousands of processors working at once. Squeezing good performance out of such machines normally depends on human experts tweaking settings by hand, a slow process full of guesswork and mistakes.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
high performance computing
parallel computing
performance modeling
- Grants
- 1 known grant, through Feb 2027
What they work on
- Build machine learning models of parallel application performance from collected software performance data, and use them to automatically select high-performance configurations instead of relying on hand tuning.
- Develop automated methods to analyze large-scale longitudinal monitoring data from parallel systems, identifying performance anomalies and their root causes, and use trained models to tune system operation.
- Apply transfer learning so performance models built on current machines can automatically tune parallel codes as they are ported to new or future architectures.
- Deliver a suite of frameworks that analyze performance data gathered at the job, system, and facility levels and make decisions targeting operational efficiency metrics in HPC and data center systems.
More about the research
This project asks whether machine learning (ML), trained on the large volumes of performance data these systems already generate, can take the human guesswork out of that loop so that parallel software and the systems running it become self-tuning: able to measure their own behavior and adjust themselves.
Why it matters
Scientific simulations and data analyses in many fields wait on supercomputer time. If parallel software and HPC facilities can tune themselves from their own performance data, individual jobs run faster, more jobs fit in the same machine, and data centers waste less energy on poorly configured runs. The project's techniques could also spare application teams the manual re-tuning normally required each time code moves to a new architecture.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Mahmut Kandemir
Pennsylvania State University, PA·Computer Science
Characterize diverse HPC workloads on real and simulated computational storage devices to identify which computations benefit from running inside storage, and what performance tradeoffs result.
Timothy Zhu
Pennsylvania State University, PA·Computer Science·Early-career
Develop software approaches for managing partially broken consumer and datacenter machines so their remaining working components can be used as-is, rather than repaired or discarded.
Shivaram Venkataraman
University of Wisconsin-Madison, WI·Computer Science·Early-career
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).
The school
University of Maryland: 49 labs on PI Pages, $1.5B a year on research, 602 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.