Introduction

Artificial intelligence, particularly deep learning, is an emerging computational technology in the field of medicine. Historically radiology is a clinical practice by professional physicians with visual examination on medical images. The increasing complexity in computer aided diagnosis now can provide more quantitative measurements and act as new assessment tools in radiology.

To facilitate AI projects, The Department of Imaging and Interventional Radiology has adopted a stack of high computational capacity devices including General Purpose Graphical Processing Units. In order to consolidate current and upcoming machines, a GPU cluster system has been established to support research works and provide training opportunities. The cluster has been equipped with the most modern industrial level GPUs to encourage frontier research. We anticipate that the platform can help to promote computational analysis and resource sharing techniques within the department.

Policies

The DIIR GPU Cluster adopts a fair usage policy for resource allocation according to the share portion of equipment holders. Department provides free yet limited resources in first come first out strategy. Running job limits are equally distributed among research teams. Quota outrunning users may still enjoy the resource yet a lower priority in the job queue.

Department members are welcome to contribute to the cluster in terms of management and hardware devices. The hardware provider may enjoy the volunteer hardware management services (excluding fee for necessary peripheral devices and maintenance) and will receive the highest computation priority in the respective donated devices. In return, the volunteer management team could enjoy the all-time second priority privilege in the donated devices. Non-device donating users may still use the devices if any of them are idling.

Resource consumptions including CPU, memory and GPU running hours will be recorded at per user level. Any user abusing or hacking the cluster may be suspended from the system.

Hardware Resources

The following computational resources are available in current cluster configuration:

Resources

Specification

Quantity

GPU

Tesla V100 32GB GRAM

4

CPU

Intel Xeon E5-2698v4 2.2GHz, 20 Cores

2

Memory

 

256GB

The computational capacity is planned to be expanded with Geforce and Titan grade GPU devices.

Supporting Libraries

Researchers are encouraged to install software and libraries in the cluster necessary for their work. The libraries currently supported by the system include but are not limited to the following: