KubeCon + CloudNativeCon China 2025

With the rapid growth of AI applications, optimal GPU utilization is essential, particularly in GPU sharing and job scheduling. Balancing performance, flexibility, and isolation is as challenging as the “Impossible Trinity”. Technologies such as vCUDA, MPS, and MIG are promising attempts, but each has its pros and cons. Managing clusters with multiple sharing techniques adds complexity due to differing resource names and configurations.
In this talk, we will demonstrate how to combine these methods easily. Users specify the memory and core count without managing GPU types or sharing methods. Based on user preferences and GPU resources, the best node and method will be selected. Requests are automatically translated into optimal profiles, and GPUs are dynamically partitioned.
This approach streamlines GPU management, enhances utilization, and improves scheduling. By integrating Volcano and HAMi, the solution strengthens GPU pooling and scheduling, optimizing AI workload management.