On the Deployment and Characterization of CUDA Teaching Laboratories

Abstract

When teaching CUDA in laboratories, an important issue is the economic cost of GPUs, which may prevent some universities from building large enough labs to teach CUDA. In this paper we propose an efficient solution to build CUDA labs reducing the number of GPUs. It is based on the use of the rCUDA (remote CUDA) middleware, which enables programs being executed in a computer to concurrently use GPUs located in remote servers. To study the viability of our proposal, we first characterize the use of GPUs in this kind of labs with statistics taken from real users, and then present results of sharing GPUs in a real teaching lab. The experiments validate the feasibility of our proposal, showing an overhead under 5% with respect to having a GPU at each of the students’ computers. These results clearly improve alternative approaches, such as logging into remote GPU servers, which presents an overhead about 30%.