[openib-general] RMPP implementation

[Fab Tillier]

> From: Sean Hefty [mailto:mshefty at ichips.intel.com]
> Sent: Wednesday, November 17, 2004 9:08 AM
> 
> Fab Tillier wrote:
> >>1.  Issue one send at a time.  Additional sends are not transfered until
> >>  the first send completes.
> >
> > Isn't #1 the simplest to implement?  Turnaround on the send queue should
> > be
> > pretty quick, so send performance should be fine.  I say do whatever is
> > simplest, and then optimize from there, and to me that means #1 at the
> > moment.  What are the reasons to *not* do #1?
> 
> It's simpler to implement, and would definitely be the easiest to do on
> redirected QPs.  The only disadvantage is that it lowers the throughput
> between two clients.  Also, this is a relatively small decrease in
> complexity with respect to the rest of RMPP.

I agree that it will lower the throughput, but by how much?  I would expect
it to be minimal.  It also allows more concurrent transfers to progress.
I'm thinking that the SA is likely the primary user of RMPP sends, and thus
responding to more queries in parallel is probably better than responding to
queries serially but faster for each query.  The send completion delay is
likely to be less than the RMPP timeouts, so might as well keep many
requestors going than getting a response to any one client quickly. 

- Fab