If you read a majority of my posts you may think I’m a Cisco hater. I’m not. I’ve made a career from them and it has treated me well. However I am not a zealot and I have no problem pointing out things that really suck. Speaking of suck leads me to this post. It’s 12:30AM and I’m dealing with an upgrade of a stack of 9, yes 9, 3750-X’s. First, a stack above two switches is completely stupid. There are too many bugs especially in the HULK LED processes (link status) that cause extremely high CPU utilization. IOS upgrades may or may not fix this issue. My biggest issue is that the CPU is almost always taxed whether it be the HULK LED or the IGP or STP or something else. In my experience stacking stinks. I love it with 2 switches but anything more I would use a 4500/6500 chassis. It’s just too resource intensive and unreliable. I like the platform performance, but honestly I try and sell the 3560-X instead of the 3750-X if there will be more than two switches. I would rather deal with STP which is easy to troubleshoot and tweak rather than stack which Cisco hides behind smoke and mirrors (where is the debug guys?).A single administrative point?? Big deal when it compares to reliability and performance. Yes it may cost me more for things such as performance monitoring (think Solarwinds elements) but it’s a small price to pay. Here’s my advice: stack with no more than two switches (two switches are a good redundancy point), anything that requires more density look at the chassis switches. Think I’m crazy? Check out the Cisco support forums and you’ll realize I’m not as crazy as you think! Good luck stackers.
Update: So the problem I was facing this night was that one switch would not join the stack. Nothing in the logs about incorrect versions…actually nothing in the logs other than the switch was removed from the stack. The root cause was that the reload never ‘occurred’ on this particular switch. Needless to say a 45 minute drive to the client site (in snow) and a simple power cycle resolved the problem. I wish switched PDU’s were cheaper.
Update #2: I did a router & switch refresh at a client and I was finally able to summarize and filter incoming routes. The 3750-X stacks are now only peering with a router and sharing < 10 routes. It made a significant difference in CPU utilization. It's still high, approx 30%, but it is behaving better.