Tsinghua University's Storage Research Group from the Department of Computer Science and Technology has been honored with the Distinguished Artifact Award at the 24th USENIX Conference on File and Storage Technologies (FAST 2026). The accolade recognizes their pioneering paper, "GPU Checkpoint/Restore Made Fast and Lightweight," which introduces a novel scheme to enhance GPU efficiency in modern computing environments.
The award-winning research, whose first author is Shaoxun Zeng, PhD student (2021 cohort), and supervised by Associate Professor Youyou Lu, addresses critical challenges in GPU management. Their proposed solution, named GCR (GPU Checkpoint/Restore), separates data and control paths to speed up saving and restoring GPU states, and enables efficient incremental checkpointing. This innovation supports key applications in high-demand GPU workloads, such as rapid scaling of elastic tasks, multi-task switching, and fault-tolerant computing. By boosting overall GPU cluster utilization, GCR holds substantial promise for data centers and AI-driven systems. Compatible with popular frameworks like vLLM, DeepSpeed, and Transformers, as well as various GPU models, the scheme minimizes average performance interference to under 1% while slashing checkpoint save times to at most 28% of traditional methods and restore times to at most 13%. The Distinguished Artifact Award highlights the paper's exceptional open-source implementation, praised for its completeness, reproducibility, and high standards in functionality and performance.
The Distinguished Artifact Award certificate
Tsinghua's Storage Research Group has a longstanding reputation for advancing storage technologies through innovative research and international accolades. In recent years, the lab has emphasized open-source contributions, fostering a vibrant ecosystem in the storage community.
Held annually, FAST is a premier academic forum in file and storage systems, established over two decades ago and classified as a Class-A conference by the China Computer Federation (CCF).
Editor: Li Han
