A hybrid GPU cluster and volunteer computing platform for scalable deep learning

Deep learning is a very computing-intensive and time-consuming task. It needs an amount of computing resource much greater than a single machine can afford to train a sophisticated model within a reasonable time. Normally, GPU clusters are required to reduce the training time of a deep learning model from days to hours. However, building large dedicated GPU clusters is not always feasible or even ineffective for most organizations due to the cost of purchasing, operation and maintenance while such systems are not fully utilized all the time. In this regard, volunteer computing can address this problem as it provides additional computing resources at less or no cost. This work presents the hybrid cluster and volunteer computing platform that scales out GPU clusters into volunteer computing for distributed deep learning. The owners of the machines contribute unused computing resources on their computers to extend the capability of the GPU cluster. The challenge is to seamlessly align the differences between GPU cluster and volunteer computing systems so as to ensure the scalability transparency, whereas performance is also another major concern. We validate the proposed work with two well-known sample cases. The results show an efficient use of our hybrid platform at sub-linear speedup.