ProCurve 3400cl and 2610 in an HA set-up for linux hosts, suggestions?

[Kumba]

I have a scenario whereby I want to set up an HA network environment for some servers. My network tree is as follows:

[provider-feed-stp] --> [3400CL Core 1] --> [2610 Access 1] --> [Linux Endpoint A eth0]

                                                 |    CX    |

[provider-feed-stp] --> [3400CL Core 2] --> [2610 Access 2] --> [Linux endpoint A eth1]

The provider feed is STP failover so it will flip-flop back and forth to whatever switch is last active. This I do not control and it is fixed. There are no options for RSTP, LACP, etc.

The two 3400cl's will act as my two core switches. They will be connected by two 10gbe CX4 ports in an expansion module. I am open to suggestions to the best way to set these ports, but I am assuming that the procurve trunking option would be the way to go.

The 2610's will act as my access switches. There will be multiple sets of these. My initial idea is that the two gigabit ports from each 2610 will be ran to one 3400cl switch in trunking mode. My second thought is that one gigabit port will be ran to each core.

The last piece of the puzzle is the Linux host with LACP/Bonding across eth0/1. This appears to be pretty stable in the 2.6.34 kernel but I am open to suggestions on other solutions as well.

What I want to accomplish is a minimal impact from a switch failure anywhere in my network. With that in mind, what configuration suggestions would you recommend with the above switches?

[Helper]

hi,

HA network topology will be each 2610 connected to each core.

Linux LACP/Bonding or any other LACP configuration could not be split over multiple switches, you need to use bonding with L3/L4 load-sharing at the linux side.

Bye,

[paulgear]

As Helper said, you can't do distributed trunking on those switch models, but if you don't expect to push the bandwidth you could configure spanning tree and still have every switch connected to each other switch.  Then the Linux servers would connect in a non-LACP bonding mode, such as balance-tlb or round-robin.

In the scenario where you crossbar with spanning tree, one of the redundant paths will always be blocked.  If you think that you really need the bandwidth, you should configure multiple links between the switches as trunks, and connect the switches in a ring rather than a crossbar scenario.  Then spanning tree will block one of the trunks only, and your switches will have the full bandwidth of the multiple links.  The Linux servers can be set up the same way regardless.