I had a very similar problem with iSCSI failover. It's addressed in this question. You can see my accepted solution that I discovered on my own for info on how I solved it.
Basically it involved setting
node.conn[0].timeo.noop_out_interval = 0
node.conn[0].timeo.noop_out_timeout = 0
node.session.timeo.replacement_timeout = 86400
so that the iSCSI session has enough time to recover before it reports errors up the chain to the kernel.
You're asking about doing a small iSCSI SAN, and you're on the right track. We do something very similar with Dell servers and MD3000i arrays.
In the diagram provided, you show two links from server to switch. I think you can get away with one link, unless you're bonding them for greater throughput. As shown, the arrangement protects against failure of the server NIC, the cable, and the port on the switch. A better (high-dollar) approach would be to add a second switch, and connect each Server to each switch, and cross connect the switches. That protects against loss of an entire switch, but adds the complexity of Spanning Tree, which is the Layer2 protocol for preventing the loop that otherwise appears in the network when 2 switches are introduced. From there, the two switches are commonly attached to two SAN heads, which themselves are cross-connected.. but that's larger scale than you've asked about. Go single-path the whole way, and accept the marginal increased risk in trade for ease of care & feeding.
Regarding ease of care & feeding: Think long and hard about the relative likelihood of hardware failure versus wetware failure. I feel like I see 5:1 ratio of human goofs versus actual HW fail, and so if you're not going to do the mega-buck fully-redundant-everything, keep it simple.
If you enable Jumbo Frames, you've got to do Jumbo Frames everywhere on that network. You've sketched out a dedicated storage network, so you can do it - I'm not so fortunate.
If you ARE bonding those server NICs for throughput, consider adding more bonded interfaces from the switch to the SAN head. If each of N servers is doing X traffic, the san needs to keep up with NX traffic, minus some small oversubscription fudge factor.
I believe the "copper connection" you've asked about is simply copper CAT6 twisted-pair ethernet, as used for iSCSI. As you move into the higher-end SAN world, you see more optical fiber connections, and HBA cards with various modular physical connectors - SFP, GBIC, etc.
Tangentially, how are you dividing up user sessions between the Citrix servers? Is there any sort of active loadbalancing in place (Netscaler?) If you have HA Xen servers, how does the failover process look from the user perspective? Make sure you've identified the mechanism by which this SAN actually improves things for the users.
Edited to add: You might also price out a more traditional shared direct-attach storage cluster. I do these with Windows, so I don't have the details around Xen/Linux, but it's a SAS disk shared between the two nodes. Example being the Dell MD3000 (not the "i" model). (With Dell, you need the proper HBA too, SAS 5/e, iirc) If you're never going to add more compute nodes, the SAS cluster might be easier & cheaper to build. Whatever you end up doing: Validate & Test, test, test. In my experience folks build a cluster to add "high availability" without defining what that means in real terms, and then don't validate that it protects them against the failures they were expecting (hoping, really) against.
Best Answer
From what I can tell, you can do this from XenCenter (and the CLI) but you have to set the MTU before using the interface.
For example, on my servers, if I go to Networking tab, select the interface used for my Management interface in the Networks area, under Network Settings, it states: