[e2e] TCP un-friendly congestion control

[Guglielmo Morandin]

On Friday, June 6, 2003, at 12:05 PM, Craig Partridge wrote:

>
> In message <3EE0E340.4030102 at cisco.com>, Guglielmo Morandin writes:
>
>> Craig, you will take 18 rtt if the cwnd doubles every rtt.
>
> Actually, if it grows by 50% each RTT.
You are right.
>
>> But after a packet drop the congestion window is supposed to grow by
>> just one packet every rtt, isn't it?
>
> Only after hitting half the congestion window.
>
I was assuming an optimistic case in which slow start always works, and
the window oscillates between 50% and 100% of the average bandwidth.

> 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.

> CWND runs at 500 Mbps and it takes only a few RTTs to grow to that
> size.
>
> After CWND (at which times we're at 500 Mbps) we grow linearly to 1 
> Gbps.
>

Yes, but assuming packet size 1500, to be able to grow to 1Gbps you 
need roughly 4000 rtts: more than six minutes.

Given the linear increase, the average number of packets you will send 
per window is about 6000.
So your drop rate must be less than 1/24E6 packets.
In bits, that is 3.5E-12. It seems very low to me.

If you run the same exercise for a tcp able to sustain 7.5 Gbps, having 
an available capacity of 10Gbps,
there must be leass than a drop every 4000s: one hour and six minutes, 
and the BER must be less than 3.5E-14.

If you are also sharing bandwidth with some other flow (say some short 
lived tcp connection that
occasionally pops up and will try to ramp up with slow start), drop 
rate might be much
higher than one packet every six minutes.

> That says our average data rate is just under 750Mbps for a 1 Gbps 
> link.
> (Assuming no error loss -- which starts to become a factor on some
> channels at these rates).
>
>

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.

Thanks
-Guglielmo