Optimization of data-intensive computing networks

• Online Access: http://hdl.handle.net/2047/D20398270

The importance of real-time computation on large quantities of data has recently grown substantially, as we have developed the infrastructure and the tools for storing and gathering data. Computing platforms and distributed systems provide three fundamental data-related functionalities: (a) modifying data (i.e. computation), (b) transferring data (i.e. communication between processors), and (c) storing data (i.e. caching reusable data). In this thesis, we address several important questions on joint optimization of computation, communication, caching, and admission control in data-intensive systems, such as IoT-enabled health-care, machine learning applications at the edge, and industrial automation. As explained thoroughly during this thesis, the problem of in-network caching and computation is combinatorics in nature. A major bulk of research papers and studies in this area are heuristic methods with the goal of enhancing network performance. The importance of this thesis is that we provide solutions, algorithms, and platforms with performance guarantees, and we stay away from heuristics. Specifically, we propose a framework for joint computation scheduling, caching, and request forwarding within such decentralized computing environments. We characterize the stability region of a ``genie-aided'' computing network where data required by computation are instantly accessible, and develop a throughput optimal control policy for this model. Based on this, we develop a practically implementable distributed and adaptive algorithm, and show that it exhibits superior performance in terms of average task completion time when compared to several popular policies. In a separate line of work, we formulate the problem of optimal admission control in cache networks, whereby both admissible input rates and content placements are optimized jointly. Given the demand rate of content requests and the path which requests follow, our goal is to determine a jointly optimal content placement and input rate allocation which makes the network stable, and maximizes the aggregate utility of requests. We formulate this as a maximization problem with non-convex constraints, and propose solving this problem via (a) a Lagrangian barrier algorithm and (b) a convex relaxation. We prove different optimality guarantees for each of these two algorithms; our proofs exploit the fact that the non-convex constraints of our problem involve DR-submodular functions. This is a big deviation from existing works in DR-submodular maximization, as we have multiple such functions in our constraint set. In our final piece of work, we study a data-intensive computing network where there are costs (delay, money, etc.) associated with fetching the results and data objects from upstream links. We formulate the problem of joint optimal computation scheduling and caching to minimize the average cost of network operation. Although the problem is NP-complete, we show that it can be solved in different regimes efficiently with optimality guarantees. Specifically, in a regime with low CPU power and higher result cost (compared to data cost), we show that a surrogate problem of DR-submodular maximization subject to a down-closed convex set can be solved instead of the original problem, and we provide optimality guarantees on the value obtained by solving the surrogate problem.