Doctoral Thesis
Eliminating Adverse Control Plane Interactions
in
    Independent Network Systems
Network system operation is typically divided into control and data
planes--- while the data plane is responsible for processing individual
messages or packets, the control plane computes the configuration of
devices and optimizes system-wide performance. Unfortunately, the control
plane of each network system or protocol layer typically operates
independently, resulting in poor interactions between control planes
across systems. I categorize such systems using four general control
plane coordination mechanisms that can overcome these problems, based on
how much information can be shared between control planes. If no
information is shared, control planes simply "react" to data plane changes
as a rudimentary form of coordination. If all information can be shared,
"transparency" in decision making can remove most poor interactions. In
many scenarios, however, only some information can be shared (e.g.,
between control planes running in different companies). Coordination in
these scenarios is more specialized; control planes with separate data
plane resources can use "priority ranking" (i.e., providing a list of
preferences for resources without needing to show how those preferences
were decided; e.g., BGP routing between ISPs), and control planes with
shared data plane resources can use "hierarchical partitioning" (i.e.,
making coarse-grained decisions globally, and fine-grained decisions
locally; e.g., internet-wide BGP + OSPF routing).
While systems utilizing control plane coordination exist today, they
have been designed in an ad hoc fashion. I propose a set of recipes that
show when it's appropriate to use different coordination mechanisms, based
on key properties (information sharing and shared resources) in varied
scenarios (layering, administrative separation, and internet-scale
systems).
I use these recipes to guide system design in a variety of contexts, as
a case study in control plane coordination. Transparency can boost
performance in layered systems, as I show with Etalon, in the context of
reconfigurable datacenters. Priority ranking can better optimize for cost
in administratively separate systems, as I show with VDX, in the context
of content brokering. Hierarchical partitioning can improve responsiveness
in internet-scale systems, as I show with VDN, in the context of live
video streaming. Through this case study we find that these coordination
mechanisms not only solve a variety of problems, but can be efficiently
implemented.
Document: Matthew K. Mukerjee. Eliminating Adverse Control Plane
Interactions in Independent Network Systems.
Slides: pdf
keynote
Selected Work
Reconfigurable Datacenter Networks: Etalon
As datacenter networking demands have increased, CMOS manufactures have
struggled to build switches with simultaneously higher bandwidth and port
count. Thus, researchers have proposed augmenting datacenters with very high
bandwidth circuit technologies (e.g., 60GHz wireless, optics) to add bandwidth
on demand. Modifying circuits, however, has a non-trivial reconfiguration delay,
leading to end-to-end challenges. In this work, we identify three key
challenges: rapid bandwidth fluctuation, poor demand estimation, and
difficult-to-schedule workloads. These challenges all derive from assumptions
made by the application and endhost stack about the network. Using cross-layer
knowledge, we solve these problems directly with: dynamic in-network queue
resizing to mask bandwidth fluctuations, proper demand estimation using endhost stack
ADUs, and rewriting application logic for easier-to-schedule demand. We evaluate
the efficacy of these solutions by building Etalon, an open-source
reconfigurable datacenter emulator for use on public testbeds, finding they can
improve circuit utilization by 2x, flow-completion time by 8x, and
HDFS write times by 9x.
Publication (NSDI '20):
Matthew K. Mukerjee, Christopher Canel,
Weiyang Wang, Daehyeok Kim, Srinivasan Seshan, Alex C. Snoeren.
Adapting TCP for Reconfigurable Datacenter Networks.
Slides: pdf
keynote
Etalon source code: github
Mitigating the Impact of Brokers on Content Delivery: VDX
Various trends are reshaping Internet video delivery: exponential growth in
video traffic, increasing expectations of high video quality of experience
(QoE), and the proliferation of varied CDN deployments (e.g., cloud
computing-based, content provider-owned datacenters, and ISP-owned CDNs). More
fundamentally, content providers are shifting delivery from a single CDN to
multiple CDNs, through the use of a content broker. In our work, we show brokers
invalidate many traditional delivery assumptions by not communicating their
decisions with CDNs (e.g., shifting traffic between CDNs mid-stream invalidates
short- and long-term CDN traffic prediction). Using data from both a CDN and a
broker, we provide the first analysis of these issues, and design a proper
marketplace-like interface (VDX) that provides better client performance and
fairly pays CDNs based on their per-cluster delivery costs.
Publication (CoNEXT '17):
Matthew K. Mukerjee, Ilker Nadi Bozkurt,
Devdeep Ray, Bruce Maggs, Srinivasan Seshan, Hui Zhang.
Redesigning CDN-Broker Interactions for Improved Content Delivery.
Best Paper Award
Publication (HotNets '16):
Matthew K. Mukerjee, Ilker Nadi Bozkurt, Bruce Maggs,
Srinivasan Seshan, Hui Zhang.
The Impact of Brokers on the Future of Content Delivery.
Slides (CoNEXT): pdf
keynote
Slides (HotNets): pdf
keynote
Practical Centralized Control for Live
Video Delivery: VDN
Live video delivery is difficult to control due to failures, flash
crowds, and under-provisioning at Internet-scale. Traditional
caching-based solutions are not helpful due to the live
aspect. Pure centralization in the wide-area is not practical due to the
simultaneous need of high availability, low latency, and highly
optimized quality. We design a system, VDN, that combines the quality
benefits of centralized optimization with the high availability and low
latency of distributed control, which we dub hybrid control.
Using large-scale simulation and a wide-area testbed, we show that VDN
can offer ~2x improvement in quality over today's DNS-based systems and
~100ms join times, while providing CDN operators expressive policy
management.
Publication (SIGCOMM '15):
Matthew K. Mukerjee, David Naylor, Junchen Jiang, Dongsu Han, Srinivasan Seshan,
Hui Zhang. Practical, Real-time Centralized Control for CDN-based Live Video Delivery.
Slides: pdf
keynote
Mobile Brain-Computer Interfaces: NeuroPhone
How can we incorporate commodity (toy / < $300)
electroencephalography (EEG) headsets into meaningful mobile
applications? We build NeuroPhone, a
brain-powered address book, to provide a cursory glance into
this seemingly Sci-Fi future. A phone presents pictures of contacts, and
the user is told to anticipate the picture of the person they wish to
call. When the picture appears, their brain elicits a P300
response recognized by the EEG headset, initiating the call.
Publication (SIGCOMM 2010 workshop MobiHeld):
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.
For my full publication list, see my
CV.