面向深度学习 Kubernetes 负载饱和调度算法设计与实现

Kubernetes was a primary technology in building deep learning cloud platform. While using load balancing strategy in the scheduling process, it would generate resource fragments, increase the task waiting time. Without considering the scoring of external expansion resources such as GPU, Kubernetes could not adapt to the business scenario of deep learning cloud platform. Aiming at the problems, a load saturation scheduling strategy was proposed to improve the Kubernetes scheduling process, which could reduce resource fragmentation and improve resource utilization. The policy also supported scoring the external expansion resources specified by users, which could adapt to the business scenario of deep learning cloud platform. The experimental results showed that the load saturation scheduling policy could reduce the task waiting time by 23. 40%, and increase the GPU utilization and the GPU memory utilization by 14. 15%, and 6. 85%, repectively.