Data centers are at the core of most traditional online services such as the Web and email, and more recent cloud computing. With an increase in application agility and customers’ stringent requirements, data centers face demanding latency requirements to satisfy the variety of applications. Introducing an insignificant delay to user-facing applications may result in huge losses for data center operators and a waste of computing resources. Networking delay is one of the major contributors to the latency issue, hence dealing with networking overheads is essential in order to satisfy data center operators’ targets.

This thesis discusses methods of handling data center delays, investigates data flow scheduling as one of the methods, introduces machine learning with a focus on deep...[ Read more ]

Data centers are at the core of most traditional online services such as the Web and email, and more recent cloud computing. With an increase in application agility and customers’ stringent requirements, data centers face demanding latency requirements to satisfy the variety of applications. Introducing an insignificant delay to user-facing applications may result in huge losses for data center operators and a waste of computing resources. Networking delay is one of the major contributors to the latency issue, hence dealing with networking overheads is essential in order to satisfy data center operators’ targets.

This thesis discusses methods of handling data center delays, investigates data flow scheduling as one of the methods, introduces machine learning with a focus on deep reinforcement learning, provides a discussion on deep learning applications in data centers, and proposes a data flow scheduling mechanism for data centers by exploiting the state-of-the-art deep reinforcement learning techniques.

The proposed flow scheduling system, AuTO, borrows contemporary ideas from deep reinforcement learning to schedule flows with an objective to minimize the average flow completion time. AuTO is distinct from other scheduling solutions as it adapts its decisions to match the current data traffic and improves with time.

Furthermore, AuTO demonstrates that deep reinforcement learning can be used to solve data center scale problems and that human heuristics-based data flow scheduling can benefit from feedback in dynamic environments.