Re: [Ips] Storage Maintenance (storm) BOF reminder & requests

Caitlin Bestler <[email protected]> Tue, 24 Mar 2009 16:46:04 -0700
Newsgroups gmane.ietf.rddp
Message-ID <[email protected]>
Robert D. Russell wrote:
> 
> 
> In any case, it might be useful to have a "Standard"
> way of bringing up initial connections in RDMA mode
> without the necessity to go through the negotiation.
> This may be faster, and in many situations it is
> already known that RDMA mode will be used, so the
> transition is just an extra, unnecessary step.
> 

Any extra unnecesary step is strictly in the local API.
Setting up an RDMA connection requires some exchange of
ULP information so that the two RDMA endpoints can be
configured compatibly (the server end, for example would
not know which Protection Domain to assign until it knew
who it was connecting with).

There is no real overhead in allowing an indefinite amount
of TCP exchange before the MPA Request/Response if you don't
use it.

In retrospect, if we had been confident that iSER would be
the only ULP requesting this capability then it might have
made sense to have iSER change. But the complexity of this
option will always be there, so the only real issue is if
it is adding round-trips or network bandwidth.

If the ULP wants to jump directly to the MPA Request/Response
it is free to do so, and there aren't any real optimizations
from that exchange that would justify a change.

The one exception is the one David already identified: MPI
and other cluster protocols really need a "nop" at the MPA
layer so that the first message from the active side requirement
of MPA can be met without involving the application.