RE: Towards Consensus on TCP Connections

[David Robinson <David.Robinson@EBay.Sun.COM>]

> iSCSI intertwines LUN states to save TCP overhead.  In doing so, one tag
> error sorting millions of tags within the mid-stream point of aggregation
> (the iSCSI design) necessitates a catastrophic system reset.  With FC over
> IP, the end point handles errors in accordance to the FC protocol such as
> FCP-2 with simple frame aligned objects.  I count an additional complex
> state machine in the middle of the communication path for iSCSI in addition
> to new adapters with complex (INSANE) state machines at the client.  A
> switch or a router would be a far safer choice for aggregation.

I think you greatly over estimate the complexity, risk, and recovery
issues. What is being described is a specific case of common
multiplex/demultiplex scheme that is well understood and widely
deployed. The complexity of doing it at the iSCSI session layer
is no greater than doing it at the TCP transport layer.  We have years
of experience building reliable TCP port mux/demux software. Furthermore
the same technique is inherent in the Layer-2/3 switches that today
can process 250M+ packets/sec.  I don't see why doing it for IPS
would be any more difficult. I am no FC/FCP expert but I am willing
to bet such a scheme exists in existing fabrics.

Although I have argued in the past for a session per LUN, complexity
of doing the mux/demux at either the session (iSCSI) or transport (TCP)
layer is not a factor in either design.

	-David