The emergence of the mega data center has resulted in the basic building block of ever larger data centers changing from a rack comprising tens of servers to a self-contained modular shipping container that holds upto a thousand servers. These self-contained modular blocks include networking, power and cooling equipment besides servers. However, provisioning bandwidth between these containers at a large scale is still a significant challenge. Traditional approaches to provisioning bandwidth use electrical packet switches with a scale-up architecture and are often highly oversubscribed. More recent proposals such as those using clos networks promise full bisection bandwidth between servers, albeit at a high cost and power consumption. We present Helios, a hybrid architecture for modular data centers that combines electrical packet switching and optical circuit switching in a single network and dynamically provisions bandwidth between the modular containers on demand. We investigate this design from an architectural standpoint by building a fully functional prototype and explore its implications for data center networks and the challenges that it introduces. Our prototype shows the feasibility of building such a system and achieving high performance at considerably lower cost. Additionally it uncovers several issues that pose new challenges for designing large data center networks and problems that arise when circuits are rapidly reconfigured

Modern web services are complex systems with several components that impose stringent performance requirements on the network. The networking subsystem in turn consists of several pieces, such as the wide area and data center networks, different devices, and protocols involved in a user's interaction with a web service. In this dissertation we take a holistic view of the network and improve efficiency and functionality across the stack. We identify three important networking challenges faced by web services in the wide area network, the data center network, and the host network stack, and present solutions. First, web services are dominated by short TCP flows that terminate in as few as 2-3 round trips. Thus, an additional round trip for TCP's connection handshake adds a significant latency overhead. We present TCP Fast Open, a transport protocol enhancement, that enables safe data exchange during TCP's initial handshake, thereby reducing application network latency by a full round trip time. TCP Fast Open uses a security token to verify client IP address ownership, and mitigates the security considerations that arise from allowing data exchange during the handshake. TCP Fast Open is widely deployed and is available as part of the Linux Kernel since version 3.6. Second, provisioning network bandwidth for hundreds of thousands of servers in the data center is expensive. Traditional shortest path based routing protocols are unable to effectively utilize the underlying topology's capacity to maximize network utilization. We present Dahu, a commodity switch design targeted at data centers, that avoids congestion hot-spots by dynamically spreading traffic uniformly across links, and actively leverages non -shortest paths for traffic forwarding. Third, scalable rate limiting is an important primitive for managing server network resources in the data center. Unfortunately, software-based rate limiting suffers from limited accuracy and high CPU overhead at high link speeds, whereas current NICs only support few tens of hardware rate limiters. We present SENIC, a NIC design that natively supports tens of thousands of rate limiters -- 100x to 1000x the number available in NICs today -- to meet the needs of network performance isolation and congestion control in data centers

Solving "Big Data" problems requires bridging massive quantities of compute, memory, and storage, which requires significant amounts of bisection bandwidth. Topologies like Fat-tree, VL2, and HyperX achieve a scale-out design by leveraging multiple paths from source to destination. However, traditional routing protocols are not able to effectively utilize these links while also harnessing spare network capacity to better statistically multiplex network resources. In this work we present Dahu, a switch mechanism for efficiently load balancing traffic in multipath networks. Dahu avoids congestion hot-spots by dynamically spreading traffic uniformly across links, and forwarding traffic over non-minimal paths where possible. By performing load balancing primarily using local information, Dahu can act more quickly than centralized approaches, and responds to failure gracefully. Our evaluation shows that Dahu delivers up to 50% higher throughput relative to ECMP in an 8,192 server Fat-tree network and up to 500% improvement in throughput in large scale HyperX networks with over 130,000 servers.

Pre-2018 CSE ID: CS2013-0992

Data centers lower costs by sharing the physical infrastructure among multiple services. However, the data center network should also ideally provide bandwidth guarantees to each service in a tunable manner while maintaining high utilization. We describe {\em NetShare}, a new statistical multiplexing mechanism for Data Center networks that does this without requiring changes to existing routers. NetShare allows the bisection bandwidth of the network to be allocated across services based on simple weights specified by a manager. Bandwidth unused by a service is shared proportionately by other services. More precisely, NetShare provides weighted hierarchical max-min fair sharing, a generalization of hierarchical fair queuing of individual {\em links}. We present three mechanisms to implement NetShare including one that leverages TCP flows and requires no changes to routers or servers. We show experiments using multiple Hadoop instances and a network of Fulcrum switches and show that the instances can interfere without NetShare and yet complete faster with NetShare when compared to the alternative of static reservation.

Pre-2018 CSE ID: CS2010-0957