Re: TCP limitations (was Re: ISCSI: Urgent Flag requirement violates TCP.)

["Randall R. Stewart" <randall@stewart.chicago.il.us>]

Vern:

Just a comment ...

Vern Paxson wrote:
> 
> > Interesting argument. How would you end up paying the extra copy but not
> > paying the extra memory?
> 
> I think you have to pay both, unfortunately.
> 
> If you had perfect framing, then you could avoid the extra copy.  But with
> the Urgent pointer approach, you don't have perfect framing for packets
> that don't begin new ISCSI messages (or if there's coalescence).  So you
> need the extra memory to keep these around somewhere, and a copy to move
> them to where you want them once you know where that is.  I'd argue that
> this copy isn't all that expensive, because it only happens when you have
> a sequence hole, which likely means that the associated stream is about
> to slow down considerably anyway.
> 
> With a different approach to framing, such as a TCP option that tells
> you where each payload is with respect to its parent PDU, or if you used
> SCTP streams, then I think you can avoid the extra copy, and with it the
> extra memory.  But the first of these is outside the scope of the WG's
> charter, other than making a recommendation about them.

I don't even think that an implementation would need to use SCTP streams
to get this effect. Since each DATA chunk has a bit marker on it to
specify if it is the Begining of a message. The streams only make it
easier to multiplex non-related commands without head of line blocking,
a valuable option to iSCSI I might add but not needed for framing....

R

> 
>                 Vern

-- 
Randall R. Stewart
randall@stewart.chicago.il.us or rrs@cisco.com
815-342-5222 (cell) 815-477-2127 (work)