[LU-11144] Dynamic Discovery is not triggered for router peers

• Status: Closed
• Project: Lustre
• Type: Bug
• Priority: Minor
• Reporter: Chris Horn
• Assignee: Amir Shehata (Inactive)
• Resolution: Fixed
• Labels: LNet
• Severity: 3

Description

It appears to me that dynamic discovery will not be triggered for LNet routers under normal operating conditions (i.e. without an explicit lnetctl discover of the routers).

The discovery logic is early in lnet_select_pathway() and based on the peer associated with the dst_nid. The only LNetGet where dst_nid is a router is going to be the router checker traffic. But that traffic goes on the LNET_RESERVED_PORTAL, so it will never trigger discovery.
Seems like we need to check for discovery of the router peer after selecting it later in lnet_select_pathway().

Issue Links

• is related to: LU-11297 Align LNet routing with Multi-Rail an...
• is related to: LU-11143 Multi-Rail/Dynamic Discovery break LN...

Comments

[Chris Horn]

I think what I was missing with router peers and dynamic discovery is the DD push
router should queue other peers for discovery based on where it is forwarding traffic, and the discovery thread should push the router's MR info to those peers as part of that. Right?

So while the description in this ticket is, I believe, accurate it's probably by design and not necessarily a problem.

[Amir Shehata (Inactive)]

both LU-11143 and LU-11144 are related.

I address them here:

https://wiki.whamcloud.com/display/LNet/Routing+and+MR+integration

Might be a good idea to use that link for feedback on the proposals

[Cory Spitz]

ashehata, ready to resolve this and LU-11143?

[Amir Shehata (Inactive)]

I believe this issue has been resolved in the new routing code.

[Cory Spitz]

ashehata, will you be resolving this issue then? Can you point at a specific commit or LU that resolved it?

[Chris Horn]

Resolved with MR routing feature