Open Cloud deep-learning architecture for big data analytics and real-time applications

Machine-learning and deep-learning has come a long way and is impacting us on an indescribable scale. With the availability of big data and high performance compute nodes, training a deep neural network model is feasible, fast and accessible. With key players open sourcing whole or part of their deep-learning framework, like Google’s TensorFlow and UC Berkeley’s Caffe, deep-learning research is currently not restricted to academia or big industry. To meet the huge hardware demand caused by the deep-learning framework’s high performance, compute nodes with specialized GPUs are required. Cloud solutions are widely available for anyone to leverage to serve their needs. However, most of the services have traditional cloud based servers, which provide virtualized instances. Despite all of the tools available, a holistic, high performance cloud machine-learning package is far from reality. This thesis propose a cloud deep-learning platform that enables end-users to easily build deep-learning models on big compute nodes that work on any type of data, of any size. Furthermore, the proposed architecture is implemented on bare metal OpenStack cloud nodes with NVIDIA M40 and K80 GPUs connected through high-speed interconnects with a scalable object storage. In our experimental platform, significant speedup was achieved using different deep-learning algorithms on popular datasets.