Re: [Tsvwg] ECN & PMTU

["Eric A. Hall" <[e-mail]>]

[I promise not to get sucked into this again, but a few clarifications are
in order.]

Jacob Heitz wrote:

> Source quenching was always controversial.
> It is considered higher impact, because it adds
> an extra packet to the network. Setting a CE
> bit adds nothing that was not there before anyway,
> so is considered to be safer, or lower impact.

...unless you also consider the TCP retransmissions to be expensive.

> Yes, it's slower than having the congested router
> send an ICMP.

> The end receiver sends an ACK anyway. Let it modify
> or withhold that ACK. This is what the receiver can
> do. It would send the ACK anyway, thus nothing
> extra is added to the network.
> 
> But consider these points:
> 
> A network is made of different makes/models of routers.
> A router may consider itself congested in all
> directions at once when it runs low on buffer memory.
> Then it may send ICMPs every which way.

Is it going to start dropping packets anyway?

> If a path is congested at many routers, each may send
> an ICMP in the reverse direction, resulting in a
> multitude of ICMPs reaching the sender.

Good point.

> When congestion collapse looms, safety first is the
> golden rule: add nothing more that was not already
> there anyway.

On half-duplex links, that is a good idea.

-- 
Eric A. Hall                                        http://www.ehsco.com/
Internet Core Protocols          http://www.oreilly.com/catalog/coreprot/