Network ZEN: Stackable or Chassis Switches

(1) Failure rate way beyond acceptable if you’re a 100% uptime shop. Most of the failures experienced have been due to stackwise port problems.

(2) Stackwise incompatibilities between IOS versions makes upgrading without taking down the entire stack next to impossible without temporary upgrades to middle versions. You want to get from 12.2(25)SEE to 12.2(50)SE2, but you can’t reload the whole stack at once because your environment has zero tolerance for downtime? You can’t do it in one step…if you reload one switch on the new 12.2(50) code, and he won’t join the 12.2(25) stack because of stackwise version incompatibilities between old and new IOS. You’d have to do interim upgrades to lesser IOS versions to get to 12.2(50). Multiply that problem by 100 stacks you need to upgrade, and you’d rather take a 3750 stack to corporate in San Jose and set it on fire first.

(3) When replacing failed stack members, very often the shiny new member won’t join the stack. He’ll just sit there as “provisioned”, and/or will hang on load waiting for the stack master to send him port information. This is using an IOS version identical to the others in the stack he’s joining. Nothing like having a finite change control window to replace a failed switch and the stupid switch won’t join the stack while the clock goes tick…tick…tick.

(4) Stackwise ports flapping excessively and eventually taking the entire stack down, unless you unplug the failed port, which of course breaks stackwise backplane redundancy and puts the stack at risk for splitting until you can replace the failed switch. Comedy ensues if the stack actually does split and you use cross-stack etherchannels or layer 3 functionality.

When dealing with Cisco on these issues, we asked them how other data centers were dealing with their 3750s, since a 6500 core/dist and 3750 access layer was a Cisco-recommended design back in the day, at least from the SE’s working our account 4 or 5 years ago. All we got were blank stares and nervous silence, as they told us that no other data centers really use 3750s at the access layer, at least not in mission critical environments. 6500s were the better choice (duh).

The 3750s are great in concept, but I’ve seen too many problems with the stackwise ports over the years to ever want to see another 3750 again.

As far as VSS, it’s a similar concept with a very different execution. VSS lab testing has generated warm fuzzy feelings for me thus far (at least when using the SXI code train, SXH was not so great). It’s on my list to create a “dual-active” scenario and test recovery before I’m completely overwhelmed by the warm fuzziness, though.