Perhaps the PrologSlurmctld could be used. See http://www.schedmd.com/slurmdocs/prolog_epilog.html
Quoting Twilight Zoners <[email protected]>: > Hi all, > > > I've setup and configured slurm 2.5.0 with 1 slurm controller and 9 > nodes (total 10 computers in a cluster, all RHEL6 64 bit) . > Things i've set up: > - using munge > - all computers are joined to windows Active Directory domain, so > users can use their Active directory credentials to login to any > RHEL6 computer in the cluster. > - samba + winbind > - NFS share > > I have successfully set all the RHEL computers to use the same UID > and GID for each users, using RID in samba config: > idmap uid = 17000000-33554431 > idmap gid = 17000000-33554431 > idmap backend = rid:domain=100000-100000000 > > So when a user log in to node01, node02, etc, the user will always > get the same UID and GID, thus, the user can run the srun command. > > > However, i am facing a challenge here. > If everytime a new user (never logged in to any computer in the > cluster) try to login to the controller and run a srun command, for > example "srun -n30 -c1 pwd", the user will get this error: > > > srun: error: Task launch for 430.0 failed on node node01: User not > found on host > srun: error: Task launch for 430.0 failed on node node02: User not > found on host > > > I believe this is due to, the user account is not yet "created" on > the node01 and node02. > > So the "temporary" solution to this is, i ask the user to login to > all the slurm nodes in order to force all the nodes to create the > user's profile/account. > > This cluster will grow in future and we will add more slurm nodes. > I cannot imagine if i need to add 100 nodes, what the users will > react if they need to login to additional 100 computers, just to > create their accounts on each node. > > I tried to implement NIS, but i think NIS only sync the local linux > users account. In my case, the domain user accounts are not saved on > the local linux account, but in winbind database > (/var/lib64/samba/*.tdb). > > > I believe i am not the only one using this approach (using windows > AD for authentication) for slurm. > > Can anyone give me a clue how to solve this challenge? > > > Thanks in advance....
