Re: Towards Consensus on TCP connections

[Jeff Fellin <jkf@research.bell-labs.com>]

Folks,
After reading all the discussion and debate on this issue, I decided to
research if the SCSI standards (T10/X3) have done with developing other
transport protocols, and how those protocols dealt with connectivity
issues between the endpoints. For the most part the other transport
protocols SCSI over Scheduled Transport (SST), which is a working draft
and the SCSI-3 Generic Packetized Prococol (SCSI-GPP), which is a proposed
technical report have a concept of a Logical or Virtual connection. This
Logical or Virtual connection has at least one connection through the
network, but may have more up to endpoint configuration parameters. So,
it appears the other groups have not come to a consensus as to the approach.

Most of the drafts are available on the T10's web site:
	www.t10.org

Except for the definition of the Scheduled transfer protocol, which is on
T11's web site:
	www.t11.org

The most interesting thing I came up with is the SCSI-GPP definition is
written to allow connections over IP networks. Annexes E and F describe
the mapping and processing capabilities. If this is a work in progress
should we take advantage of there previous start and work with T10 to
have a better definition of a protocol for storage over IP?

What I understand from a quick read of the SST and SCSI-GPP protocol
definitions is they have dealt with the issues of security, error recovery,
connection management, transfer of SCSI bus control signals (bus reset,
abort tasks.

If we don't use the SCSI-GPP than I propose changing the one LUN per
TCP connection group to be a SCSI target id that multiplexes all LUN's
of the target id onto the connection group. I believe this will decrease
the amount of connections and provide data channels for large transfers.


Sincerely,
	Jeff Fellin
	MH 2A-352
	(908) 582-7673
	fellin@lucent.com