|
Hi there
Firstly - do you mean NFS (as in your subject) or iSCSI (as in your
message body)?
NFS is all but impossible to load balance due to the ephemeral nature of
several parts of the protocol.
iSCSI, being based on TCP and using well-known and defined ports, should
be fairly easy - however I don't see any benefit at all in load
balancing several iSCSI targets (servers) back to one initiator
(client).
Because iSCSI is a connected protocol - ie. there are very long-lived
sessions at play - then you'd only ever see four sessions on your
director (when correctly configured, of course); this would correspond
to four mounted LUNs (devices) on your client.
At this point, how your application accesses the data on these LUNs
(which you say is identical...) is entirely up to you.
I can see that, for example, using a gigabit ethernet network you move
the bottleneck back to the disks on each server, so combining four would
allow you some benefit - however, do you know what your actual maximum
throughput of one target/initiator pairing is?
It's an interesting idea but I'm not sure that LVS is the way to do it.
If you really need the throughput, why not just mount the 4 LUNs without
any intervening hardware and see how it goes from there?
Graeme
|
