Unformatted text preview: of identifying unfriendly ows, and then present simulation
results of the identi cation algorithm in 8 ; we are not aware
of other realizations of the identi cation approach.
One can think of the process of identifying unfriendly
ows as occurring in two logically distinct stages. The rst,
and relatively easy, step is to estimate the arrival rate of
a ow. The second, and harder, step is to use this arrival
rate information along with the dropping rate and other
aggregate measurements to decide if the ow is unfriendly.
Assuming that friendly ows use a TCP-like adjustment
method of increase-by-one and decrease-by-half, one can derive an expression see 8 for details for the bandwidth
share S as a function of the dropping rate p, round-trip
time R, and packet size B : S Rpp for some constant .
Routers do not know the round trip time R of ows, so must
use the lower bound of double the propagation delay of the
attached link; this allows ows further away from the link to
behave more aggressively without being identi ed as being
unfriendly.12 10 As we discuss later, if ows can tolerate signi cant levels of loss,
the situation changes somewhat.
11 If identi cation were perfect, and this management goal achieved,
all ows would get their max-min fair allocations. However, we are
not aware of any algorithm that can achieve this management goal.
12 We are not delving into some of the details of the approach layed
out in 8 where ows can also be classi ed as very-high-bandwidth
but not necessarily unfriendly, and as unresponsive and therefore
unfriendly. Simulation 2
Algorithm UDP Simulation 1
TCP-1 TCP-2 TCP-1 TCP-2
0.906 0.280 0.278 0.565 0.891
0.554 0.468 0.478 0.729 0.747 Table 5: Simulation 1 The throughputs in Mbps of one
UDP and two TCP ows along a 1.5 Mbps link under
REDI 8 , and CSFQ, respectively. Simulation 2 The
throughputs of two TCPs where TCP-2 opens its congestion window three times faster than TCP-1, under REDI,
and CSFQ, respectively.
To see how this occurs in practice, consider the following
two experiments using the identi cati...
View Full Document
- Fall '10
- Eugene Ng