Re: Towards Consensus on TCP Connections

[julian_satran@il.ibm.com]

David writes:
___________________

.....

It is an established practice in the storage industry
to achieve both 1) and 2) via the use of multiple
SCSI connections, and evidenced by numerous products
that do this for host access to storage.  Hence a
"no" answer to (B) does not preclude the use of
multiple TCP connections for bandwidth or fault
tolerance/recovery as long as one is willing to use
multiple SCSI connections to do so.

So, to resolve issue (B), we need to discuss whether
there are compelling reasons for doing 1) and/or 2)
within a single SCSI connection, as opposed to across
multiple SCSI connections.  This specific issue needs
further discussion -- I invite everyone to have at it ...

--David

___________________

That is a fair description of the status-quo except for the following
points:

i) a good number of mainframe class controllers aggregate bandwidth and
links underneath
the channel layer (the rough equivalent of SCSI) for reasons similar to
those
neatly stated in Michael Krause comment - anything above the channel layer
is oblivious
to the channel configuration (NB the aggregation scheme used there is even
more radical
than the one in the current iSCSI draft - even the allegiance mechanism is
missing)

ii) parallel SCSI buses have different failure modes than a SAN; with
parallel buses at low speed
 - it is safe to assume that if a command fails it is the device/controller
that is failing and retry is seldom an issue. As a result of this history
applications and devices HANDLE BADLY recovery at SCSI level - the most
notorious being the tapes. The FC community has realized this after the
first version of FCP was out and gone a long way to enable recovery at the
channel level (introduced command numbering to solve this recovery issue).
Considering that failures in a complex  network have a substantially higher
probability than on parallel buses and that percolating them to the upper
layers looks like a bad idea  we had only few design choices - expect TCP
to introduce a recovery mechanism or provide one of our own.

Another way (and a more abstract one) to look at it is that the SCSI stack
was designed
with simple conceptual layers (device, transport, physical interconnect)
and as  the physical interconnect gets more complex you have to adjust the
transport in order to maintain simplicity at
the device (SCSI command) level.



Julo