RE: New List: rdma@cisco.com: to discuss RDMA

["Douglas Otis" <dotis@sanlight.net>]

Randall,

With protocols like SCSI, the data payload is contained within an additional
structure.  Although one could attempt to constrain data to a stream and
make exchanges dynamically allocated to streams, it would be easier to
extract payload from the encapsulation.  From this extraction, either
generically or specifically, transport the payload as described.  Yes, such
optimization would be implementation specific but very possible with SCTP.

Doug

> -----Original Message-----
> From: randall@stewart.chicago.il.us
> [mailto:randall@stewart.chicago.il.us]
> Sent: Thursday, September 28, 2000 5:17 AM
> To: Douglas Otis
> Cc: Charles Monia; csapuntz@cisco.com; Jim Williams; ips@ece.cmu.edu
> Subject: Re: New List: rdma@cisco.com: to discuss RDMA
>
>
> Douglas:
>
> I may be missing something here, but sendmsg() and recvmsg() both
> have scatter/gather arrays within the berkley sockets api.
> I am not sure that one can uses sendmsg() recvmsg() with TCP.. I
> have never tried. As far as SCTP goes, we have defined a sockets
> mapping... the draft is released but real rough.. an next version
> is in the works.... the current "rough" cut can be found at:
>
> http://search.ietf.org/internet-drafts/draft-stewart-sctpsocket-si
> gtran-00.txt
>
> Now it may be that I am missing a key point in this discussion.. were
> you
> thinking other than the iovec structure... or were you thinking a
> modification
> that allows the pages to be stolen with no copy... the no copy becomes
> a implemenation issue.. I have done something like it in the past
> where if the application gives a 4k page in the send, the O/S underneath
> does a swap of a page from kernel space with the user space one...
>
> But as I say, this is real implemenation dependant :)
>
> R
>
> Douglas Otis wrote:
> >
> > Charles,
> >
> > With respect to SCTP, features added by a TCP option for RDMA
> are not needed
> > to support alignment and out of sequence processing that
> ultimately alters
> > the TCP API.  SCTP adds these features without disruption or
> modification to
> > TCP.  The intent of VI is to allow scatter/gather function to
> handled by the
> > target.  A safer scheme would be to adhere to SCSI conventions
> and implement
> > zero copy and out of sequence processing using SCTP and SAM structures
> > related to locally pre-arranged transfer structures.  This
> would keep the
> > initiator or client in intimate control of memory and not
> reliant on targets
> > eliminating boundary checking.  Such a feature will slightly
> impact SCTP to
> > add a means to generally encapsulate a data payload associated with
> > structure of pointers equipped with scatter-gather lists and relative
> > offsets.  Specifically, this could be defined as the FCP structure data
> > structure in SCSI implementations.  I would whole-heartedly endorse such
> > features within SCTP.  I would dissuade such features from
> being added to
> > TCP.  One could view the scatter-gather list structure as a
> token to be used
> > by the target in more conventional VI methods.
> >
> > Doug
> >
> > > -----Original Message-----
> > > From: owner-ips@ece.cmu.edu [mailto:owner-ips@ece.cmu.edu]On Behalf Of
> > > Charles Monia
> > > Sent: Wednesday, September 27, 2000 11:49 AM
> > > To: csapuntz@cisco.com; Jim Williams
> > > Cc: ips@ece.cmu.edu
> > > Subject: RE: New List: rdma@cisco.com: to discuss RDMA
> > >
> > >
> > >
> > >
> > > > -----Original Message-----
> > > > From: csapuntz@cisco.com [mailto:csapuntz@cisco.com]
> > > > Sent: Wednesday, September 27, 2000 10:16 AM
> > > > To: Jim Williams
> > > > Cc: ips@ece.cmu.edu; csapuntz@cisco.com
> > > > Subject: Re: New List: rdma@cisco.com: to discuss RDMA
> > > >
> > > >
> > > >
> > > > Does anybody on the list object to specifying an RDMA mechanism for
> > > > use with iSCSI? Does anybody on the list object to mandating an RDMA
> > > > mechanism? Please include your reasons.
> > > >
> > >
> > > I'd object to mandationg the use of RDMA in iSCSI. However, I
> > > would support
> > > structuring the spec so that an RDMA transport mechanism could be used
> > > underneath (I guess that's motherhood). If the iSCSI folks
> decided this
> > > wasn't a priority issue, that's ok with me too.
> > >
> > > Charles
> > >
>
> --
> Randall R. Stewart
> randall@stewart.chicago.il.us or rrs@cisco.com
> 815-342-5222 (cell) 815-477-2127 (work)
>