Network flow optimization for distributed clouds

Internet applications, which rely on large-scale networked environments such as data centers for their back-end support, are often geo-distributed and typically have stringent performance constraints. The interconnecting networks, within and across data centers, are critical in determining these applications' performance. Data centers can be viewed as composed of three layers: physical infrastructure consisting of servers, switches, and links, control platforms that manage the underlying resources, and applications that run on the infrastructure. This dissertation shows that network flow optimization can improve performance of distributed applications in the cloud by designing high-throughput schemes spanning all three layers. At the physical infrastructure layer, we devise a framework for measuring and understanding throughput of network topologies. We develop a heuristic for estimating the worst-case performance of any topology and propose a systematic methodology for comparing performance of networks built with different equipment. At the control layer, we put forward a source-routed data center fabric which can achieve near-optimal throughput performance by leveraging a large number of available paths while using limited memory in switches. At the application layer, we show that current Application Network Interfaces (ANIs), abstractions that translate an application's performance goals to actionable network objectives, fail to capture the requirements of many emerging applications. We put forward a novel ANI that can capture application intent more effectively and quantify performance gains achievable with it. We also tackle resource optimization in the inter-data center context of cellular providers. In this emerging environment, a large amount of resources are geographically fragmented across thousands of micro data centers, each with a limited share of resources, necessitating cross-application optimization to satisfy diverse performance requirements and improve network and server utilization. Our solution, Patronus, employs hierarchical optimization for handling multiple performance requirements and temporally partitioned scheduling for scalability