[e2e] new high speed TCP protocol

rhee at unity.ncsu.edu rhee at unity.ncsu.edu
Fri Aug 22 14:52:23 PDT 2003


I sent this email to e2e several days ago, but it seems it 
did not make it to this list. So I am resending it. If you 
received this post multiple times, my sincere apology.


Previously we talked about the fairness problems running variants of 
high speed TCP, in particular HSTCP (high speed TCP by Floyd) and STCP 
(scalable TCP by Kelly) under drop-tail routers in fast long distance 
networks. We finally nailed down its cause. It is RTT-unfairness 
where flows with different RTTs are subject to huge unfair use of 
network bandwidth. The RTT unfairness of TCP has been known; however, 
we find that the RTT unfairness of HSTCP and STCP is much more 
severe than TCP. For instance, a STCP connection with a short 
RTT can dwarf other STCP connections with larger RTTs (dropping their 
throughput to almost zero). 

Recognizing this limitation, we develop a new (loss based) CC protocol 
that exhibits similar RTT unfairness as TCP while providing 
scalability under low loss rates and TCP-friendliness under high 
loss rates (the same features as HSTCP and STCP except for RTT unfairness). 

Our report and the NS test suite validating our findings can 
be found in the following link.


The current NS implementation of STCP and HSTCP is too slow to do 
any meaningful experiment. So, we modified the NS implementation to 
speed things up. If you are looking to do high speed TCP simulation, 
you might find our work a little helpful (its speed improvement 
is in the order of 10 times).

We appreciate any comments you might have on our work. 

Injong Rhee
Computer Science Dept
North Carolina State Univ.
rhee at csc.ncsu.edu

More information about the end2end-interest mailing list