One interesting find I made while
working with the Seneca students on Fedora ARM was that a loopback filesystem hosted on top of an NFS share can outperform the NFS share. Yes, it's counter-intuitive, because that would seem to introduce additional layers of processing, but I think it makes sense.
When using NFS, file metadata management is performed by NFS. When loopback-mounting a filesystem in a file hosted on NFS, the file metadata management takes place entirely on the local system -- NFS merely provides a data store. In this sense, it's not much different from iSCS, because the loopback filesystem can't be readily accessed by two separate hosts at once.
In fact, on a small ARM system such as an OpenRD-Client, loopback-ext3-on-NFS over GigE handily outperforms both Class 6 SD and a local USB-PATA drive.
Why not just use iSCSI? Well, for reasons I haven't yet determined, the Fedora iSCSI initiator doesn't work reliably on ARM-- under heavy load, it sends invalid opcodes to the target. This sounds like an alignment issue, but alignment fixups don't cure it. Investigation continues...