On 28 nov 2005, at 12.25, Florian Weimer wrote:
* Kurt Erik Lindqvist:
Previous router generations couldn't process packets on the fast path as soon as they contained IP options (IPv4) or extension headers (IPv6). Has this really changed?
Not for previous version no...
What do you mean?
That deployed hardware is deployed hardware. It won't get magically upgraded.
Do ASIS exist which can skip over an arbitrary number of IPv6 extension headers, at line rate?
Yes.
(The whole issue may not be a real problem, it only shows that the claim that IPv6 has been designed for efficient forwarding is crap.)
Well...that argument goes both ways. You could also say that the hw vendors didn't follow the development as IPv6 was certainly defined when the current hw was designed.
Are you sure? The extension header concept seems to stem from SIPP (1994), which predates ASIC forwarding (flow-based forwarding was state of the art back then) and layer 4 filters:
Current hardware was designed in last 1-2 years so given the above my argument still holds.
To locate the transport layer header, you must perform limited processing for all options. With a chained header (like the one in SIPP and IPv6) of almost arbitrary length, this task can't really be parallelized in hardware.
With IPv4, you can at least skip over all IP options in a single step (violating tons of RFCs, and perhaps your peering contract).
I am certainly not an ASICs designer and will never pretend to be one, but AFAIK we can do deep packet inspection at 40G line rate, at an (extremely) high price. CRS-1 for example seem to claim just this.
Personally I don't think v6 gives as much increased forwarding capacity as MPLS - none.
But this is a pity because one of the explicit design goals was to use a simpler header format, permitting more efficient forwarding. 8-(
Look, IPv6 is longer (more) addresses. Nothing more nothing less. It' won't give you increased performance over v4, but it won't make it worse either. - kurtis -