M Mustafa Rafique
Rochester Institute of Technology, NY·Computer Science
Long scientific computer programs often save snapshots of their state partway through, called checkpoints. These were first used so a program could restart after a crash.
- Fields
Computer Science
Engineering
- Topics
checkpointing
checkpoint-restart
high-performance computing
- Grants
- 1 known grant, through Sep 2027
What they work on
- Develop composable providers that capture and assemble checkpoints of distributed data structures across modules and programming languages, arranging the saved data to remove redundancy and reduce size.
- Design multi-level co-optimized caching and prefetching that manages the life cycle of checkpoints on mixed storage systems while checkpoints are saved and reused concurrently.
- Build checkpointing tools for large AI models integrated with PyTorch and DeepSpeed, giving users fast, scalable checkpointing through a familiar programming interface.
- Co-design, tune and evaluate VLCC-States with industry and national research laboratory partners.
More about the research
Today, checkpoints also help programs finish faster even when nothing fails. Training an artificial intelligence (AI) model may go off course, and saving checkpoints lets researchers step back and try another path. Checkpoints also let a job pause and resume, move to another machine, replay a bug, or check that results can be reproduced. Current tools are not flexible, fast or scalable enough for all these uses.
Why it matters
Faster, more flexible checkpointing could shorten time-to-solution for scientists running weather prediction, fluid dynamics, seismic imaging and financial models, and for teams training large AI models. Being able to backtrack, fine-tune from earlier models, migrate jobs and replay errors could cut computing costs and development effort across many research communities.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Xian-He Sun
Illinois Institute of Technology, IL·Computer Science
Engineer IOWarp, a modular and scalable data management platform for AI-infused scientific workflows, using NVMe SSDs, CXL memory devices and CPU-GPU co-design to reduce data access times on HPC systems.
Remzi Arpaci-Dusseau
University of Wisconsin-Madison, WI·Computer Science
Characterize dependencies across blockchain computing, storage and network resources to build a foundation for classifying faults that cause transaction data loss.
Yong Chen
Texas Tech University, TX·Computer Science
Build and operate a renewable-energy-powered high-performance computing facility, and study the design tradeoffs and barriers of running large-scale scientific computing on variable renewable power.
The school
Rochester Institute of Technology: 9 labs on PI Pages, $78M a year on research, 65 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.