Going forward my 10GbE designs will be doing more to answer questions around physical network setup and configuration. Therefore you will see more detail in the diagram then I normally give, especially in regards to how the physical switches are uplinked and interconnected. I had some network design input from Cisco engineers on this to ensure that redundancy and throughput are not compromised once vSphere traffic gets onto the physical switches.
There were many discussions around the use of Load Based Teaming and Etherchannel, neither of which is used in the following design. LBT is not used because licensing does not allow for it. For more information on LBT please check out this link.
LACP is not used as it would not be good design practise. There are very few implementations where LACP/Etherchannel would be valid. For a comprehensive writeup on the reasons why please check out this link.
The following design is based around a segmented 10GbE networking infrastructure where multiple physical switches are interconnected using high speed links. All traffic is segmented with VLAN tagging for logical network separation.
It is assumed that each host has four 10GB NICs provided by 2 x PCI-x Dual Port expansion cards. All NICs are assigned to a single virtual standard switch and traffic segregation is performed by pinning each VMkernel to a specific uplink. This is where design for 10GbE diverges from standard design in 1GB configurations. Typically for 1GB setup you would need at least two virtual switches or three when using iSCSI storage; switch1 for Management and vMotion, switch2 for VMs and switch3 for storage.
In order to gain the performance increase of Jumbo Frames for the storage layer all networking components will need to have Jumbo Frames enabled end-to-end from the hosts through the network and to the storage arrays. There is definitely a performance increase by incorporating Jumbo Frames and this is outlined in the following link. It is important to note that enabling Jumbo Frames on the single switch will allow all traffic to transmitt at 9000MTU. This means that Management, vMotion, FT and Storage will all use Jumbo Frames. VMs will not use Jumbo Frames unless this feature is enabled on the network adapter inside the OS of the VM.
Trunking needs to be configured on all uplinks where all ports on the physical switch allow all VLANs through. Trunking at the physical switch will enable the definition of multiple allowable VLANs at the virtual switch layer. It is important to note that the colours used in the diagram show how traffic will flow under normal circumstances. However all VLANs need to be able to use all uplinks in the event of a NIC failure.
If you are running Cisco equipment then you might be able to use the Rapid Spanning Tree Protocol (802.1w) standard. This means that you do not need to configure trunk ports with Portfast as the physical switches will automatically identify this port correctly. If running any other type of equipment the safest option would probably be to disable STP and enable Portfast on each trunk port, but please refer to your manufacturer manual.
Running vCenter and the vCenter database on the same hosts that it manages is not a problem in this design. So you do not need to run a Management cluster but I would say that it is usually a good design decision to do so. If you built this solution and then upgraded to Ent+ and started using virtual distributed switches then a Management cluster would be required.
This design is based around scenarios where Enterprise Plus licensing and network based bandwidth limiting/control is not available. Because SIOC and NIOC are not available in this design there is no way to guarantee bandwidth for particular traffic types. vMotion in vSphere 5 would be quite happy to consume 8Gb of an uplink and in situations where other traffic is running on that uplink it would be constricted by vMotion.
*** Updates ***
05/05/2012 - Minor update to Jumbo Frames paragraph. Thanks to Eric Singer for his observations.