Gaia Scheduler: A Kubernetes-Based Scheduler Framework

摘要

This paper proposed a topology-based GPU scheduling framework. The framework is based on the traditional kubernetes GPU scheduling algorithm. In existing algorithms, GPU can only be completely allocated (In other words, a GPU is the smallest unit of resource allocation), the GPU resources are not fully used and the load is uneven. In our algorithm, GPU cluster topology is restored in a GPU cluster resource access cost tree, and different GPU resource application scenarios are scheduled and dynamically adjusted based on the resource access cost tree to obtain an optimal scheduling effect. The kubernetes GPU cluster load is effectively improved. Experimental results show that performance on load balance and resource utilization are also improved in this way. GaiaGPU has been widely used in the production practice of Tencent, the application of GaiaGPU has increased the resource utilization of GPU cluster by about 10%.