⚡ Lightning Talk: Advanced GPU-Orchestrated Workflows and HPC Integrations on K8s for Distributed AI/ML at Scale
Lightningat KubeCon + CloudNativeCon China 2025 · on Tuesday, 2025-06-10· at 15:37· Level 16 | Grand Ballroom I·
summary
Shows concrete GPU scheduling wins: switching to MIG and MPS for instance- and context-level sharing, plus gang scheduling with Volcano, raises GPU utilization from 58% to 83% and cuts average queue time from >3 minutes to 47 seconds. Describes RDMA/GPU-Direct integrations and frameworks like Horovod/Ray on Kubernetes, and adds low-overhead visibility via eBPF-driven instrumentation to debug multi-node training and tune placement policies.
AI-written summary · about
related talks
09mLightning Talk: Optimize the Ray Schedule and Autoscaling on Kubernetes Cloud: A Hetero...
10mLightning Talk: K8s Issue #52757: Sharing GPUs Among Multiple Containers
12mLightning Talk: Benchmarking Your Distributed ML Training on the K8s Platform
09mLightning Talk: Bringing Multi-Tenancy to GPU-Based Workloads
07mProject Lightning Talk: From Static Slices To Elastic GPUs: Dynamic MIG With HAMI
05mLightning Talk: Sharing a GPU Among Multiple Containers
38mBeyond Kubernetes: Adapting To Spe...
07mLightning Talk: Environmentally Sustainable AI via Power-Aware Batch Scheduling