[e2e] TCP un-friendly congestion control

[Cannara]

It's actually more than looking carefully at BER, and assuming "90% of
connections never see a bit error and 10% do" -- the latter is particularly
hard to match with reality when many connections share an errored link -- a
situation wireless is making more common.  When one looks at proprietary mods
to TCP, as in the some of the storage industry, it's clear that attention to
buffering, window handling, etc. can bring TCP transfers up to full link
capacity.  However, as soon as losses occur, TCP recovery is poor, unless
error knowledge is included.  The problem with TCP is that it's an error-loss
amplifier.  A 10% loss will bring any connection to its knees.  A 5% loss on a
typical WAN link (e.g, that includes a wireless link) will cause a common
application running on TCP to incur a 50% time penalty.  Until a version of
TCP, or something else, deals properly with error loss, using TCP as a general
transport in modern communications is guaranteed to cause performance
problems.

Alex

Craig Partridge wrote:
> 
> In message <4CA02219-9918-11D7-AEC6-003065DC1D5E at cisco.com>, Guglielmo Morandin
>  writes:
> 
> >> So run that logic.  Suppose the available capacity is 1 Gbps (and if
> >> it isn't, why are we having this discussion?) and so around the point
> >> we hit 1 Gbps we regularly get a congestion loss.
> >>
> >Isn't your assumption a little too strict to be useful in reality? If I
> >do have guaranteed
> >1Gbps capacity, I don't need tcp. I could just shape my flow at a rate
> >slightly
> >lower, and retransmit packets that are dropped because of bit errors.
> 
> Well, run that logic.  If you have 1 Gbps guaranteed, you can set your
> TCP window to the right size and run congestion loss free too....
> 
> I agree that one has to look very carefully at BER (though remember that
> bit errors are not evenly distributed -- what you're likely to see is
> a situation where, say, 90% of connections never see a bit error and 10%
> do).
> 
> >I doubt that 750Mbps can be sustained by standard tcp in a real
> >network, unless available capacity is much bigger.
> >But bandwidth is not THAT cheap.
> 
> TCP running at 700+Mbps over real networks was demonstrated by Cray over
> a dozen years ago.
> 
> Craig