Best place to deploy CSS

[mchockalingam]

Hi All,

I was aked to look into the possibility of using a pair of 11503 CSS to provide load balancing to a pair web servers hosted on a IBM blade server. These blade servers are connected to Layer2/3 switch with routed links to the distribution layer.

Any suggestions?

[Gilles Dufour]

the best place is to be as close as possible to the servers.

Don't use one-armed solution.

So, between your distribution and access layer introduce the CSS.

Add new vlan between distribution and CSS.

If you had

Distr ----- Acc ---- Servers

<------- vlan x ------>

Now you should have

Distr ----- CSS ----- Acc ----- servers

<- vlan y -><------ Vlan x ------>

Like this you keep the servers unchanged.

The CSS becomes the default gateway.

You need to modify the distribution layer .

Physically the topology should be sth like this

CSS1 ==== D1 ======= D2====CSS2

| \______/ |

| /\_____ |

|/ \ |

A1 A2

Don't try to create redundant connections between the CSS and the other switches.

It is best to let the CSS failover to the backup in case of link issues.

This is just my personal view.

I'm open to discussion if you think there is sth wrong with this solution.

Gilles.

[michael.e.reid]

Hi Gilles,

In my company we do the same as you but we are also looking at the possibilty of source natting so that the servers can still use their original gateway. IS this not a good idea as I heard that the CSS is not a particularly good router ?

Also, why are you not a fan or the one armed design ?

[mchockalingam]

Gilles,

I was just looking at the following paper on how IBM blade servers are integrated into Cisco IGESM (Intelligent Gigabit Ethernet Switching Module)

http://www.cisco.com/application/pdf/en/us/guest/netsol/ns304/c649/cdccont_0900aecd8010e768.pdf

and basically each chassis containing several blade servers have 2 Cisco's 2950 integrated switch with upto 8 uplinks to the distribution layer. We are using 2 uplinks to the distribution layer. These uplinks are trunks and not routed links as I mentioned before. Currently all blade servers are on the same vlan and some needs to be load balanced and some not.

Does this mean, I need to create a separate vlan and move all the servers that needs load balancing to that vlan? If not, all traffic to the blade servers will end up going through the CSS and not sure how much latency will be introduced due to that?

[Gilles Dufour]

the CSS is *not* a good router in terms of routing protocols. We usually do not recommend to run ospf or rip on the CSS.

But the CSS routes perfectly well traffic.

The problem with source nat is that the servers will only see 1 client ip address == the CSS.

This could lead to issues and will definitely prevent you to get statistics about origin of requests.

Gilles.

[mchockalingam]

Gilles,

Is it possible to run trunking on both sides of the CSS, meaning on the Access layer side as well as the distribution layer side?

[mchockalingam]

Never mind. I found the following link that explains trunking on the CSS in detail.

http://www.cisco.com/en/US/products/hw/contnetw/ps789/products_configuration_example09186a00801e16c7.shtml