XIA, Carnegie Mellon University
I work under the umbrella of the eXpressive
Internet Architecture (XIA) project, a large-scale
research project funded by the National Science Foundation (NSF) to effectively design, implement, and
evaluate a clean-slate redesign of core internet functionality. XIA seeks
to mainly improve the evolvability of the network by providing a simple framework to
allow deployment of future means of communication, in addition to providing incredibly
flexible routing and intrinsic security.
Next-Generation Optical Circuit-switched Datacenters, CMU
Advisor: Srini Seshan, Dave Andersen,
Michael Kaminsky, Alex Snoeren, and George Porter
A range of new datacenter switch designs combine wireless or optical circuit
technologies with electrical packet switching to deliver higher performance at
lower cost than traditional packet-switched networks. These hybrid networks
propose scheduling large traffic demands via a high-rate circuit-switched fabric
and handle any remaining traffic with a lower-rate, traditional packet-switched
interconnect. Unfortunately, the circuit switch is technologically constrained
by non-trivial reconfiguration time, requiring clever tricks to achieve
high utilization of the hybrid network.
We've investigated multiple problems within the space of hybrid circuit/packet
networks, including: 1) techniques for building efficient direct schedulers, 2)
sending traffic indirectly through other nodes to reduce the number of
necessary configurations, 3) rewriting applications to make better use of the
underlying network through application-layer bursts.
He Liu, Matthew K. Mukerjee, Conglong Li, Nicolas Feltman, George Papen,
Stefan Savage, Srinivasan Seshan, Geoffref M. Voelker, David G. Andersen,
Michael Kaminsky, George Porter, and Alex C. Snoeren.
Scheduling Techniques for Hybrid Circuit/Packet Networks. CoNEXT '15.
Global Video Control Plane, Carnegie Mellon University
Srini Seshan, Bruce Maggs, and Hui Zhang
Video is the use-case for the Internet, expected to comprise 80-90% of
all traffic by 2019. Live video is also rapidly overtaking the web, with
delivery records in the 7+ Tbps range (Akamai world cup delivery) and estimates
of 50 Tbps by this year. Despite video delivery already being a complex
optimization problem, it's further complicated by the different entities in the
larger video ecosystem: e.g., content brokers (the Conviva’s of the world),
CDNs, ISPs, content providers, etc., each optimizing for video delivery
independently. There is no standard interface between them, leading each entity
to act in their own best interests based on limited information about other
entities and the network. For example, occasionally CDNs make incorrect
assumptions about how ISPs will react to large volumes of traffic, hurting user
performance. We wish to have each entity share a minimal amount of data that
allows them to continue to act in their own interest, but lower their own cost
and increases end-to-end performance by understanding other entities’ decisions
Video Delivery Network (VDN), Carnegie Mellon University
Srini Seshan, Dongsu Han, and Hui Zhang
Live video delivery is difficult to control due to failures, flash crowds, and
under-provisioning at Internet-scale. Users want very high quality and
synchronized low latency (as many users interact with each other in real-time
via social media or incorporated stream chat functionality), but traditional
solutions (i.e., caching) are not helpful due to the live aspect. The
obvious solution of centralization and replication is not practical---the
combined need of high availability, low latency, and highly optimized quality is
difficult to impossible to simultaneously obtain in a purely centralized
Internet-scale solution. Thus, current systems use a DNS-based distributed
We design a system that combines the quality benefits of a centralized control
plane (using centralized optimization) with the low latency and resilience to
failures of a distributed control plane, which we dub a hybrid control
plane. Through careful design and integration of both approaches, we show our
system benefits from both, while avoiding problems from competing decisions. We
evaluate our system using client-side traces from a large-scale video analytics
company in a variety of scenarios to measure quality improvement, as well as on
EC2 to show low latency (join times) and resilience to failures. We find a
~2x improvement in quality with ~100ms join times despite WAN
conditions, in addition to providing flexibility to CDN operators in terms of
quality, policy expression, and cost.
Matthew K. Mukerjee, David Naylor, Junchen Jiang, Dongsu Han, Srinivasan Seshan,
Hui Zhang. Practical, Real-time Centralized Control for CDN-based Live Video Delivery. SIGCOMM '15.
Understanding Incremental Deployment, Carnegie Mellon University
Srini Seshan and Peter Steenkiste
The focus of this work is building a fundamental understanding of incremental
deployability of new network architectures. Although IPv6 has been around for decades,
ease of deployment over IPv4 is still a major concern. My work focused on distilling down
four key problems that network architectures need to solve in order to be deployable. We
further examine a variety of specific mechanisms that solve these problems, creating a design space of
options. We evaluate a select few of these options (modeling current IPv6 deployment
techniques as well as XIA) across the US using PlanetLab as a testbed. We find that certain
mechanisms prominently featured in XIA (multiple discrete identifiers and fallbacks in
forwarding) as well as control plane centralization can help aid in virtually seamless
deployment of new network architectures.
Matthew K. Mukerjee, Dongsu Han, Srinivasan Seshan, and Peter Steenkiste.
Understanding tradeoffs in incremental deployment of new network architectures. CoNEXT '13.
BiFocals, Cornell University
Daniel Freedman and Ken Birman
I worked with members of the BiFocals project to research into the cause and effect of
high-speed 10 GbE fiber-optic wide-area network burstiness. The burstiness in question
happens at a timescale so small (order of microseconds) that conventional computer
science techniques (userland software, kernel time-stamping, NIC level time-stamping,
etc.) are entirely unable to see these bursts. We thus used high-precision physics
equipment to measure the packets in-flight on the actual fiber. These bursts provide
an instantaneous data rate of 10 Gbps, potentially overwhelming commodity endpoint
servers. We show through experimentation that various common endpoint configurations can
provide radically different loss for the same bursty stream.
NeuroPhone, Dartmouth College
Andrew Campbell, Tanzeem Choudhury, Rajeev Raizada
This work focused on probing the intersections of neuroscience and
mobile. Given the existence of commodity ("toy") electroencephalography (EEG) headsets ($<300) what are possible
applications in the space of mobile? Our system (NeuroPhone) provides a cursory glance
at a possible applications in that space, a "brain-powered" address book. The mobile
phone presents pictures of contacts on the display and the EEG headset recognizes which
contact the user wished to call, due to the contact they expect to appear eliciting a specific
brain response ("P300"). This work eventually lead to an NSF EAGER grant.
Campbell, A. T., T. Choudhury, S. Hu, H. Lu, M. K. Mukerjee, M. Rabbi,
R. D. S Raizada. NeuroPhone: Brain-Mobile Phone Interface using a Wireless
EEG Headset. SIGCOMM 2010 - MobiHeld 2010, August 2010.