The growing complexity of communication networks and the explosion of network traffic have made the task of managing these networks exceedingly hard. A potential approach for striking this increasing complexity is to build an autonomous self-driving network that can measure, analyze and control itself in real time and in an automated fashion with- out direct human intervention. In this thesis, we focus on realizing such an autonomous network leveraging state-of-the-art networking technologies along with artificial intelli- gence and machine learning techniques. Toward this goal, we exploit different learning paradigms to automate network management. First, we propose supervised machine learning methods to detect increases in delays in mobile broadband networks. Further, considering the challenges of supervised learning in networking applications, we present a novel real-time distributed architecture for detecting anomalies in mobile network data in an unsupervised fashion. It also involves a collaborative framework for knowledge sharing between the distributed probes in the network to improve the overall system accuracy. Second, we propose a novel deep reinforcement learning based control framework for op- timizing resources utilization while minimizing performance degradation in multi-slice Radio Access Network (RAN) through a set of diverse control actions. We explore both centralized and distributed control architectures. Last, we design a framework for timely collecting telemetry, detecting and attributing outages in mobile networks. We evaluate our framework on a software defined virtualised testbed that resembles a cloudified mobile network.

Processor architectures have been evolving quickly since the introduction of the central processing unit. For a very long time, one of the important means of increasing performance was to increase the clock frequency. However, in the last decade, processor manufacturers have hit the so-called power wall, with high heat dissipation. To overcome this problem, processors were designed with reduced clock frequencies but with multiple cores and, later, heterogeneous processing elements. This shift introduced a new challenge for programmers: Legacy applications, written without parallelization in mind, gain no benefits from moving to multicore and heterogeneous architectures. Another challenge for the programmers is that heterogeneous architecture designs are very different with respect to caches, memory types, execution unit organization, and so forth and, in the worst case, a programmer must completely rewrite the application to obtain the best performance on the new architecture.

Multimedia workloads, such as video encoding, are often time sensitive and interactive. These workloads differ from traditional batch processing workloads with no real-time requirements. This work investigates how to 
use modern heterogeneous architectures efficiently to process multimedia workloads. To do so, we investigate both simple and complex workloads on multiple architectures to learn about the properties of these architectures. When programing multimedia workloads, it is very important to know how the algorithms perform on the target architecture. In addition, achieving high performance on heterogeneous architectures is not a trivial task, often requiring detailed knowledge about the architecture. We therefore evaluate several optimizations so we can learn how best to write programs  for these architectures and avoid potential pitfalls. 
We later use the knowledge gained to propose a framework design and language called Parallel Processing Graph (P2G). The P2G framework is designed for multimedia workloads and supports heterogeneous architectures. To demonstrate the feasibility of the framework, we construct a proof-of-concept implementation. Two simple workloads show that we can express multimedia workloads in the system. We also demonstrate the scalability of the designed solution.