Efficient Access to Shared GPU Resources: Mechanisms and Use Cases

This talk explores efficient ways to share GPU resources in Kubernetes, addressing the challenges of limited and expensive GPUs. Speakers Diogo Guerra and Diana Gaponcic from CERN discuss mechanisms like time-sharing and Nvidia's Multi-Instance GPU (MIG), presenting benchmark results to guide workload assignment. They cover how to manage GPUs centrally for optimal utilization in applications such as continuous integration, machine learning, and batch processing.