RE: Towards Consensus on TCP Connections

["VonStamwitz, Paul" <paulv@corp.adaptec.com>]

> (A) Should iSCSI require a TCP connection per LUN?

David Robinson is not the only one who advocates a session/LUN. I've stated
publicly that I feel that this is the preferred method of operation and the
current version of SEP mandates it. Julian has stated several times that
this is the simplest method. That being said, I do appreciate the concerns
regarding resource requirements. Randy's example used 64B for state as a
wild guess, and a low one at that (our implementation is currently using
about 0.5K/TCP connection), and that is before you factor in how you
allocate send/receive buffers. I still purport that a vast majority of
targets will have the resources to handle session/LUN, even those with many
LUs because they have lots of memory. But, I agree that the protocol must
not require it. That is why the next version of SEP will include the LUN
field in the header. What I advocate is the ability to establish multiple
sessions between initiator and target and assign 'n' LUs to each session,
where 'n' is 1 to all.

Somesh states "the way it [iSCSI] is currently defined allows implementors
to implement one connection per lun or one connection per (arbitrary)
connection of luns. I think this is more flexible as it would allow either
implementation or migration over time from one form to the other." This is
true, but the draft assumes one session per target and there is no provision
to negotiate how many sessions and how to map LUNs to sessions.

Another point, if the targets support ACLs for authorization purposes, it
will know ahead of time what the maximum number of connections.

Paul