[ofa-general] Multipath and IB Bonding

[Nicolas Morey Chaisemartin]

Hello everyone!

In our current project, we are working  with nodes containing 2 IB QDR 
HCA (2 cards with 1 port each).
Both of these ports will be connected on the same interconnect/subnet.
Currently, most of our applications (Lustre, NFS...) only use one the 
ports as both are on the same subnet.
It works, however, half of the available bandwidth is not used and when 
a link is lost, the applications are not able to fail over on the second 
port.

To provide a unique solution to this problem, we are exploring the 
possibility of "IB Bonding". More precisely, virtualizing the libverbs  
(kernel mode only) so the applications (kernel modules in fact) see only 
one interface and QP (virtual) while there may be many underneath.
As we only work with RC QP we will only try to virtualize this protocol. 
we don't have any needs for virtual RD, UC or UD yet.

We have some ideas about the way to implement it but before starting, 
we'd like to have some opinions from you about what might go wrong, what 
might be impossible and so on...
Eventually, if people are interested in this, we'd be glad to share our 
results and get some help on this.

Best Regards

Nicolas Morey-Chaisemartin