A Heterogeneous Resource Scheduling Framework in Container Cloud

This paper innovatively proposes a heterogeneous resource scheduling framework in Container Cloud, implementing an integrated training and inference data analysis operator supermarket system. On one hand, it enables professionals to host their training and inference operators in container cloud, allowing business personnel to use them with a single click. On the other hand, it achieves flexible scheduling of underlying heterogeneous resources, whereby the system allocates different resource quotas based on task priority and data volumes. In particular, addressing the issue of GPU resource wastage during the operator execution process, this paper implements a GPU sharing method.