Network Modeling, Simulation, and Management
A doctor cultures
a sore throat and then tries different antibiotics on the cultures to see
which work best. Rensselaer researchers have pioneered a similar process for
diagnosing and treating computer network problems. With support from DARPA,
the National Science Foundation, and other funding sources, they have
developed methods to run rapid simulations to identify problems and then to
apply automated traffic management techniques to solve the problems. Current
research continues to improve simulation techniques, network management, and
quality of service.
Genesis: Many Processors
Sharing the Work
Simulation has
become an accepted scientific tool for understanding the dynamics of a
complex system, and the Internet is one of the most complex artifacts
created by technology. With decentralized control and unintentional and
deliberate disruptions, its dynamics are very difficult to understand.
Attempts to simulate it are riddled with challenges. Simulations based on
the lowest network level of a single packet can account for all the affects
and dynamics of the Internet but tend to be slow. Packets change their
status in millisecond time, and their parallel simulation requires
synchronization of the simulating processors at that time scale, which kills
the parallel efficiency. (They are like workers who spend so much time
coordinating what they want to do that little time is spent working).
Simulations based on flows of packets are fast because they synchronize at
the time scale of flows, seconds, or even minutes, but cannot represent
important events happening to individual packets. A new approach to this
problem has been proposed by
Dr. Boleslaw Szymanski,
professor of computer science and founding director of the Center for
Pervasive Computing and Networking, who was elected an IEEE fellow for his
work on parallel and distributed systems. With strong support from DARPA, he
and his team developed Genesis (The GeneralNEtworkSimulation Integration
System), a novel approach to scalability and efficiency of parallel network
simulation. The system divides the Internet into domains, with a separate
processor used to simulate each domain over a given time interval. Each
domain does the packet-level simulation but synchronizes entire flows over
the time interval. When each simulation is finished, the domains exchange
information about the flows and begin a new iteration over the same time
interval. The iterations continue until all simulations converge, and then
the system moves on to a new time interval. This approach has proven
successful in many difficult tasks, such as accurately modeling the TCP
(Transmission Control Protocol), which constantly adjusts to current network
conditions.
In other work,
Dr. Szymanski is collaborating with a group that includes the University of
Genoa in Italy, the University of Girona in Spain, and the National Academic
Network (NASK) in Poland to develop SNAP, the Service Negotiation and
Adaptive Delivery Platform. With researchers in Girona and Warsaw, he is
looking at means to use dynamic pricing to negotiate Quality of Service.
Prices are set to avoid congestion and yield rational use of the system,
which also monitors the service delivered and gives discounts, if needed, to
compensate for times when service deteriorates. The goals are to deliver the
expected service despite changing environments and to maintain customer
satisfaction. Dr. Szymanski is also providing the simulation tools needed to
measure and monitor service.
ROSSNet: The More Work,
the Better
Christopher Carothers, assistant professor of computer
science, has built ROSSNet (Rensselaer’s Optimistic Simulation System), a
very fast simulation method that uses parallel processors. With AT&T
support, Dr. Carothers and
Shivkumar Kalyanaraman,
associate professor of electrical, computer, and systems engineering (ECSE),
are using real-time simulations to optimize very complex systems in which
networks using different protocols and operating at different time scales
must be able to communicate with each other through the BGP (Border Gateway
Protocol). ROSSNet provides them a system that is insensitive to changes in
the topology of the network, that uses far less memory per connection than
other systems, and that performs very rapidly. In fact, rather than slowing
down as domains and processors are added, the system speeds up. Carothers
explains that this is because the work on each processor remains constant,
but the overhead per processor decreases.
Automated Traffic
Management
Dr. Kalyanaraman is an expert in network traffic
management whose work has already won him recognition by MIT’s Technology
Review as one of the top 100 “Visionaries for the Millennium.” With support
from both government and industry, he, his colleagues, and his students are
working on a number of projects to reduce congestion, automate network
management, and improve quality of service. These include:
Good Results Fast Using the slogan “good results fast,” Dr.
Kalyanaraman has built a recursive random search algorithm that continuously
collects real-time data on network conditions and runs simulations to
discover improved settings for the millions of network parameters. His
approach is designed to choose the best simulations from the many that are
possible. His system then works very quickly, not to seek optimal settings,
which would take too long, but to discover improved settings within a
limited time frame. The improved settings can be given to network operators
or set automatically to continuously improve network performance by
adjusting parameters to changing conditions.
Overlay
Services Just as the Internet developed over telephone lines, Dr.
Kalyanaraman is working on “overlay” next-generation services that can be
created on top of the current Internet. Such services, which could be
delivered by network service providers or could be created in ad hoc
peer-to-peer organizations, would scavenge bandwidth from computers not in
use to gain bandwidth and would choose the best routes to ensure far better
quality of service. Using overlay systems, Dr. Kalyanaraman’s methods can
deliver reliable, broadband capability for such consumer-oriented services
as video conferencing and chats and sending home movies to friends and
relatives.
Going Bananas
BANANAS is an evolutionary architectural Internet framework with a new
coding scheme that gives messages far more flexibility in the route they
take. Instead of being committed to one route, messages can be sent on
alternate paths when congestion is slowing traffic, just as drivers may
choose to switch to another route when one highway is badly congested.
Wireless Traffic Modeling
With support from
DARPA,
Biplab Sikdar, ECSE assistant
professor, is working to understand traffic patterns in wireless networks.
In one project, he developed a model to characterize the arrival times of
packets in wireless networks that use the IEEE 802.11 MAC (media access
control) protocol. His model showed significant differences between traffic
patterns in wired and wireless networks. While this study assumed the nodes
in the network were stationary, he now is looking at what happens when the
various nodes become mobile. The insights he is developing can lead to
improved network performance and can provide the modeling tools for more
accurate performance evaluation.
Can We Understand the
Internet?
Given the immense
size of the Internet and the rapid changes it constantly undergoes, is it
ever possible to accurately measure its properties and totally understand
its behavior?
Petros Drineas, assistant professor
of computer science and an expert in the design and analysis of algorithms,
and
Bulent Yener, associate professor of
computer science with strong experience in quality-of-service issues, are
looking at this basic question. They are interested in the concept of
“evasiveness,” the difficulty of obtaining accurate information about
specific properties. If it is not possible to obtain 100 percent of the
accurate information about the topology of the Internet, their goal is to
understand how much accurate information is needed for various management
tasks.
Contacts: