Re: Concensus Call on Urgent Pointer.

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

David,

> I like the challenge and the way you put it.  But I would
> also like to preface it with a better statement of
> the "framing problem" we are trying to solve.  Some will
> (and have) argued that simply placing the message
> length in the iSCSI header is sufficient, the rest is
> just a performance optimization and not all agree that
> there is a performance problem, just an implementation
> detail.

I think the problem being solved is one of buffer and resource utilization,
in the face of dropped frames. When buffers are already allocated by higher
layers, the off-loaded TCP is unable to use them, in the face of dropped
frames. To hold these frames whose identity is not established in some
"anonymous" buffer at the NIC seems wasteful, when something like UP can
allow TCP to place it in the approriate upper level owner. In that model,
the NIC is only using up buffers for those PDUs which would not be complete
unless the missing frame comes in. Message length in the iSCSI header is
useful also, so all subsequent PDUs can also be placed in its higher-level
buffers, but not sufficient by itself.

The only other alternative I can come up with somehow force iSCSI PDU's
header to be always at the start of a TCP payload, immediately following the
TCP header. That seems to violate the stream nature of TCP, which enables
"standard" TCP implementations to operate without knowledge of framing of
upper layers.

Venkat Rangan
Rhapsody Networks Inc.
www.rhapsodynetworks.com