Essentially you need two vmkernel ports (on the same vSwitch) and manually bind both to the software iCSI initiator via SSH:

esxcli swiscsi nic add –n [vmkernel-port] –d vmhba33

Each vmkernel port should be bound to a single vmnic only via client. Then change the LUN’s multipath properties to vmware round-robin.

Finally, set the properties of each LUN through SSH to change the default IO’s limit of 1000 used for load balancing between these connections to a much lower value, say 3 (needs to be scripted as it will be lost on host reboots):

esxcli nmp roundrobin setconfig –device [lunid] –iops 3 –type iops

This will provide full bandwidth even to individual VMs – I measured 220MB/s read to a WIndows VM running on ESX 4u1 configured in this way with iSCSI storage.

HTH

The VMWare person is right. By splitting the ports and assignig datadtores to each, it will give you multiple paths to the storage forcing the use of 2 physical ports. In doing so, doubling the speed. Each datastore has to access the San or nas throught a different IP for that to work correctly.

This is the reason why a lot of the VMware community have changed to doing software iscsi through the VMs operating system to access storage instead of relying on ESX. Basically creating a base disk in esx to install the OS and then software iscsi or NFS mont more storage through the VMs OS. This spreads the storage traffic through all NICs in the infrastructure. Make sense?

We had 4, 1 Gb nics dedicated for the environment. The engineer said to split two and two. Leave two of the nic’s in vSwitch0 and then drop another 2 nics into vSwitch1. He also stated that you should have the first vSwitch dedicated to particular data stores, and the second to different or you would see performance degradation.

I am looking into if the VMKernel port really does just use one active nic at a time. I am in full agreement, you should be able to pin NFS/iSCSI traffic to particular interfaces much like you can pin VMotion to a particular interface.