[libfabric-users] feature requests
Oucharek, Doug S
doug.s.oucharek at intel.com
Tue Jun 6 17:04:44 PDT 2017
I should make another clarification:
LNet only uses addresses to establish routing from the session manager. The address lists I keep referring to are lists of established endpoints (we use RC - reliable connections) and not the actual addresses. So, when the application (Lustre) refers to an address, we map that to a list of established endpoints (each could be to a different provider) and round robin sending messages over that set of endpoints. We have “credits” on each endpoint so we can do weighted round robin.
> On Jun 6, 2017, at 4:57 PM, Hefty, Sean <sean.hefty at intel.com> wrote:
>> We just recently added Multi-Rail support to LNet (Lustre networking
>> layer). We designate a “primary address” which is mapped to a list of
>> addresses for Multi-Rail. The application only uses the primary
>> address to send/receive from the peer.
> How did you handle the full address exchange? Does a failure to the primary address result in any issues?
More information about the Libfabric-users