[Ofvwg] OFVWG meeting notes - 9/15/2015

[Liran Liss]

Parav presented his recent patch-set that adds RDMA cgroup support.

Controlling RDMA resources in general (without device affinity) is simpler to implement and administer, as this scheme does not need to handle HW hot plug/unplug situations.
However, controlling resources per-device is more meaningful since the resources themselves are per-device.
So, the general scheme holds well when a single device is used, but in the general case, a per-device scheme seems better.

Another approach to consider: use the general scheme as the interface, but apply it on a per-device basis.

The question of whether controlling resources per-port is reasonable. However, it seems that this doesn't add any further control because in multi-port devices, resources are not restricted to specific ports.

It was generally preferred that RDMA resources should have a distinct cgroup.

The notion of soft limits was discussed. However, it seems that there is no natural definition for this, since HW resources are not migrated on demand between processes.

Regarding abstractions, Verbs objects are a good candidate for granular per-resource tracking.
Even for devices that don't have HW limits for specific resources - the name-space itself could be controlled.

More abstract notions could be considered, such as device "resource pools".
Providers can control their actual HW resources more precisely, but it is not clear to the user what is actually controlled.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/ofvwg/attachments/20150915/a26f5e42/attachment.html>