RE: A question on Zero Copy

[Douglas Otis <dotis@sanlight.net>]

Y.P.

If you imagine even RAID controller devices have memory management that
operate in a manner to resolve data discontinuities and header structure
placement, then post mapping replaces look-ahead filtering.  Such post
mapping has greater freedom.  The next device to access these structures
will see a coherent image upon completion of a post parse.  Restricting PDUs
to word units would be helpful in this area.  There is not any deterministic
means to predict a response from the Initiator with respect to iSCSI other
than in a general sense.  Commands and data may occur in any order should
there be outstanding R2T and an open command window.  I would consider a
RAID controller to be a dedicated appliance.  The typical computer has an
MMU with pages larger than the typical SCSI block where there is also no
page alignment.  If the data being sent was a consistent FC frame, then
reconstruction and alignment of this frame would be easy with but a single
frame delay.  The hardware to handle an FC frame is already in place.  iSCSI
makes this task less bound.

Doug


> > From: Douglas Otis [mailto:dotis@sanlight.net]
> > Sent: Wednesday, December 06, 2000 2:03 PM
> >
> > Y.P.
> >
> > In Julian's defense, I suspect most targets will be dedicated appliances
> > where memory handling is specialized to the point of not being
> influenced
> > greatly by discontinuities created in blind placement of
> > segments.  I doubt
> > there is anything within the protocol that would allow a target advanced
> > knowledge of segment content.  Otherwise, you are right about needing
> > content directed zero copy witin the target as well.
> >
> > Doug
>
> I believe you are referring to disk and tape drives as "dedicated
> appliances".  In such context you are totally correct.  However, I believe
> most iSCSI devices will be RAID or SAN-like boxes with JBODs.  The LUN and
> block address of an iSCSI command are mapped into a cache memory location.
> R2T is returned with a target task tag.  The incoming data
> messages will be
> directed per target task tag to cache memory which later will be
> flushed to
> JBODs.  In such context, you could say the LUN and iSCSI block
> address of an
> iSCSI command as "advanced knowledge of segment content."  Of course, the
> unsolicited write data do not apply.
>
>