Database Reference
In-Depth Information
Time
Figure 4.3
Theoretical TCP behavior on a link with no concurrent
trac.
acknowledgments and indication of packet loss. In TCP, packet loss is taken as
an indication of congestion. These algorithms are described in more detail in
RFC 2581. *
We can summarize that TCP works by increasing the congestion window
with one segment each RTT (round-trip time for the connection) when no
congestion is detected, and decreases the window to half when congestion is
detected. This algorithm can be described as AIMD (additive increase multi-
plicative decrease). This behavior is illustrated in Figure 4.3.
A long history of utilization has shown that TCP is extremely success-
ful in supporting the Internet trac while avoiding a crash of the system.
However, its characteristics create performance problems for situations where
window sizes become very large, such as when using high-capacity wide area
networks. First, the response to a congestion event (reducing the window to
half) means that on a network link with no concurrent trac, TCP will not
be able to sustain utilization of the full link capacity. Second, the slow in-
crease of the congestion window during the congestion avoidance stage means
that on long-distance, high-capacity links, the time needed to increase the
congestion window to reach the link capacity can be measured in hours. Also
in practice, bandwidth utilization as a fraction of the total capacity will be
lower on higher capacity networks since the probability of a single packet loss
occurrence that is not caused by congestion is higher when the number of
packets traversing the network increases, and the algorithm treats all packet
losses as an indication of congestion.
* http://www.faqs.org/rfcs/rfc2581.html. Accessed July 16, 2009.
Search WWH ::




Custom Search