[e2e] High Packet Loss and TCP

[Jim Kurose]

Christian Huitema wrote:

>There are two issues, the behavior of TCP and then the possible measures
>to fix it.
>
>At high loss rate, TCP will tend to operate with a window of 1 or 2, in
>a cycle of send one or two packets, wait for a timer, and repeat until
>an ack is received -- possibly a delayed ack. Each failed transmission
>will result in an increase of the timer value, until the timer
>eventually reaches some pre-configured maximum. Eventually, the TCP
>connections will settle to something like sending one packet every 4
>seconds, and receiving only a few of those.
>  
>
Exactly.  The paper that Jonathan Smith referred to in the earlier 
message on this thread is:
  Modeling TCP Throughput: A simple model and its empirical evaluation 
(1998)
  Jitendra Padhye, Victor Firoiu, Don Towsley, Jim Kurose,  
  http://citeseer.nj.nec.com/padhye98modeling.html
which points out the importance of timeout when modeling throughput as a 
function of loss.

Jim

>One can argue that this can lead to congestion collapse: if too many
>connections try to use a too small link, then they will all pack-off to
>this one packet every 4 second behavior, which may well exceed the link
>capacity. Also, the slow progress will lead to many job aborted midway,
>which is a loss of resource. It has been shown that, in these
>circumstances, admission control helps: preferably drop the SYN packets,
>so that the remaining connections have some bandwidth and can actually
>complete.
>
>-- Christian Huitema
>  
>