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 operates within a robot’s control loop, making latency, energy consumption, and task success joint concerns for system design.
Approach. We characterized four VLA architectures on an edge GPU server and two onboard SoCs through stage-level profiling, roofline analysis, DVFS sweeps, and 43,200 closed-loop episodes.
Findings. Model architecture and hardware balance jointly determine stage bottlenecks. Across the evaluated configurations, the preferred operating point changes with latency, energy, and task-success targets, revealing trade-offs that inform runtime and accelerator design.
Problem. Kubernetes schedules pods based on requested resources but does not account for two defining properties of distributed training: a job must acquire all of its GPUs together, and different jobs achieve very different throughput from the same resources. Existing ML schedulers address these issues through training-code modifications, execution history, GPU sharing, or dynamic reallocation, making them difficult to adopt in existing clusters.
Approach. ProfilingSense lets each job reveal its own resource efficiency through a short, isolated pre-flight run at its requested GPU count. Without modifying user code, it extracts iteration throughput from standard training logs and combines measured throughput-per-resource with waiting-time-based aging. It then gang-schedules jobs in priority order, changing only when jobs run rather than how they train or how many GPUs they receive.
Findings. We profiled five deep-learning workloads on a six-GPU testbed and used the measurements to replay one week of the Microsoft Philly trace with 2,312 jobs at an 864-GPU scale. ProfilingSense reduced average job-completion time by 18.1% in the physical evaluation and by 22.1% in the trace-driven simulation relative to the default Kubernetes scheduler. Its resource-efficiency ordering further reduced average job-completion time by 4.5% and maximum waiting time by 12.7% over FIFO gang scheduling, showing that meaningful scheduling gains are possible from execution order alone, without code changes, prior history, GPU sharing, or dynamic reallocation.