Caltech Computer Scientists Develop FAST Protocol to Speed up Internet
March 18, 2003
Caltech computer scientists have developed a new data transfer
protocol for the Internet fast enough to download a full-length DVD
movie in less than five seconds.
Caltech computer scientists have developed a new data transfer
protocol for the Internet fast enough to download a full-length DVD
movie in less than five seconds.
The protocol is called FAST, standing for Fast Active queue management
Scalable Transmission Control Protocol (TCP). The researchers have
achieved a speed of 8,609 megabits per second (Mbps) by using 10
simultaneous flows of data over routed paths, the largest aggregate
throughput ever accomplished in such a configuration. More
importantly, the FAST protocol sustained this speed using standard
packet size, stably over an extended period on shared networks in the
presence of background traffic, making it adaptable for deployment on
the world's high-speed production networks.
The experiment was performed last November during the Supercomputing
Conference in Baltimore, by a team from Caltech and the Stanford
Linear Accelerator Center (SLAC), working in partnership with the
European Organization for Nuclear Research (CERN), and the
organizations DataTAG, StarLight, TeraGrid, Cisco, and Level(3).
The FAST protocol was developed in Caltech's Networking Lab, led by
Steven Low, associate professor of computer science and electrical
engineering. It is based on theoretical work done in collaboration
with John Doyle, a professor of control and dynamical systems,
electrical engineering, and bioengineering at Caltech, and Fernando
Paganini, associate professor of electrical engineering at UCLA. It
builds on work from a growing community of theoreticians interested in
building a theoretical foundation of the Internet, an effort in which
Caltech has been playing a leading role.
Harvey Newman, a professor of physics at Caltech, said the fast
protocol "represents a milestone for science, for grid systems, and
for the Internet."
"Rapid and reliable data transport, at speeds of one to 10 Gbps and
100 Gbps in the future, is a key enabler of the global collaborations
in physics and other fields," Newman said. "The ability to extract,
transport, analyze and share many Terabyte-scale data collections is
at the heart of the process of search and discovery for new scientific
knowledge. The FAST results show that the high degree of transparency
and performance of networks, assumed implicitly by Grid systems, can
be achieved in practice. In a broader context, the fact that 10 Gbps
wavelengths can be used efficiently to transport data at maximum speed
end to end will transform the future concepts of the Internet."
Les Cottrell of SLAC, added that progress in speeding up data
transfers over long distance are critical to progress in various
scientific endeavors. "These include sciences such as high-energy
physics and nuclear physics, astronomy, global weather predictions,
biology, seismology, and fusion; and industries such as aerospace,
medicine, and media distribution.
"Today, these activities often are forced to share their data using
literally truck or plane loads of data," Cottrell said. "Utilizing the
network can dramatically reduce the delays and automate today's labor
intensive procedures."
The ability to demonstrate efficient high performance throughput using
commercial off the shelf hardware and applications, standard Internet
packet sizes supported throughput today's networks, and requiring
modifications to the ubiquitous TCP protocol only at the data sender,
is an important achievement.
With Internet speeds doubling roughly annually, we can expect the
performances demonstrated by this collaboration to become commonly
available in the next few years, so the demonstration is important to
set expectations, for planning, and to indicate how to utilize such
speeds.
The testbed used in the Caltech/SLAC experiment was the culmination of
a multi-year effort, led by Caltech physicist Harvey Newman's group on
behalf of the international high energy and nuclear physics (HENP)
community, together with CERN, SLAC, Caltech Center for Advanced
Computing Research (CACR), and other organizations. It illustrates the
difficulty, ingenuity and importance of organizing and implementing
leading edge global experiments. HENP is one of the principal drivers
and co-developers of global research networks. One unique aspect of
the HENP testbed is the close coupling between R&D and production,
where the protocols and methods implemented in each R&D cycle are
targeted, after a relatively short time delay, for widespread
deployment across production networks to meet the demanding needs of
data intensive science.
The congestion control algorithm of the current Internet was designed
in 1988 when the Internet could barely carry a single uncompressed
voice call. The problem today is that this algorithm cannot scale to
anticipated future needs, when the networks will be compelled to carry
millions of uncompressed voice calls on a single path or support major
science experiments that require the on-demand rapid transport of
gigabyte to terabyte data sets drawn from multi-petabyte data stores.
This protocol problem has prompted several interim remedies, such as
using nonstandard packet sizes or aggressive algorithms that can
monopolize network resources to the detriment of other users. Despite
years of effort, these measures have proved to be ineffective or
difficult to deploy.
They are, however, critical steps in our evolution toward ultrascale
networks. Sustaining high performance on a global network is extremely
challenging and requires concerted advances in both hardware and
protocols. Experiments that achieve high throughput either in isolated
environments or using interim remedies that by-pass protocol
instability, idealized or fragile as they may be, push the state of
the art in hardware and demonstrates its performance limit.
Development of robust and practical protocols will then allow us to
make effective use of the most advanced hardware to achieve ideal
performance in realistic environments.
The FAST team addresses the protocol issues head-on to develop a
variant of TCP that can scale to a multi-gigabit-per-second regime in
practical network conditions. The integrated approach that combines
theory, implementation, and experiment is what makes their research
unique and fundamental progress possible.
Using standard packet size that is supported throughout today's
networks, the current TCP typically achieves an average throughput of
266 Mbps, averaged over an hour, with a single TCP/IP flow between
Sunnyvale near SLAC and CERN in Geneva, over a distance of 10,037
kilometers. This represents an efficiency of just 27 percent. The FAST
TCP sustained an average throughput of 925 Mbps and an efficiency of
95 percent, a 3.5-times improvement, under the same experimental
condition. With 10 concurrent TCP/IP flows, FAST achieved an
unprecedented speed of 8,609 Mbps, at 88 percent efficiency, that is
153,000 times that of today's modem and close to 6,000 times that of
the common standard for ADSL (Asymmetric Digital Subscriber Line)
connections.
The 10-flow experiment sets another first in addition to the highest
aggregate speed over routed paths. It is the combination of high
capacity and large distance that causes performance problems.
Different TCP algorithms can be compared using the product of achieved
throughput and the distance of transfer, measured in
bit-meter-per-second, or bmps. The world record for the current TCP is
10 peta (1 followed by 16 zeros) bmps, using a nonstandard packet
size. The Caltech/SLAC experiment transferred 21 terabytes over six
hours between Baltimore and Sunnyvale using standard packet size,
achieving 34 peta bmps. Moreover, data was transferred over shared
research networks in the presence of background traffic, suggesting
that FAST can be backward compatible with the current protocol. The
FAST team has started to work with various groups around the world to
explore testing and deploying FAST TCP in communities that need
multi-Gbps networking urgently.
The demonstrations used a 10 Gbps link donated by Level(3) between
StarLight (Chicago) and Sunnyvale, as well as the DataTAG 2.5 Gbps
link between StarLight and CERN, the Abilene backbone of Internet2,
and the TeraGrid facility. The network routers and switches at
StarLight and CERN were used together with a GSR 12406 router loaned
by Cisco at Sunnyvale, additional Cisco modules loaned at StarLight,
and sets of dual Pentium 4 servers each with dual Gigabit Ethernet
connections at StarLight, Sunnyvale, CERN, and the SC2002 show floor
provided by Caltech, SLAC, and CERN. The project is funded by the
National Science Foundation, the Department of Energy, the European
Commission, and the Caltech Lee Center for Advanced Networking.
One of the drivers of these developments has been the HENP community,
whose explorations at the high-energy frontier are breaking new ground
in our understanding of the fundamental interactions, structures and
symmetries that govern the nature of matter and space-time in our
universe. The largest HENP projects each encompasses 2,000 physicists
from 150 universities and laboratories in more than 30 countries.
Rapid and reliable data transport, at speeds of 1 to 10 Gbps and 100
Gbps in the future, is a key enabler of the global collaborations in
physics and other fields. The ability to analyze and share many
terabyte-scale data collections, accessed and transported in minutes,
on the fly, rather than over hours or days as is the current practice,
is at the heart of the process of search and discovery for new
scientific knowledge. Caltech's FAST protocol shows that the high
degree of transparency and performance of networks, assumed implicitly
by Grid systems, can be achieved in practice.
This will drive scientific discovery and utilize the world's growing
bandwidth capacity much more efficiently than has been possible until
now.
Copyright 2003 California Institute of Technology.
Email prmedia@caltech.edu with comments, questions or
suggestions