[67022] in North American Network Operators' Group
RE: CIsco 7206VXR w/NPE-G1 Question
daemon@ATHENA.MIT.EDU (Lincoln Dale)
Sat Jan 31 00:14:37 2004
X-Envelope-From: ltd@interlink.com.au
X-Envelope-To: nanog@merit.edu
Date: Sat, 31 Jan 2004 16:13:33 +1100
To: alex@pilosoft.com
From: Lincoln Dale <ltd@interlink.com.au>
Cc: nanog@merit.edu
In-Reply-To: <Pine.LNX.4.44.0401301134470.15221-100000@paix.pilosoft.com
>
Errors-To: owner-nanog-outgoing@merit.edu
At 03:51 AM 31/01/2004, alex@pilosoft.com wrote:
>Keep in mind, 72xx is still flow-based
72xx NPE-xxx is NOT flow-based -- unless you explicitly configure it to be.
(i.e. disable CEF, enable flow switching).
CEF is prefix-based switching - where all possible prefixes (routes/RIB)
are already programmed into the forwarding table (FIB).
anything not programmed into the FIB doesn't exist in the RIB, ergo there
is no route therefore is dropped.
i believe the words you're looking for is "NPE-xxx is SOFTWARE-based
forwarding". this part is true enough - but a NPE-G1 has far more cpu
cycles to switch/route than previous NPE-400/300/225/200/150 et al.
software-based forwarding isn't so bad -- it means that platforms such as
the 7200 typically have lots of features.
this is different to the NSE-xxx which is part software-based forwarding
and part PXE-based forwarding.
the exact features accelerated by PXE varies depending what code release is
used.
your said:
flow-based means router's performance is based on number of flows
established, and first packet of each 'flow' is processed differently
[slower] from all other within the flow, and things like nachi
will kill
it.
no, this isn't true. (at ieast, it isn't unless you explicitly configure
it that way). for a service-provider, you wouldn't want to use it in any
forwarding mode other than CEF, unless there is very good reason to.
to provide you with a summary of forwarding paths and their uses:
CEF switching:
prefix-based pre-populated FIB
dCEF switching:
distributed version of CEF - typically each linecard has its own
FIB and therefore switching decisions are distributed per linecard
Fast switching:
destination-based demand switching. a 'route cache' exists of
destinations to be forwarded to. the first packet to a destination
is "process switched", which installs the route-cache entry.
subsequent packets are switched in the "fast" (aka interrupt)
path.
Process switching:
all packets received (at interrupt level) are queued for process-level
to route.
then there's Flow Switching, whose definition has changed over time:
Flow Switching:
a variation on Fast-switching, but where a flow-entry is created based
on a 5-tuple (srcip/dstip/proto/srcport/dstport/TOS). first
packet is process-
switched, which installs the flow entry, subsequent packets are
switched
at interrupt level
now, Flow Switching has changed over time. you can enable both CEF+Flow
and Flow simply becomes an accounting method that is useful for netflow -
but you continue to have packets switched using CEF.
as to the exact level of forwarding used for each packet, that varies --
if you enable a feature that isn't in the CEF path, then the packet is
switched using the next-lower-layer that supports the 'feature'.
for service-provider type environments, there aren't too many features
necessary for /most/ deployments that aren't already covered in CEF on
7200, so you're mostly ok there.
this is just a brief description of how a 72xx works - and there are many
permutations and differences between different platforms and boxes.
if you want the full rundown, Phil Harris normally gives a "Router
Architecture" presentation at every Networkers i've ever attended, and it
covers all this and more.
cheers,
lincoln.
disclosure: my other email address is ltd@cisco.com, but i work in Fibre
Channel not IP these days.
