Shut down management services and power off the HPE Cray EX management Kubernetes cluster.
Understand the following concepts before powering off the management non-compute nodes (NCNs) for the Kubernetes cluster and storage:
mon
processes. If one of the storage nodes goes down, then Ceph can rebalance the data onto the remaining nodes and object storage daemons (OSDs) to regain full protection.kubelet process on each node retrieves information from the etcd
cluster about which pods must be running. If a node becomes unavailable for more than five minutes, then Kubernetes creates replacement pods on other management nodes.The sat bootsys command automates the shutdown of Ceph and the Kubernetes management cluster and performs these tasks:
etcd and which freezes the state of the Kubernetes cluster on each management node.kubelet on each management and worker node.containerd on each management and worker node.mon process.An authentication token is required to access the API gateway and to use the sat command. See the “SAT Authentication” section of the HPE Cray EX System Admin Toolkit (SAT) product stream
documentation (S-8031) for instructions on how to acquire a SAT authentication token.
To check the health and status of the management cluster before shutdown, see the “Platform Health Checks” section in Validate CSM Health.
Check the health and backup etcd clusters:
Determine which etcd clusters must be backed up and if they are healthy.
Backup etcd clusters.
Check the status of NCN no wipe settings.
Make sure that metal.no-wipe=1. If any management NCNs do not have that set, then review
Check and Set the metal.no-wipe Setting on NCNs before proceeding.
ncn-m001# /opt/cray/platform-utils/ncnGetXnames.sh
Example output:
+++++ Get NCN Xnames +++++
=== Can be executed on any worker or master ncn node. ===
=== Executing on ncn-m001, Thu Mar 18 20:58:04 UTC 2021 ===
=== NCN node xnames and metal.no-wipe status ===
=== metal.no-wipe=1, expected setting - the client ===
=== already has the right partitions and a bootable ROM. ===
=== Requires CLI to be initialized ===
=== NCN Master nodes: ncn-m001 ncn-m002 ncn-m003 ===
=== NCN Worker nodes: ncn-w001 ncn-w002 ncn-w003 ===
=== NCN Storage nodes: ncn-s001 ncn-s002 ncn-s003 ===
Thu Mar 18 20:58:06 UTC 2021
ncn-m001: x3000c0s1b0n0 - metal.no-wipe=1
ncn-m002: x3000c0s2b0n0 - metal.no-wipe=1
ncn-m003: x3000c0s3b0n0 - metal.no-wipe=1
ncn-w001: x3000c0s4b0n0 - metal.no-wipe=1
ncn-w002: x3000c0s5b0n0 - metal.no-wipe=1
ncn-w003: x3000c0s6b0n0 - metal.no-wipe=1
ncn-s001: x3000c0s7b0n0 - metal.no-wipe=1
ncn-s002: x3000c0s8b0n0 - metal.no-wipe=1
ncn-s003: x3000c0s9b0n0 - metal.no-wipe=1
Shut down platform services.
ncn-m001# sat bootsys shutdown --stage platform-services
Example output:
The following Non-compute Nodes (NCNs) will be included in this operation:
managers:
- ncn-m001
storage:
- ncn-s001
- ncn-s002
- ncn-s003
workers:
- ncn-w001
- ncn-w002
- ncn-w003
Are the above NCN groupings correct? [yes,no] yes
Executing step: Create etcd snapshot on all Kubernetes manager NCNs.
Executing step: Stop etcd on all Kubernetes manager NCNs.
Executing step: Stop and disable kubelet on all Kubernetes NCNs.
Executing step: Stop containers running under containerd on all Kubernetes NCNs.
WARNING: One or more "crictl stop" commands timed out on ncn-w003
WARNING: One or more "crictl stop" commands timed out on ncn-w002
ERROR: Failed to stop 1 container(s) on ncn-w003. Execute "crictl ps -q" on the host to view running containers.
ERROR: Failed to stop 2 container(s) on ncn-w002. Execute "crictl ps -q" on the host to view running containers.
WARNING: One or more "crictl stop" commands timed out on ncn-w001
ERROR: Failed to stop 4 container(s) on ncn-w001. Execute "crictl ps -q" on the host to view running containers.
WARNING: Non-fatal error in step "Stop containers running under containerd on all Kubernetes NCNs." of platform services stop: Failed to stop containers on the following NCN
(s): ncn-w001, ncn-w002, ncn-w003
Continue with platform services stop? [yes,no] no
Aborting.
In the preceding example, the commands to stop containers timed out on all the worker nodes and reported WARNING and ERROR messages.
A summary of the issue displays and prompts the user to continue or stop. Respond no stop the shutdown. Then review the containers running on the nodes.
ncn-m001# for ncn in ncn-w00{1,2,3}; do echo "${ncn}"; ssh "${ncn}" "crictl ps"; echo; done
Example output:
ncn-w001
CONTAINER IMAGE CREATED STATE NAME ATTEMPT POD ID
032d69162ad24 302d9780da639 54 minutes ago Running cray-dhcp-kea 0 e4d1c01818a5a
7ab8021279164 2ad3f16035f1f 3 hours ago Running log-forwarding 0 a5e89a366f5a3
ncn-w002
CONTAINER IMAGE CREATED STATE NAME ATTEMPT POD ID
1ca9d9fb81829 de444b360808f 4 hours ago Running cray-uas-mgr 0 902287a6d0393
ncn-w003
CONTAINER IMAGE CREATED STATE NAME ATTEMPT POD ID
Run the sat command again and enter yes at the prompt about the etcd snapshot not being created:
ncn-m001# sat bootsys shutdown --stage platform-services
Example output:
The following Non-compute Nodes (NCNs) will be included in this operation:
managers:
- ncn-m001
storage:
- ncn-s001
- ncn-s002
- ncn-s003
workers:
- ncn-w001
- ncn-w002
- ncn-w003
Are the above NCN groupings correct? [yes,no] yes
Executing step: Create etcd snapshot on all Kubernetes manager NCNs.
WARNING: Failed to create etcd snapshot on ncn-m001: The etcd service is not active on ncn-m001 so a snapshot cannot be created.
WARNING: Failed to create etcd snapshot on ncn-m002: The etcd service is not active on ncn-m002 so a snapshot cannot be created.
WARNING: Failed to create etcd snapshot on ncn-m003: The etcd service is not active on ncn-m003 so a snapshot cannot be created.
WARNING: Non-fatal error in step "Create etcd snapshot on all Kubernetes manager NCNs." of platform services stop: Failed to create etcd snapshot on hosts: ncn-m001, ncn-m00
2, ncn-m003
Continue with platform services stop? [yes,no] yes
Continuing.
Executing step: Stop etcd on all Kubernetes manager NCNs.
Executing step: Stop and disable kubelet on all Kubernetes NCNs.
Executing step: Stop containers running under containerd on all Kubernetes NCNs.
Executing step: Stop containerd on all Kubernetes NCNs.
Executing step: Check health of Ceph cluster and freeze state.
If the process continues to report errors due to Failed to stop containers, then iterate on the above step. Each iteration should reduce the number of containers running. If necessary,
containers can be manually stopped using crictl stop CONTAINER. If containers are stopped manually, then re-run the above procedure to complete any final steps in the process.
Shut down and power off all management NCNs except ncn-m001.
ncn-m001# sat bootsys shutdown --stage ncn-power
Example output:
Proceed with shutdown of other management NCNs? [yes,no] yes
Proceeding with shutdown of other management NCNs.
IPMI username: root
IPMI password:
The following Non-compute Nodes (NCNs) will be included in this operation:
managers:
- ncn-m002
- ncn-m003
storage:
- ncn-s001
- ncn-s002
- ncn-s003
workers:
- ncn-w001
- ncn-w002
- ncn-w003
The following Non-compute Nodes (NCNs) will be excluded from this operation:
managers:
- ncn-m001
storage: []
workers: []
Are the above NCN groupings and exclusions correct? [yes,no] yes
Monitor the consoles for each NCN.
Use tail to monitor the log files in /var/log/cray/console_logs for each NCN.
Alternately attach to the screen session (screen sessions real time, but not saved):
ncn-m001# screen -ls
Example output:
There are screens on:
26745.SAT-console-ncn-m003-mgmt (Detached)
26706.SAT-console-ncn-m002-mgmt (Detached)
26666.SAT-console-ncn-s003-mgmt (Detached)
26627.SAT-console-ncn-s002-mgmt (Detached)
26589.SAT-console-ncn-s001-mgmt (Detached)
26552.SAT-console-ncn-w003-mgmt (Detached)
26514.SAT-console-ncn-w002-mgmt (Detached)
26444.SAT-console-ncn-w001-mgmt (Detached)
ncn-m001# screen -x 26745.SAT-console-ncn-m003-mgmt
Check the power off status of management nodes.
NOTE:
read -sis used to read the password in order to prevent it from being echoed to the screen or preserved in the shell history.
ncn-m001# USERNAME=root
ncn-m001# read -r -s -p "NCN BMC ${USERNAME} password: " IPMI_PASSWORD
ncn-m001# export IPMI_PASSWORD
ncn-m001# for ncn in ncn-m00{2,3} ncn-w00{1,2,3} ncn-s00{1,2,3}; do
echo -n "${ncn}: "
ipmitool -U "${USERNAME}" -H "${ncn}-mgmt" -E -I lanplus chassis power status
done
From a remote system, activate the serial console for ncn-m001.
external# ipmitool -I lanplus -U "${USERNAME}" -E -H NCN-M001_BMC_HOSTNAME sol activate
From the serial console of ncn-m001, shut down Linux.
ncn-m001# shutdown -h now
Wait until the console indicates that the node has shut down.
From a remote system that has access to the management plane, power off ncn-m001.
external# ipmitool -I lanplus -U "${USERNAME}" -E -H NCN-M001_BMC_HOSTNAME chassis power status
external# ipmitool -I lanplus -U "${USERNAME}" -E -H NCN-M001_BMC_HOSTNAME chassis power off
external# ipmitool -I lanplus -U "${USERNAME}" -E -H NCN-M001_BMC_HOSTNAME chassis power status
CAUTION: The modular coolant distribution unit (MDCU) in a liquid-cooled HPE Cray EX2000 cabinet (also referred to as a Hill or TDS cabinet) typically receives power from its management cabinet PDUs. If the system includes an EX2000 cabinet, then do not power off the management cabinet PDUs. Powering off the MDCU will cause an emergency power off (EPO) of the cabinet and may result in data loss or equipment damage.
(Optional) If a liquid-cooled EX2000 cabinet is not receiving MCDU power from this management cabinet, then power off the PDU circuit breakers or disconnect the PDUs from facility power and follow lock out/tag out procedures for the site.
Return to System Power Off Procedures and continue with next step.