I'm trying to validate the deterministic model for TCP throughput in simple scenarios.
According to the model, the throughput equation is throughput [pkts/sec]
= (1/RTT) * sqrt[1.5/p]
where p is the loss rate over the TCP flow path.
The topology consists of a source and destination nodes connected through
a 20 Mbps/1ms duplex
link with RED queue. Attached to the source node are a set of FTP sources
sending 1000 bytes
packets.
Varying the number of flows, measurements show that TCP flows equally
share the link capacity,
however, substituting the measured loss rate (via ns trace ) and RTT
over the link in the
throughput equation yields an overestimate of the flow share. The computed
share multiplied
by the number of flows increasinlgy departs from the total link capacity
as the number of flows increases.
Since the RTT does not change significantly, the problem seems to be
inaccurate measurement of
the loss rate. An earlier thread on this list mentioned a problem with
source node (as opposed to link)
dropping packets for certain BW-delay combinations. These drops may
not be accounted for in the link trace,
but also I don't see them in the nam animation.
I appreciate your insight.
Regards to All,
-Mahmoud
-- Mahmoud S. Elhaddad Computer Science Dept. NC State University EGRC 325, (919)513-4201