In November 2007, Caltech hosted a TCP evaluation "round table", attended in person and using
Caltech's EVO teleconferencing software, by the world leaders in TCP benchmarking. At this
meeting, initial work was begun on standardizing tests suitable for use in simulation and testbed
studies. These tests were selected to test a wide range of common scenarios which TCP handle
correctly. Work is currently under way to implement these tests on WAN-in-Lab.
A publicly available Web interface has been developed to access WAN-in-Lab's nascent
benchmarking suite. This interface allows users to upload experimental kernels, instrumented
with the Web100 tools, and to run a set of predefined tests. The results of these tests are then
processed and placed on the web in both graphical and numerical form. Protocols are tested for
RTT fairness, convergence speed both with and without existing large flows, interaction with
short flows, and fairness between flows traversing different numbers of hops.
FAST TCP:
FAST [2] is a novel flow control algorithm developed at Caltech that has been proven to be very
effective for delivering high and stable throughput over long-distance networks over the last six
years, notably in tests carried out in cooperation with the High Energy Physics community.
FAST has been designed from the start to be amenable to analysis. New understanding of the
dynamics of FAST has recently been obtained, using a new model developed at Caltech in
conjunction with Sweden's KTH. This model captures effects on a finer times-scale than
previous models [3]. These phenomena have a greater effect on FAST than on other TCP
variants because it responds much more rapidly when congestion occurs.
This "differential algebraic equation" model captures an effect known as ACK-clocking with
greater accuracy than previous models, and agrees with numerical results from the popular NS-2
122
Page1
Page121
Page123
Page177