Appears to be a work in progress here, but since it flatters my own approach to design where I can get away with it, I consider it worth sharing. ^_^
http://www.cisco.com/en/US/docs/solutions/Enterprise/Campus/campover.html
Among the highlights: “proper” use of the 3750 series Catalyst switches. Apart from what may be ever-so-slightly higher backplane capacity that will never be used in a my existing $dayJob environment anyway, there is one architectural difference between the 3750 and the 3650, and one architectural difference only: the ability to do a proper switch stack via the StackWise feature.
If one does not plan to stack the switches, there is no reason to shell out the extra money for a 3750 when the 3650 line will do the job more than adequately. If the money’s already been spent, however, you can do some great things to simplify and optimize your network. These include reduced sevice requirements (no need to run STP* or HSRP), simplified configuration and manageability (done only once on the master switch in the stack as opposed to two physical switches) and increased availability and load-balancing. Of course, this precludes physical separation of the switches by distances greater than the longest available StackWise cable; but if its physical separation one wanted, it’s cheaper to do so with the 3560.
By running etherchannels (which can span multiple switches in the stack) to your up/downstream switches, you can make use of multiple links to your “core” stack instead of having the redundant links blocked by STP. By definition, loops cannot form, reducing complexity associated with STP (especially in multi-vendor environments, with HP defaulting to MSTP, Cisco defaulting to proprietary PVST+, Dell defaulting to god knows what…), and more easily facilitating the spanning of VLANs across access-layer switches – generally not desired for your user-access, but sometimes preferred for server-access where layer-2 adjacencies are preferred for convenience/clustering’s sake. This is best illustrated on page 18 of the document, with the next page illustrating the benefit in terms of number of active, traffic-passing links vs. an STP design.
Not running HSRP elimates the risk of split-braining, which occurs with physically separated layer-3 switches when they lose connectivity to one another; both switches go active for the HSRP virtual IP address, but return traffic only goes towards one of the two switches, blackholing traffic destined for one of the switches.
And it’s certainly easy enough to do; assuming your existing separate 3xxx series each have links to the same upstream and downstream devices, just take one ‘of ‘em offline, configure the other with a nice high priority (“switch priority 15”), connect the “secondary” switch via the StackWise cables and power it up. Recommend administratively disabling all the ports until the appropriate EtherChannels have been set up on all switches.
In summation, a quick general-usage guide for the current models of Catalyst switches, all of which come in PoE and non-PoE, and 12/24/48-port flavours of various speeds (10/100/1000); FYI, their QoS architectures and configuration are identical, differing only in buffer pool size and policer granularity (1Mbps policer intervals on the 2960, 8kbps intervals on the 3560/3750):
2960: Dumb layer-2 wiring closet switch. Replaces the 2950. No layer-3, no routing protocols, no VRF stuff, nothing. Can be managed via IPv4/v6.
3560: Smarter layer-3 wiring-closet/distribution/core switch; successor to the 3550. Can route IPv4 and v6 over all protocols.
3750: Smartest layer-3 1RU switch. Exactly the same as the 3560, but with the StackWise feature to enable virtual switching — presenting multiple physical switches as a single switch.
3650-E and 3750-E: Same as their counterparts, but more backplane capacity (StackWise backplane doubles from 32 Gbps to 64 Gbps), dual AC power supplies, and support for 10GigE.
Save for the E-series, all of the above feature a single AC power supply, and a proprietary DC power input. It can be fed redundant power to this DC input via the RPS 675 (EoS) and RPS 2300. Each RPS can feed up to six different devices; however, the RPS 675 can only actively back up one device. The 2300 can provide redundant power for up to two. If you load up 6 switches on a 2300 and all of their internal supplies fail, four of the switches are out of luck.
As far as dual AC supplies go (again, save for the E-series), one has to take a large step up to the 4500 and 6500-series chassis. Each features the potential for in-box power and supervisor redundancy. 4500 is relatively (of the two) low-cost choice for port-density and minimal features; 6500 fully loaded can hold most if not all of the internet routing table, all manner of line-cards for all manner of port types, and a variety of service modules (load-balancing and SSL-terminating ACE modules, firewall modules, network analysis modules, etc), and a ton of features. Not likely to see deployment in $dayJob any time soon, but the datacenter aggregation switch of choice in large environments.
*Yah, yah, you should still “run” STP so as to prevent accidental loops from forming, I know, I know: http://www.cisco.com/web/strategy/docs/gov/turniton_stpt.pdf
