trying to calculate nhc.conf

Former-commit-id: 6c4d2f97

CICD/files/.gitignore

+nhc.conf
 ssh_known_hosts
 slurm.conf
 slurmdbd.conf

CICD/files/nhc.conf

-#######################################################################
-###
-### Filesystem checks
-###
-
- * || check_fs_used / 90%
-# * || check_fs_iused / 100%
-# * || check_fs_iused /glusterVolume 100%
-
-#not that useful at this stage as Nagios should be monitoring servers.
-# just check the file servers are happy 
-#* ||  check_fs_used  '/usr/local' 95%
-#* ||  check_fs_used '/home'       95%
-#* ||  check_fs_used '/projects'   95%
-#* ||  check_fs_used  '/scratch'   95%
-
-
-#* ||  check_fs_used '/'           100%
-
-# 
-#  New syntax:  check_fs_mount [ -0 ] [ -r ] [ -s src ] [ -t fstype ] [ -o options ] [ -O mount_options ] [ -e cmd ] [ -E cmd ] -f fs
-#
-
-# m3a0[16-20] nodes are currently disabled from mounting /usr/local. The check is disabled on all m3a because I didn't want to be more specific -- Chris 20180212
-m3a* ||  check_fs_mount_rw  -f  '/usr/local'
-m3c* ||  check_fs_mount_rw  -f  '/usr/local'
-m3d* ||  check_fs_mount_rw  -f  '/usr/local'
-m3e* ||  check_fs_mount_rw  -f  '/usr/local'
-m3f* ||  check_fs_mount_rw  -f  '/usr/local'
-m3g* ||  check_fs_mount_rw  -f  '/usr/local'
-m3h* ||  check_fs_mount_rw  -f  '/usr/local'
-m3i* ||  check_fs_mount_rw  -f  '/usr/local'
-m3m* ||  check_fs_mount_rw  -f  '/usr/local'
-m3p* ||  check_fs_mount_rw  -f  '/usr/local'
-dgx* ||  check_fs_mount_rw  -f  '/usr/local'
-* ||  check_fs_mount_rw  -f  '/home'
-* ||  check_fs_mount_rw  -f  '/projects'
-* ||  check_fs_mount_rw  -f  '/scratch'
-* ||  check_lustre_health
-
-#check numa config
-m3a* || check_numa
-m3c* || check_numa
-m3d* || check_numa
-m3e* || check_numa
-m3g* || check_numa
-m3h* || check_numa
-m3i* || check_numa
-m3m* || check_numa
-m3p* || check_numa
-* || check_SSSD
-* || check_user_lookup
-
-#######################################################################
-###
-### Hardware checks
-###
-# Don't check_hw_eth eth0 because most of our compute nodes have eth1 not eth0, but I won't guarantee this
-# This has to do with renaming eth0 to mlx0 (i.e. the mellanox device) but is senstive to device initialisation order I suspect
-# Chris Hines 20160907
-# * || check_hw_cpuinfo 1 1 1
-# * || check_hw_physmem 4048416kB 4048416kB 3%
- * || check_hw_swap 0kB 0kB 3%
- * || check_hw_eth lo
-!dgx* || check_hw_eth mlx0
-dgx* || check_hw_eth bond0.113
-dgx* || check_hw_eth bond0.114
- * || check_ibv_devinfo
-
-
-###
-### ECC not available on m3f K1
-m3c* || check_gpu_ecc
-m3e* || check_gpu_ecc
-
-# m3c* have 4 gpus
-# m3e* have 8 gpus
-# m3f* have 3 gpus
-# m3g* have 3 gpus
-# m3h* have 2 gpus
-# m3p* have 6 gpus
-# dgx* have 8 gpus
-m3c* || check_num_of_gpu 4
-m3e* || check_num_of_gpu 8
-m3f* || check_num_of_gpu 1
-m3g* || check_num_of_gpu 3
-m3h* || check_num_of_gpu 2
-m3p* || check_num_of_gpu 6
-dgx* || check_num_of_gpu 8
-#add more here
-
-m3c* || check_nvidia_device_existance
-m3e* || check_nvidia_device_existance
-m3f* || check_nvidia_device_existance
-m3g* || check_nvidia_device_existance
-m3h* || check_nvidia_device_existance
-m3p* || check_nvidia_device_existance
-dgx* || check_nvidia_device_existance
-# Kerri Wait 20170830 Add new check for xorg.conf file to ensure vglrun works
-m3c* || check_xorg_conf_file_existance
-m3f* || check_xorg_conf_file_existance
-m3g* || check_xorg_conf_file_existance
-m3h* || check_xorg_conf_file_existance
-m3p* || check_xorg_conf_file_existance
-
-#######################################################################
-###
-### Process checks
-###
-* || check_ps_service -S -u root sshd
-* || check_ps_service -S ntpd
-
-#Check for UVM for m3c and m3h
-m3c* || check_nvidia_uvm
-m3g* || check_nvidia_uvm
-m3h* || check_nvidia_uvm
-m3p* || check_nvidia_uvm
-#dgx* || check_nvidia_uvm
-

CICD/plays/computenodes.yml

@@ -72,6 +72,7 @@
   roles:
   - { role: slurm-common, tags: [ slurm, slurmbuild ] }
   - { role: slurm_config, tags: [ slurm_config, slurm ] }
+  - { role: calculateNhcConfig, tags: [ nhc, slurm ] }
   - { role: nhc, tags: [ nhc, slurm ] }
   - { role: slurm-start, start_slurmd: True, tags: [ slurm, slurm-start ] }
   - { role: vncserver, tags: [ other ] }