High Throughput Data Center Topology Design
(To appear in) NSDI '14 with P. Brighten Godfrey, Alexandra Kolla
[ Tech Report]
We prove the near-optimality of random graphs for the problem of high throughput topology design with homogeneous network equipment. We then use random graphs as building blocks for a systematic exploration of the design space for heterogeneous networks with switches with different port-counts and line-speeds. Our design methods yield a 40% throughput improvement on a real-world deployed data center topology.
Practical DCB for Improved Data Center Networks
(To appear in) INFOCOM '14 with Brent Stephens, Alan Cox, John Carter, Colin Dixon, Wesley Felter
Based on the use of lossless Ethernet (DCB), we propose a new TCP variant, TCP-Bolt, which reduces flow completion times by as much as 70%. This paper details the difficulties posed by a lossless network -- large buffering delays, unfairness, head of line blocking, and routing deadlock -- and proposes simple, practical techniques to overcome them.
Ensuring Connectivity via Data Plane Mechanisms
NSDI '13 with
Junda Liu, Aurojit Panda, P. Brighten Godfrey, Michael Schapira, Scott Shenker
We present a novel mechanism that routes around failures at data plane speeds. Our scheme (provably) ensures route-connectivity as long as the underlying physical topology is connected.
On the Resilience of Routing Tables
We motivate the study of the resilience of static routing tables which incorporate failover paths, and present the first positive and negative results on the resilience achievable with such tables.
Jellyfish: Networking Data Centers Randomly
A random graph based topology with higher capacity and resilience, shorter path lengths, lower costs, and greater operational flexibility (construction at arbitrary sizes, and easier incremental expansion) than traditional topologies.
Proteus: A Topology Malleable Network
A data center network designed to adjust the topology to the traffic demands. Based on use of optical technology, such a network provides freedom from the headaches of "clever" workload placement, and traffic engineering.
Information-Centric Networking: Seeing the Forest for the Trees
HotNets '11 with Ali
Ghodsi, Teemu Koponen, Barath Raghavan, Scott Shenker, James Wilcox
[ PDF ]
A skeptical overview of information (or content) centric network proposals. In particular, we raise questions about the utility of ubiquitous caching espoused by many of the proposals, and the extent of differentiation among them.
Intelligent Design Enables Architectural Evolution
HotNets '11 with Ali Ghodsi,
Teemu Koponen, Barath Raghavan, Scott Shenker, James Wilcox
[ PDF ]
We claim that only a few "intelligent" design changes are needed in the Internet's architecture to support evolvability, and they follow directly from the well known principles of indirection, modularity, and extensibility.
DISCO: Scalable Routing on Flat Names
A routing protocol that guarantees delivery of packets within close to the lowest possible latency, given only an arbitrary location-independent (DNS-like) name. We build upon compact routing theory, closing the heretofore open problem of achieving the latency and state trade-offs in the literature with a distributed protocol.
Verifiable Network-Performance Measurements
CoNext '10 with Katerina
Argyraki, Petros Maniatis
[ PDF ]
An inter-domain system to detect and assess Service Level Agreement (SLA) violations by contractual partners. Our system ensures that domains cannot abuse it to significantly exaggerate their performance.