[e2e] (no subject)

[Wu-chang Feng]

The basic problem is that TCP is too aggressive when its windows are 
small (in terms of rate) and too conservative when its windows are 
large.  The only way to ensure that an aggregation of small-windowed TCP 
flows across a bottleneck link does not grow by X% per round-trip time 
is to use a multiplicative increase algorithm which makes the per-RTT 
increase a min of (MSS, CWND*X%).  This is the SubTCP modification.  It 
makes TCP more conservative and as a result, trades off packet loss and 
router buffer space for fairness.  

Wu

Tan Koan-Sin wrote:

>On Fri, Aug 24, 2001 at 05:45:30PM -0700, wuchang at cse.ogi.edu wrote:
>
>>For sure, SubTCP is unfair and will not converge.  However, it was meant to 
>>operate in a TCP congestion region which is dominated by RTOs. In this 
>>region, TCP is an exponential decrease, hyper-exponential increase 
>>algorithm which is even worse (as demonstrated by the Morris work).  
>>
>
>Why not make TCP use AIMD or other binomial congestion control algorithms       
>that will converge when congestion window size is smaller that 4 packets?
>