Switch fabric performance

pavelalex1 · ‎05-04-2007

I am trying to calculate the “limits” for using the Catalyst 3750G stack as a core for 2-tier LAN, where several 2960G play edge switches role.

This solution is designed by Cisco as best for a branch up to 200 users. For a LAN with 250-500 users Cisco recommend Catalyst 4500 as a core. I suppose this number is pure theoretic. I want to find out some practical guidance how to calculate realistic demands for core switch in a concrete situation.

For example I have a LAN with 250 with users and up to 5 servers, all of them with 1Gb Ethernet (this is not significant, it is easier to calculate). Let’s suppose there isn’t traffic between desktops, the typical hub-to-spoke configuration.

As best practice, for 250 with 1:20 oversubscribing (recommended values are 1:15 to 1:20) I must have 12Gb downlink throughput. But there are only 5 servers and I suppose (may be am wrong?) I don’t need these 12G uplinks. So, how thick must be common downlink throughput for these 5 servers? 12Gb or only 5Gb is enough?

And let’s suppose I have more serves – 20:-). I suppose in this care I have to ensure this 12Gb throughput from edge switches to core. For most switches Cisco provide the switch fabric performance values. In this term how “big” must be the core to ensure these 12Gb downlinks:

- 12Gb switch fabric;

- 24Gb (because I have 1Gb full duplex);

- 48Gb (because core must process traffic from uplinks and “responding” one from servers)?

As far about my example with Catalyst 3750G. I meant a stack from one 3750G-12S, where all uplinks are connected and one 3750G-24T, where serves are connected. As edge switches are intended 2960G. In this case all traffic between desktops and servers pass through 32Gb StackWise bus. This “32Gb” is enough or not for previous examples?

Danilo Dy · ‎05-04-2007

Hi,

It maybe theoretic or marketing strategy, but you won't be able to get the right number until you see it for yourself. To do this, you may need to contact a vendor for POC.

Monitor all aspects of performance - throughput, cpu, memory, buffer, latency, etc.. while using a stress test tools to simulate a expected load (or until it breaks) if you put the device in production.

Dandy

pavelalex1 · ‎05-06-2007

I agree with this.

But there must be some playing rules and general advices? For example, oversubscribing rules 20:1 and 4:1 for desktops and for servers was developed in someway.

In fact I want to understand two things:

1 – Must I take into consideration number of servers in designing LAN or it is preferable to follow the oversubscribing rules?

2 – What means switch fabric performance? It is his overall throughput or maximum volume of information it is capable to process (summing incoming and outgoing packets)?