Ceci est une ancienne révision du document !
Explanations will be long, take a coffee, a tea, or quit ;o)
Last week, we experienced a threshold effect (again) while adding new nodes to E5 cluster. It lead us to reboot a big part of debian nodes. Doing this open a Murphy's box (you know the law ? It's the same, with a ribbon on it).
- an electrical problem appear where we add the new nodes, on the same
distribution unit of a block of x5570, including their scratch servers. -> unexpected reboots while trying to figure out what's going on.
- new E5 nodes cannot connect to “old” infiniband switchs (firmware problem)
- new E5 infiniband switch cannot connect to “old” infiniband switchs (firmware
problem) -> old switchs need a new firmware, and reboot -> a special machine, with a very early OS, with experimental libs is needed to upgrade switchs
- old E5 nodes bios is incompatible with new infiniband card firmware…
- with all theses, scratch has become incoherent, checkfs needed…
And now ? Upgrades are (mostly) OK, scratch has been checked and is OK. Some E5 nodes are not OK, and have been pushed away from queuing system.