There are 3 basic types of switching (in order of increasing performance).

        process switching
        fast switching
        autonomous switching

Process and fast switching support inbound and outbound, simple and
extended, access lists. Of course, for fast switching, such lists only
restrict traffic on the particular fast-switched interface.

Autonomous switching is done in the switch processor, a microcoded device that
is capable of switching IP, IPX, and bridging packets in the 100kpps range.
This is known as the "SP" card on the 7000 and the CBUS controller on the AGS+.
Encapsulation support is rather limited (Ethernet, HDLC, HSSI...).

The cisco 7000 also supports:

        silicon switching

Silicon switching is done in the silicon switching engine (creative, eh? ;-).

The silicon switch processor (SSP) is the board which combines both the
switch processor and a silicon switching engine.

The SSP supports simple and extended outbound access lists in 10.3 and later.
The SSP supports simple and extended inbound access lists in 11.1 and later.

The cisco 75xx series supports:

	"optimal" switching (cruddy name, eh?)
	"flow" switching
	"distributed" switching

* "optimal" switching (cruddy name, eh?)

The 7500 platform does not have a separate SP or SSP card, rather the RISC
processor on the "integrated route/switch processor card (IRSP)" handles
switching directly, similar to the 4000 series routers.  There are several
hardware and software enhancements made though to increase the throughput to
a level that is several times above what you would normally get from "fast"
switching.  Everything that "fast" switching supports is supported in
"optimal" switching.

* "flow" switching

Basicly the "optimal" switching method, however things have been front-ended
with an additional small "flow" cache.  This flow cache contains information
about source/destination addresses & ports which allow the router to make more
informed queueing decisions and process access lists faster.  This is a win in
routers that would tend to carry a reasonably small number of flows at any one
time, such as what you would expect in a corporate network or in a smaller
internet service provider network.  It's unclear if there are any advantages
in a large internet backbone.

* "distributed" switching

cisco has announced a new type of interface-processor card, called a "VIP"
available in the 7500 platform that is intelligent enough to switch packets
with no intervention on the part of the IRSP card.  This once again separates
switching from routing, as in the earlier CBUS/SP/SSP design.


The first packet of every session or connection is always Process Switched.
The route table is consulted (this resides in DRAM on the CPU) and the
"result" is cached in the system memory cache. If the protocol can only be
process switched, then it will continue this way and interrupt the CPU for a
route table lookup each time. [comment: Process Switching is brutally slow
compared to other switching methods. Some features (usually new features do
this for the first few software releases) force every packet to be process
switched. If you can't avoid process-switching every packet, at least get a
router with a fast CPU, such as the 75xx, 4500, and 4700. The 4700 is
currently the fastest at process-switching packets, with the 4500 and 75xx
tied for second. The 75xx can optimum-switch, however, so it's a lot faster
than either of the 4x00s, if you can use it).

The second and subsequent packets of each session are capable of being Fast
Switched (more session types are becoming fast-switchable), and will consult
only the route-cache. This still involves a memory lookup on the board, but
the packet can be transferred from the source card directly to the
destination card without requiring full storage on the CSC [the CSC refers
to the CPU card, basically].



There are some undocumented commands that are useful for obtaining
per-interface statistics on what sort of switching was performed.

For instance:

frobozz-magic-robot>sh int atm4/0 switch
ATM4/0    
         Throttle count:          0
     Protocol       Path    Pkts In   Chars In   Pkts Out  Chars Out
           IP    Process     104851    7669968     116378   11180988
            Cache misses      35826
                    Fast          0          0          0          0
               Auton/SSE          0          0          0          0
frobozz-magic-robot>sh int atm4/0 stat
ATM4/0
          Switching path    Pkts In   Chars In   Pkts Out  Chars Out
               Processor     105024    7679155     116422   11184108
         Route cache/FIB          0          0          0          0
       Distributed cache          0          0          0          0
                   Total     105024    7679155     116422   11184108