If you want to bill accurately, bill off the Layer 2 ports; that's what is always churning the traffic. I've not looked at the accuracy on a scientific level, but I've never found what I believed to be a serious discrepency when billing/polling the physical ports.
The reporting of the Layer 2 and 3 devices, virtual or otherwise appears to be correct; I argue that Cisco attempting to 'populate' the SVI counters with information they are actually not seeing would be 'breaking' the implementation. Remember folks, we're talking about multi layer switching/routing here; the SVI isn't processing all of the traffic and should not lie and say that it is.
Hudson Delbert J Contr 61 CS/SCBN wrote:
What about the cases where the customer has more than 1 port on your
switch, you must then aggregate the traffic from N ports, discount the
data between the local hosts and only bill for the actual up/down from the
switch to the core, no?
That seems complex, of course perhaps only 1 port per customer makes some
sense in these cases too, eh?
This too is a discussion argued a number of times previously. Personally, I prefer the architecture where one port belongs to one VLAN; this is obviously not appropriate in all situations, but it is in mine.
Nothing in this world is free, and the bandwidth that a customer uses across my network is not either, regardless if it's in between their own two servers. In instances where a customer has multiple machines which require communication between one another, it is held at the customers discretion to purchase a private switch and second NIC(s), so our billing system remains ignorant, or get billed for the traffic.
If you are someone who enjoys living dangerously, there are also a variety of Flow based accounting systems and Probes which would allow you to bill based on the flow/IP accounting, rather than SNMP on your access devices. This can be done either through your choice Layer 3 device or a third-party promiscuous probe.
I'm sure that everybody here has their own idea on best how to do this, and what is 'right' for them; my argument is only that falsifying data through propagation from multi layer switching does not at all seem to be the best way.
Christopher L. Morrow wrote: