L2/L3 together

[tsims]

My SOP has been to use L2 for access switches and servers. The servers and switches are directly connected to two 6509. I have been thinking about running L3 to the closet because I would like to have less Spanning-tree dependecy but I would still have it because of the servers. The servers are not on seperate switches. Does anyone see a problem with this?

Second question, the access switches are daisy chained would that cause any problems. I would think you would make each connection in the chain and connecting back to the 6509 a point to point link.

Basically, I like working in L3 environment.

[Georg Pauwen]

Hello,

I would not recommend running L3 to the closet(s), it would most likely considerably slow down your network. Besides, there are spanning tree feature enhancements which make failover almost seamless (such as uplinkfast, RSTP). Besides, whatever you run at L3 will take time to failover as well.

Regarding the daisy-chained access switches: obviously you are looking at a single point of failure (that is, whatever is downstream of the failing switch will fail as well).Better to connect each access switch directly to the 6509 switches...

Regards,

GP

[tsims]

I do not know if I would buy L3 being slower. You would still have your vlans locally on the access switch. There would be equal cost paths and the convergence would be almost instant like in your distribution and core layers. Though having said that I will probably stick with L2 for access switches in this case even without the need to really have vlans across different switches. Just to have one environment.

The daisy chain has the switch at the top going to one 6509 and the switch at the bottom going to the other. I use to run HSRP but now will probably run GLBP.

Thanks

TS

[scottmac]

To add an additional note to GP's post, the other major drawback (aside from the single point of failure), is a bandwidth bottleneck.

All of the traffic from the farthest switch must travel through a single active link, all the traffic from the intermediate switch would be added, then the combined traffic has to cross another single active link back to you resources.

It will work, but it's waaaay down the list of acceptable practices.

The actual affect will depend a lot on the scope and nature of the traffic on your network, where your resources are located, etc.

Good Luck

Scott

[jroyster]

Generally you want to make sure and use pure L3 links between core and distribution layers so I'd be all for it. Just makes it so much nicer (convergence time in seconds, predictable traffic flows/routing, much more control, etc).

I avoid altogether and make sure there are no blocked ports anywhere. Also having that layer3 boundary at the distribution layers assists with containing failures, broadast/multicast storms, and all the "bad stuff" that can happen.

here's a great SRND design document...

http://www.cisco.com/application/pdf/en/us/guest/netsol/ns432/c649/cdccont_0900aecd801a8a2d.pdf

for all of them check out

www.cisco.com/go/srnd

[dgahm]

There was a good session at Networkers this year called Deploying A Fully Routed Enterprise Campus Network RST-2031

The takeaway for me was that advances in multilayer switching have removed the performance penalty, and that even with RSPT, routing protocols like EIGRP or OSPF offered superior convergence times. Layer 3 is also somewhat simplier to deploy.

Bottom line, the old L2/L3 model is still viable, but a fully routed network may be a better choice if you have the proper hardware, and available IP address space.

[tsims]

Yes, I was there but not in that session. I have since looked at that session online and it really prompted my initial post.

The only issue I have is that our servers are not on seperate set of switches which connect to the core. So, I will need to do L2/L3 for them because they connect into the same distribution with our "closet" access switches. So, I might as well keep the same L2/L3 model until I can get the servers moved.

I would rather use EIGRP than Spanning Tree but......