On 15/09/2015 06:45, Jason Gunthorpe wrote:

No, I'm saying the resource pool is *well defined* and *fixed* by each
hardware.

The only question is how do we expose the N resource limits, the list
of which is totally vendor specific.

I don't see why you say the limits are vendor specific. It is true that
different RDMA devices have different implementations and capabilities,
but they all use the expose the same set of RDMA objects with their
limitations. Whether those limitations come from hardware limitations,
from the driver, or just because the address space is limited, they can
still be exhausted.

Yes, using a % scheme fixes the ratios, 1% is going to be a certain
number of PD's, QP's, MRs, CQ's, etc at a ratio fixed by the driver
configuration. That is the trade off for API simplicity.


Yes, this results in some resources being over provisioned.

I agree that such a scheme will be easy to configure, but I don't think
it can work well in all situations. Imagine you want to let one
container use almost all RC QPs as you want it to connect to the entire
cluster through RC. Other containers can still use a single datagram QP
to connect to the entire cluster, but they would require many address
handles. If you force a fixed ratio of resources given to each container
it would be hard to describe such a partitioning.

I think it would be better to expose different controls for the
different RDMA resources.

Regards,
Haggai