A Container Cloud Based PaaS Solution for Supplying GPU Computing Resources

With the increasing number of deep learning models started to rely on the high computing throughput of GPUs, many institutes, such as universities, have invested a lot in purchasing GPU devices and servers and building data centers. However, providing usable development environments for end-users can be quite a challenge for them due to the lack of specific expertise. In this paper, we designed and implemented a platform as a service (PaaS) solution, which enables end-users to provision an integrated multi-framework development environment with shared GPU computing resources using a few clicks. Our actual use shows that the proposed platform is capable of supporting nearly 100 development environments through a Kubernetes cluster consisting of 3 nodes and each node equipped with 2 Nvidia Tesla V100 datacenter-class GPUs with 32GB device memory, and can be scaled up and down as needed.