[openib-general] IPOIB CM with Non SRQ support

Pradeep Satyanarayana pradeep at us.ibm.com
Mon Jan 29 14:44:48 PST 2007


Hello Michael,

Yes, the code seems to get complex with lots of small changes spread 
across all over the recieve side. Plus 
special cassing them with #ifdef makes it look a little messy. It is 
unlikely I can get this out by Feb 1st.

As I was working through this I noticed a few things and here are my 
observations:

-ipoib_cm_modify_rx_rts() does not actually transition the passive side qp 
to RTS state and remains in the
RTR state. However, the active side qp does transition to RTS.

-One artifact of the current send side implemantation is that for every 
message we create a new set of tx qps.
So, if one were to use IB for the cluster heartbeat mechanism as an 
example, then for every heartbeat we
end up creating an ipoib_cm_tx structure and initiating a set of CM 
exchanges.  This might consume a lot of
 resources (even on an "idle" system). Changing this has a potential 
performance upside.

Pradeep
pradeep at us.ibm.com 

"Michael S. Tsirkin" <mst at mellanox.co.il> wrote on 01/25/2007 11:41:28 PM:

> > Quoting Pradeep Satyanarayana <pradeep at us.ibm.com>:
> > Subject: IPOIB CM with Non SRQ support
> > 
> > 
> > Michael, 
> > 
> > I am working on a prototype based on your IPOIB CM patch to 
> incorporate support for Non SRQ  as well. IPOIB CM was planned to be
> in OFED 1.2 if I remember correctly. If I were to submit a patch for
> non SRQ support, what would be the cut off date to make it
> > into OFED 1.2? 
> 
> I think it must be ready for merge by feature freeze on Feb 1st, but at 
this
> stage it really needs to be a small patch. I can't commit to merging it
> before I see it.
> 
> I have to warn you that I thought about this problem, and unfortunately
> I do not see a way to implement it in a robust fashion without 
complicating
> the code significantly. In this case, you'll just might have to maintain 
it
> as a separate patch until the code lands upstream, and propose as a 
separate
> improvement later.
> 
> -- 
> MST
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070129/187373fe/attachment.html>


More information about the general mailing list