M.S. Student, School of Computing, KAIST · CASYS Lab
I am a third-semester M.S. student in the School of Computing at KAIST, advised by Jaehyuk Huh at the Computer Architecture and Systems Lab (CASYS). I obtained my B.S. in Computer Science and Engineering at Sungkyunkwan University, where I was fortunate to work with Euiseong Seo on GPU resource scheduling for distributed deep learning on Kubernetes-based cloud platforms.
Problem. VLA inference runs inside a robot’s control loop, coupling latency and energy consumption with task success.
Approach. Characterized four VLA architectures across an edge GPU server and two onboard SoCs using stage-level profiling, roofline analysis, DVFS sweeps, and 43,200 closed-loop episodes.
Findings. Showed that model architecture and hardware balance shift stage bottlenecks and change the preferred operating point across latency, energy, and task-success objectives.
Problem. Distributed training jobs must acquire all requested GPUs together, but Kubernetes accounts for neither gang scheduling nor differences in job-level resource efficiency.
Approach. ProfilingSense profiles each job in an isolated pool, extracts iteration throughput from standard training logs, and ranks jobs by measured throughput per resource plus wait-time aging. It changes only job execution order, without modifying training code, sharing GPUs, or reallocating resources.
Findings. Using profiles of five workloads measured on a 6-GPU testbed, we replayed 2,312 Microsoft Philly jobs at 864-GPU scale. Compared with FIFO-Gang scheduling, ProfilingSense reduced peak queue length by 8.2% and P99 JCT by 5.8%.