[ofa-general] IPOIB CM (NOSRQ) patch -memory footprint

[Shirley Ma]

Hello Michael,

> I've just answered in another thread. Summary:
> I think that to enable connected mode on ehca, what we need is
>
> 1. A way to make IPoIB fall back on datagram mode when you run out of
>    resources.  This might need to be addressed at the protocol level.

My point of view this, if we run out of resource, there is nothing we can
do. There won't be any new connections, just like native RC mode. I think
this is a generic RC issue, w/o SRQ just will hit this sooner. In the
reality, our PPC cluster won't hit this because of the cluster size and
memory configuration in this cluster.

Anyway no matter what this issue we should address in the future. Can we
delay this work to find a solution for RC w/i or w/o SRQ later? We do want
IPoIB-CM mode for the performance gain and interoperability between our
xCluster(Mellanox) and pCluster(eHCA) in coming OFED-1.3. So let's keep
what it is without any parameter tuning, does this make sense?

> 2. A way to separate the noSRQ hacks more cleanly. This is not just me
>    being a micro-optimization freak. I suggested some ways to do this
better.
>
>
> --
> MST

Yes, that needs to be fixed.

Thanks
Shirley Ma
IBM Linux Technology Center
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20070524/a7a0a17f/attachment.html>