[LU-20000] CPTs and tunables during import are wrong

• Type: Bug
• Resolution: Fixed
• Priority: Medium
• Fix Version/s: Lustre 2.18.0
• Labels: IPv6
• Severity: 3
• Assignee: Cyril Bordage
• Reporter: Cyril Bordage

Description

CPTs can be set for each NI, but because of a bug, all NIs in a net will have the same CPTs.

Also, tunables parsing is wrong. But for this issue, there are other problems...

In handle_yaml_config_ni header comment, we have:

net:
   - net type: <net>[<NUM>]
     local NI(s):
       - nid: <ip>@<net>[<NUM>]
         status: up
         interfaces:
             0: <intf_name>['['<expr>']']
             1: <intf_name>['['<expr>']']
      tunables:
             peer_timeout: <NUM>
             peer_credits: <NUM>
             peer_buffer_credits: <NUM>
             credits: <NUM>
      lnd tunables:
             peercredits_hiw: <NUM>
             map_on_demand: <NUM>
             concurrent_sends: <NUM>
             fmr_pool_size: <NUM>
             fmr_flush_trigger: <NUM>
             fmr_cache: <NUM>

However, with export and net show, it is this format:

net:
   - net type: <net>[<NUM>]
     local NI(s):
       - nid: <ip>@<net>[<NUM>]
         status: up
         interfaces:
             0: <intf_name>['['<expr>']']
             1: <intf_name>['['<expr>']']
         tunables:
                peer_timeout: <NUM>
                peer_credits: <NUM>
                peer_buffer_credits: <NUM>
                credits: <NUM>
         lnd tunables:
                peercredits_hiw: <NUM>
                map_on_demand: <NUM>
                concurrent_sends: <NUM>
                fmr_pool_size: <NUM>
                fmr_flush_trigger: <NUM>
                fmr_cache: <NUM>

And in the implementation point of view, it is:

net:
   - net type: <net>[<NUM>]
     local NI(s):
       - nid: <ip>@<net>[<NUM>]
         status: up
         interfaces:
             0: <intf_name>['['<expr>']']
             1: <intf_name>['['<expr>']']
         lnd tunables:
                peercredits_hiw: <NUM>
                map_on_demand: <NUM>
                concurrent_sends: <NUM>
                fmr_pool_size: <NUM>
                fmr_flush_trigger: <NUM>
                fmr_cache: <NUM>
     tunables:
            peer_timeout: <NUM>
            peer_credits: <NUM>
            peer_buffer_credits: <NUM>
            credits: <NUM>

Issue Links

• is related to: LU-20246 lnet may not have multiple cpu_npartitions
• is related to: LU-18417 Finish IPv6 support