Networked applications generate messages that are segmented into small
er, fixed or variable size packets, before they are sent through the n
etwork, In high-speed networks, acknowledging individual packets is im
practical; so when congestion builds up, and packets have to be droppe
d, entire messages are lost. For a message to be useful, all packets c
omprising it must arrive successfully at the destination. The problem
is therefore which packets to discard so that as many complete message
s are,delivered, and so that congestion is alleviated or avoided altog
ether. In this paper, selective discarding policies, as a means for co
ngestion avoidance, are studied and compared to nondiscarding policies
. The partial message discard policy discards packets of tails of corr
upted messages. An improvement to this policy is the early message dis
card that drops entire messages and not just message tails. A common p
erformance measure of network elements is the effective throughput whi
ch measures the utilization of the network links but which ignores the
application altogether. We adopt a new performance measure-goodput-wh
ich reflects the utilization of the network from the application's poi
nt of view and thus better describes network behavior. We develop and
analyze a model for systems which employ discarding policies, The anal
ysis shows a remarkable performance improvement when any message-based
discarding policy is applied, and that the early message discard poli
cy performs better than the others, especially under high load. We com
pute the optimal parameter setting for maximum goodput at different in
put loads, and investigate the performance sensitivity to these parame
ters.