[e2e] Internet packet dynamics

[David P. Reed]

At 01:14 AM 3/15/2004, Jon Crowcroft wrote:
>Iof course the right way to do voice of DSL is to use the ATM multiplex, 
>and then pop the voice out at the POP and
>THEN packetize it in IP where the linespeeds are enough that the 
>serialisation delay, and queueing
>delay jitter  are much less:)

You are being ironic here.   But for others on the list you are too 
subtle.  Here's how to understand it:   If you track the timing of the 
"last byte" of voice in each VoIP packet (not the first byte) or "packet 
yet to be made" you find that the main difference is due to the delay in 
serialization [we spell that with a "zed" :-)].   And if you look at the 
system, the main effect here will be that your "packetizer" is a shared 
resource among many users, if you put the function in the network.   This 
introduces major delay variation that wouldn't be present if the packet 
were built at the source by a dedicated packetizer.

I write this to emphasize that the "end to end argument" is a technical 
argument, and "moving function to the edge" rather than doing it in the 
middle of the network, is NOT a religious doctrine - though the routerheads 
(who used to be the bellheads) still are looking to move functions to the 
point they control, so their marketers can lock us in. (the latter is a 
"political" statement, I know - it is worth, however, noting that there are 
vested interests in the major router companies that constantly push to lock 
function into the core of the network and into middleboxes).