RE: iSCSI: Reliability of markers (was Re: R2TDataSN and other recovery mechanisms)

["Venkat Rangan" <venkat@rhapsodynetworks.com>]

Somesh,

Thanks for your insights.

> Consider the marker mechanism as an example. If the marker
> falls within the TCP byte stream region which contains the
> header, and the header digest fails, then do you trust this
> marker or skip to the next marker. Or do we need a sum on the
> marker itself.

You point out one of the disadvantages of the current marker
approach. In this respect, the word-stuffing proposal as described
in draft-weber-fcip-encaps-00.txt is a better message boundary
identification approach. One key point you bring up is if there
is just a two-bit burst error with one bit on the Header
and the other bit on the Synch-and-Steering header, you've lost
a reliable way of synching to any part of the stream. This is
possible because the Synch-and-steering header can be right smack
in the middle of an iSCSI header. You may also have cases where the
two copies of the Next-iSCSI-PDU-start pointer disagree on
where the next pdu starts. What then?

With constant-overhead-word-stuffing, all is not lost if the synch
words get corrupted - you skip over enough words in the stream
until you find one to synch with. Also, you do not have nearly
the level of with UFL mentioned in section 1.2.8.3 of the
draft. I don't remember all the technical discussion that went
into the selection of current Synch-and-steering, but it does
seem to have some weaknesses.

Regards,

Venkat Rangan
Rhapsody Networks Inc.
http://www.rhapsodynetworks.com