How to reduce the convergence time?

[b.yan]

The condition is two 3550-12T are core, twenty 2950T-24 are access, every four 2950T-24 bulidup a group, use 1000base-T port connect one by one , then the edge two has a link to the two 3550. All 2950T-24 has use uplink fast, and 3550-12T has config primary and secondary. But when I turn off the primary 3550 to switch to the secondary , I found the STP need run over 30 second, But the custom is a mission-critical environment, so I need to reduce it to max 10 second. How can I do it? Thanx

[davehampson]

Uplinkfast should work and give a failover time of <10 secs. For uplinkfast to work you need to check STP on the 2950 uplink interfaces, one should be forwarding and one blocking. The root bridge should be your primary 3550.Then with a direct link failure between your forwarding 2950 to the primary 3550 root bridge fast failover is achievable.

[milan.kulik]

Can you describe yout topology clearly? Is my undestanding correct that in each 2950 group 4 switches are fully meshed and 2 of them are connected to the core? If yes, Uplink Fast is not sufficient, it will work fine only on that 2 switches with the uplink connection, not on the other ones. I would think about Backbone Fast but I'm not sure if 2950 supports this feature.

HTH,

Milan

[b.yan]

I don't know how to describe the topology in English. It is like a stack form which haven't command switch. Cisco says stack has two form. One is has a command switch, another hasn't. The group 2950T's topology is like which hasn't command switch's one. I don't know if I describe it clear. My English is poor.

[milan.kulik]

OK.

Is your topology following ?

C--C

| X |

A --A

| X |

A --A

Or this?

C--C

| X |

A A

| X |

A A

Or something else?

These schemes are describing just the core switches (C) and one of multiple access switches (A) groups. (The "editor" in this forum is removing additional spaces, so the second scheme means there are no "horizontal" connections between A switches.)

Milan

[b.yan]

The second is very like. But there's a line between the bottom two 2950T.

by the way: you are so bright :) Can found this way to expression the topology.

C-------C

I*******I

A*******A

I*******I

A_____A

" * " means nothing :) it is instead of space. all of 2950T groups are use this form.

[milan.kulik]

Your network seems not to be well designed. Let's just imagine following two access groups:

A1---------A2

|**********|

A3*******A4

|**********|

C1--------C2

|**********|

A5*******A6

|**********|

A7--------A8

where * means no connection.

Let's suppose C1 is the root.

The network diameter (maximum number of switches the frame has to go through on the path from one PC to another) is 6 in this case (A1 to A8, e.g.).

This is OK, the maximum value recommended is 7.

But in the case C1 fails the network diameter increases to 9 (A3 to A5)!!!

This make cause a pretty long convergence time.

UplinkFast will not help you at all in your topology. It needs another uplink in blocking stage to work but all your access switches (except A1 and A8) are having just one uplink.

So the final recommendation:

Change your network topology!

I would remove connection A1-A2 and add connections A1-A4, A2-A3, A3-C2, A4-C1 in each group (i.e. remove A7-A8 and add connections A7-A6, A8-A5, A5-C2, A6-C1 in the bottom access group in our scheme). I don't know what is possible to do in your network, you have to make the decission, the the solution mentioned above is an ideal case.

There are some good campus network guides on CCO, http://www.cisco.com/warp/public/cc/so/neso/lnso/cpso/gcnd_wp.htm e.g.

HTH,

Milan

[b.yan]

I know the theory. And I agree your word. But the custom's budget and need determine this topology. They want 1000base-T uplink, and they don't want pay any more even one dollar. But there's sth. good: all the PC client will only communication with the server which connect C1 and C2 ( use AFT or ALB). There's no any communication between two clients. I want to know if it is useful for I reduce the HELLO package of STP.

[milan.kulik]

I'm afraid there is no way to improve you network STP convergence time if you are not able to change the topology.

I don't understand exactly what you mean by "reduce the HELLO package of STP". Do you mean to increase the hello timer? If yes, I strongly do NOT recommend it. According to your network topology it could cause serious problems.

Generally, if you want to tune the STP timers you should have a really good reason for it and know very well what and why you're doing. The best way is to use some macro (set spantree root on Cat4000, e.g.) which changes all the timers regarding to network diameter. But your network diameter is 6 (near 7 - default) so I wouldn't touch the timers.

Regards,

Milan

[b.yan]

My English is too poor to express my meaning. I want to say " I want to reduce the time length of STP." STP spend about 55 seconds to listening, learning and forwarding. And I want to reduce this time. For example, reduce listening time to 1 second and reduce learning time to 1 second. That's only a try. I think I already have no idea. But your suggest help me much. I want to change the topology as follow:

C1~~~~~~~~C2

II******X*******II

A1************ A4

II******X*******II

A2*************A3

~~~~~~ is GEC

II is 1000base-T connection

X is 100base-T connection

* is nothing

because 2950T-24 has only two 1000base-T port. So some line will be 100base-T. I want to set them to backup line. Do you agree it? Thanx!

[milan.kulik]

YES,

if you change the topology this way and enable UplinkFast on your access switches it should work and convergence time should decrease very much.

Regards,

Milan