Peeking Into the GPU Black Box: Continuous Profiling on Kubernetes With eBPF - Zahari Dichev

This talk explores the challenges of monitoring GPU usage within a Kubernetes cluster, where many existing tools fail to provide transparency for containerized workloads. Zahari Dichev from Buoyant presents an end-to-end approach to GPU observability using eBPF, detailing methods for continuous profiling of workloads and their interactions with GPU devices. The session highlights techniques to trace kernel launches, detect CUDA memory leaks, identify hardware issues, and visualize workload activity across pods and nodes. By effectively bridging Kubernetes, GPUs, and eBPF, the speaker aims to provide actionable insights, transforming GPUs from a black box into a transparent component of the cloud-native stack.