Migration Aware Scheduling in GPU as a Service Systems

Systems with GPU instead of CPU can be used for parallel jobs such as deep learning jobs training. In cloud computing data centers, there are servers with GPU instead of CPU. Since the time unit price of GPU servers is ۵ or ۸ x higher than CPU servers. Optimal using of GPU servers is a challenging problem. This paper proposes a migration aware scheduling model for GPU servers that are running parallel jobs with the aim to increase acceptance rate and reduce latency. The evaluation results show that the proposed method works better in migrating jobs than other basic models such as FIFO and without migration models

کلیدواژه ها:

GPU as a service ، scheduling ، parallel job ، migration.