[Iccrg] SSDT Scope - summary

[Lachlan Andrew]

> Lachlan Andrew wrote:
> > But when we first start, we don't know if we're talking to a 300bps
> > mote or a 100GbE supercomputer.  After a single timeout, we will
> > generally (???) be using a "similar" network.
>
> In times of QoS this assumption may be very wrong. E.g. You have two
> wires to the internet. One 1MB and one 100MB, but for the 100MB only
> HTTP is allowed all other protocols have to use the 1MB link. If youre
> first connection is an HTTP one, then the 1MB link will be overshot, the
> other way round you'll have to wait for a long time to satisfy the 100MB
> if you increase just linear.
>
> The first mentioned problem, with 1MB link get 100 MB data could be
> solved in one RTT, but in this case 99MB data get dropped. This would be
> an overshot of the factor 100 and much worser than slowstart.

If only one is HTTP, then they're probably different TCP connections.

Even ignoring that:
1. going from 100Mbit/s to 1Mbit/s, you will overshoot in the first
RTT, when the rate is much less than 100Mbit/s, get a time-out, and
start a normal exponential slow-start.  I'm not suggesting going
straight   from the current rate to the final rate.

2. going from 1Mbit/s to 100Mbit/s, either way slow start will stop
when  ssthresh  is reached, when the source is sending at ~1Mbit/s.

Cheers,
Lachlan

-- 
Lachlan Andrew  Dept of Computer Science, Caltech
1200 E California Blvd, Mail Code 256-80, Pasadena CA 91125, USA
Ph: +1 (626) 395-8820    Fax: +1 (626) 568-3603
http://netlab.caltech.edu/lachlan