The MPLS WG Archive

[Randy Bush <randy@psg.com>]

hi ping,

> From various studies (Vern Paxson...), it was shown that end-to-end
> jitter can be quite high.

i think it is actually less high inter-backbone than seems to be generally
thought.  e.g. i will append (humble apologies for the non-ascii, but it is
a picture) a ippm-style ripe traffic measurement between verio denver and
ripe in amsterdam.  there are outliers.

> Within core networks, packet delay and loss may be low. But after going
> through multiple domains, NAP's and POP's

i do not mean to be pedantic, but i am not aware of how, from a policy
routing perspective, traffic would go through multiple naps.  i push the
point because the suckyness (sp?) of some of the naps is one of the worst
sources of interprovider problems.

i am not sure what 'domains' are.  if you mean autonomous systems, then
this devolves to naps, inter-provider private circuits, and intra-provider
pops.

a few naps suck, but even they are supposedly being fixed.  and in the
interim, folk tend to route around them as much as possible.

as provisioning is the biggest pain in private peering engineering,
inter-provider private circuits tend to be overprovisioned just so they
don't have to be revisited.

intra-provider pops are back to the backbone overprovision issue.

> I saw data shown that Internet packets go through 5-6 domains on average.

i become more and more curious what a 'domain' is.

> Finally, from all the conversations that I had with SP people, they all
> complained about the complexity of managing QoS with the existing
> protocols, tools and fancy architectures. I never hear any of them
> saying that they don't need QoS.

complexity in the core is not known to improve reliability.

randy

denver to ams