Today's largest data centers contain tens of thousands of servers, and they will encompass hundreds of thousands in the very near future. These machines are designed to serve a rich mix of applications and clients with significant aggregate bandwidth requirements; distributed computing frameworks like MapReduce/Hadoop significantly stress the network interconnect, which when compounded with progressively oversubscribed topologies and inefficient multipath forwarding, can cause a major bottleneck for large computations spanning hundreds of racks. Non-uniform bandwidth among data center nodes also complicates application design and limits overall system performance. Furthermore, using the highest-end, high port-density commercial switches at the core and aggregation layers incurs tremendous cost. To overcome this limitation, this dissertation advocates three major goals: First, the horizontal, rather than vertical, expansion of data center networks, using commodity off-the-shelf switch components, and rearrangeably non-blocking topologies such as fat- trees. We show that these topologies have several advantages in overall equipment and operational cost and power compared to traditional hierarchical trees. However, the corresponding increase in the degree of multipathing makes traffic forwarding more challenging. Traditional multipath techniques like static hashing (ECMP) can waste bisection bandwidth due to local and downstream hash collisions. To overcome this inefficiency, we next describe the architecture, implementation, and evaluation of Hedera : a centralized flow scheduling system for data center networks with global knowledge of traffic patterns and link utilization. Hedera computes max-min fair flow- bandwidth demands and uses one of several online placement heuristics to find flow paths that maximize the achievable network bisection bandwidth. Finally, to enable rapid network extensibility, we describe the system architecture and implementation of NetBump: a platform for data-plane modifications "on the wire." By using low-latency kernel bypass and user-level application development, NetBump allows examining, marking, and forwarding packets at line- rate, and enables a host of active queue management disciplines and congestion control mechanisms. This allows the prototyping and adoption of innovative functionality such as DCTCP and 802.1Qau quantized congestion notification (QCN). We show that augmenting top-of-rack switches with NetBumps effectively enables bypassing the slow adoption of data center protocols by commercial switch vendors

Engineering large-scale data center applications built from thousands of commodity nodes requires both an underlying network that supports a wide variety of traffic demands, and low latency at microsecond timescales. Many ideas for adding innovative functionality to networks, especially active queue management strategies, require either modifying packets or performing alternative queuing to packets in-flight on the data plane. However, configuring packet queuing, marking, and dropping is challenging, since buffering in commercial switches and routers is not programmable. In this work, we present NetBump, a platform for experimenting with, evaluating, and deploying a wide variety of active queue management strategies to network data planes with minimal intrusiveness and at low latency. NetBump leaves existing switches and endhosts unmodified by acting as a ``bump on the wire,'' examining, marking, and forwarding packets at line rate in tens of microseconds to implement a variety of virtual active queuing disciplines and congestion control mechanisms. We describe the design of NetBump, and use it to implement several network functions and congestion control protocols including DCTCP and 802.1Qau quantized congestion notification.

Pre-2018 CSE ID: CS2012-0976