RE: Connection Startup

I'm not sure we're going to converge on this discussion, so I suggest we
agree to bury the hatchet, make sure the feature works, mandatory to
implement, and optional to use, and move on.


Jim




> -----Original Message-----
> From: Caitlin Bestler [mailto:[email protected]]
> Sent: Friday, October 03, 2003 3:34 PM
> To: Jim Pinkerton
> Cc: [email protected]
> Subject: Re: [rddp] Connection Startup
> 
> 
> On Friday, October 3, 2003, at 04:48 PM, Jim Pinkerton wrote:
> 
> > The reason I bring this up so strongly is for *new* protocols, which
> > are
> > designed to run on top of IWARP, and don't care about compatiblity
with
> > InfiniBand, I see no need for this capability. It's far more optimal
to
> > embed their config data in their existing configuration handshakes,
and
> > not add a new round-trip handshake.
> 
> That is a valid concern when adding an RDMA mode to an *existing*
> protocol, such as iSCSI. And the number of *existing* protocols
> that are LLP neutral is, well, pretty much zero.
> 
> Use of an RDMA negotiation message *after* other negotiations would
> indeed be a wasteful additional round-trip. The intent is to support
> applications that only need minimal simple parameters exchanges in
> order to establish the session. When this topic has been discussed
> in the DAT Collaborative, there has been little interest in engaging
> in more complex session setup.
> 
> The only thing that *must* be negotiated before enabling RDMA is
> the essential RDMA configuration (identifying any existing session,
> RDMA read credits, maximum untagged message size, etc.) to the
> extent that it is not pre-dictated as constants for that ULP.
> 
> The general conclusion I have seen application developers who
> are working on an RDMA specific applications is that complex
> detailed negotiations can be done *after* the connection is
> established.
> 
> Now you can argue that they were merely thinking that way
> because that is the only model that InfiniBand supported.
> Or perhaps, that is the only model that InfiniBand
> supported because it is a fine model that meets the
> needs of ULP designers who are assuming RDMA capability.
>