This procedure prepares the system to remove power from all system cabinets. Be sure the system is healthy and ready to be shut down and powered off.
The sat bootsys shutdown and sat bootsys boot commands are used to shut down the system.
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.
Obtain the user ID and passwords for system components:
Obtain user ID and passwords for all the system management network switches.
If necessary, obtain the user ID and password for the ClusterStor primary management node. For example, cls01053n00.
If the Slingshot network includes edge switches, obtain the user ID and password for these switches.
Use sat auth to authenticate to the API gateway within SAT.
If SAT has already been authenticated to the API gateway, this step may be skipped.
See the “SAT Authentication” section in the HPE Cray EX System Admin Toolkit (SAT) product stream documentation (S-8031) for instructions on how to acquire a SAT authentication token.
Ensure /root/.bashrc has proper handling of kubectl commands on all master and worker nodes.
Important: During the process of shutting down the system, there will be a point when kubelet will be stopped on all the master and worker
nodes. Once kubelet has been stopped, any kubectl command on any master or worker node may not work as expected and may have a long timeout before
failing.
This issue can cause a slowdown for these sat commands which ssh from the sat pod to ncn-m001 and the
other nodes because the ssh will execute commands from /root/.bashrc.
Commands affected during the power down
sat bootsys shutdown --stage platform-servicessat bootsys shutdown --stage ncn-powerCommands affected during the power up
sat bootsys boot --stage ncn-powersat bootsys boot --stage platform-servicesHere is a sample command in /root/.bashrc which sets an environment variable using the output from kubectl which has the problem.
export DOMAIN=$(kubectl get secret site-init -n loftsman -o jsonpath='{.data.customizations\.yaml}'|base64 -d | grep "external:")
This shows one way to correct that sample command so the environment variable will be set when kubelet is available and will skip setting the variable when kubelet is not available.
if systemctl is-active -q kubelet ; then
export DOMAIN=$(kubectl get secret site-init -n loftsman -o jsonpath='{.data.customizations\.yaml}'|base64 -d | grep "external:")
fi
Ensure /root/.ssh/known_hosts does not have ssh stale host key entries for any of the management nodes.
Important: Many of the sat commands use ssh from a sat Kubernetes pod to execute commands on the management nodes. This sat pod
uses the paramiko Python library for ssh and it will access /root/.ssh/known_hosts. If /root/.ssh/config or /etc/ssh/config has
been configured to set UserKnownHostsfile to /dev/null or some other file and there are ssh host key mismatches in /root/.ssh/known_hosts, then
when a sat command tries to use ssh with paramiko, it will fail even though an interactive ssh command by the root user might succeed.
For example, the sat bootsys shutdown --stage platform-services command would show this type of error and fail.
INFO: Executing step: Stop and disable kubelet on all Kubernetes NCNs.
ERROR: Host key for server 'ncn-w003' does not match: got 'AAAAE2VjZHNhLXNoYTItbmlzdHAyNTYAAAAIbmlzdHAyNTYAAABBBCGNgIUTU7+o/+c5bD84u7/1S3xNNOd5+c/0l4vpVEehWGrjuC6IRC/KAImozzznXHhdBL7yQF2Dnh3FHGQDyko=', expected 'AAAAE2VjZHNhLXNoYTItbmlzdHAyNTYAAAAIbmlzdHAyNTYAAABBBL9bwo5gmW/vX3kUQyXIDgJa4EVtCYDyntmNt43BHTM7YKn6yFe1dV59Ervi13V20OxdVECxg2hTyeTueVKvwj4='
ERROR: Fatal error in step "Stop and disable kubelet on all Kubernetes NCNs." of platform services stop: Failed to ensure kubelet is inactive and disabled on all hosts.
To prevent this issue from happening, remove stale ssh host keys from /root/.ssh/known_hosts before running the sat command.
Check certificate expiration deadlines to ensure that a certificate won’t expire while the system is powered off.
(ncn-mw#) Check the expiration date of the Spire Intermediate CA Certificate.
kubectl get secret -n spire spire.spire.ca-tls -o json | jq -r '.data."tls.crt" | @base64d' | openssl x509 -noout -enddate
Example output:
notAfter=Dec 17 00:00:24 2024 GMT
If the certificate will expire while the system is powered off, replace it before powering off the system. See Replace the Spire Intermediate CA Certificate.
(ncn-m#) Check the Kubernetes and Bare Metal etcd certificates from a master node.
Check certificate expiration deadlines for Kubernetes and its bare-metal etcd cluster.
kubeadm certs check-expiration --config /etc/kubernetes/kubeadmcfg.yaml
Example output:
WARNING: kubeadm cannot validate component configs for API groups [kubelet.config.k8s.io kubeproxy.config.k8s.io]
CERTIFICATE EXPIRES RESIDUAL TIME CERTIFICATE AUTHORITY EXTERNALLY MANAGED
admin.conf Sep 24, 2021 15:21 UTC 14d ca no
apiserver Sep 24, 2021 15:21 UTC 14d ca no
apiserver-etcd-client Sep 24, 2021 15:20 UTC 14d etcd-ca no
apiserver-kubelet-client Sep 24, 2021 15:21 UTC 14d ca no
controller-manager.conf Sep 24, 2021 15:21 UTC 14d ca no
etcd-healthcheck-client Sep 24, 2021 15:19 UTC 14d etcd-ca no
etcd-peer Sep 24, 2021 15:19 UTC 14d etcd-ca no
etcd-server Sep 24, 2021 15:19 UTC 14d etcd-ca no
front-proxy-client Sep 24, 2021 15:21 UTC 14d front-proxy-ca no
scheduler.conf Sep 24, 2021 15:21 UTC 14d ca no
CERTIFICATE AUTHORITY EXPIRES RESIDUAL TIME EXTERNALLY MANAGED
ca Sep 02, 2030 15:21 UTC 8y no
etcd-ca Sep 02, 2030 15:19 UTC 8y no
front-proxy-ca Sep 02, 2030 15:21 UTC 8y no
Depending on which certificates will expire, one of these procedures could be used for the renewal. The first procedure will renew all certificates, but that may be more than needs to be renewed.
(ncn-mw#) Check for a recent backup of Nexus data.
Note: Doing the Nexus backup may take multiple hours with Nexus being unavailable for the entire time.
Check whether an export PVC called nexus-bak exists and is recent.
kubectl get pvc -n nexus
Example output:
NAME STATUS VOLUME CAPACITY ACCESS MODES STORAGECLASS AGE
nexus-bak Bound pvc-09b6efe6-18e3-4681-8103-53590ad49d04 1000Gi RWO k8s-block-replicated 293d
nexus-data Bound pvc-bce9db69-d1a6-491d-89fc-d458c92f2895 1000Gi RWX ceph-cephfs-external 518d
This output shows that the nexus-bak PVC was created 293 days ago.
If there is no nexus-bak PVC, then use this Nexus export procedure to create one. This procedure does check that
there is enough space available for the copy of the nexus-data PVC and provides guidance on how to clean up space if
necessary or reduce the size of the existing nexus-data PVC.
See Nexus Export.
If there is an existing nexus-bak PVC, but it is too old or the age is not recent enough to include the most recent
software update or otherwise not considered valid, then use the Nexus cleanup procedure before the export procedure.
See Nexus Cleanup, then see
Nexus Export.
(ncn-mw#) Determine which Boot Orchestration Service (BOS) templates to use to shut down compute nodes and UANs.
There will be separate session templates for UANs and computes nodes.
List all the BOS session templates.
If it is unclear what BOS session template is in use, proceed to the next substep.
cray bos sessiontemplates list --format json | jq -r '.[].name' | sort
Find the BOS session templates used most recently to boot nodes.
sat status --filter role!=management --fields xname,role,subrole,"most recent session template"
Example output:
+----------------+-------------+-----------+------------------------------+
| xname | Role | SubRole | Most Recent Session Template |
+----------------+-------------+-----------+------------------------------+
| x3209c0s13b0n0 | Application | UAN | uan-23.7.0 |
| x3209c0s15b0n0 | Application | UAN | uan-23.7.0 |
| x3209c0s17b0n0 | Application | UAN | uan-23.7.0 |
| x3209c0s19b0n0 | Application | UAN | uan-23.7.0 |
| x3209c0s22b0n0 | Application | Gateway | MISSING |
| x3209c0s23b0n0 | Application | Gateway | MISSING |
| x9002c1s0b0n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s0b0n1 | Compute | Compute | compute-23.7.0 |
| x9002c1s0b1n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s0b1n1 | Compute | Compute | compute-23.7.0 |
| x9002c1s1b0n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s1b0n1 | Compute | Compute | compute-23.7.0 |
| x9002c1s1b1n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s1b1n1 | Compute | Compute | compute-23.7.0 |
| x9002c1s2b0n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s2b0n1 | Compute | Compute | compute-23.7.0 |
| x9002c1s2b1n0 | Compute | Compute | compute-23.7.0 |
| x9002c1s2b1n1 | Compute | Compute | compute-23.7.0 |
+----------------+-------------+-----------+------------------------------+
NOTE When the Most Recent Session Template shows MISSING, it means the BOS session information was removed.
Old BOS sessions are cleaned up based on the numbers of days in cleanup_completed_session_ttl. The default value is seven days.
Check the current setting for cleanup_completed_session_ttl.
cray bos options list | grep cleanup_completed_session_ttl
Determine the list of xnames associated with the desired boot session template.
cray bos sessiontemplates describe SESSION_TEMPLATE_NAME --format json | jq '.boot_sets | map({node_list, node_roles_groups, node_groups})'
Example outputs:
[
{
"node_list": [
"x3000c0s19b1n0",
"x3000c0s19b2n0",
"x3000c0s19b3n0",
"x3000c0s19b4n0"
],
"node_roles_groups": null,
"node_groups": null
}
]
[
{
"node_list": null,
"node_roles_groups": [
"Compute"
],
"node_groups": null
}
]
(ncn-mw#) Use SAT to capture state of the system before the shutdown.
sat bootsys shutdown --stage capture-state
(ncn-mw#) Optional system health checks.
Use the System Diagnostic Utility (SDU) to capture current state of system before the shutdown.
Important: SDU may take about 45 minutes to run on a small system (longer for large systems).
sdu --scenario triage --start_time '-4 hours' \
--reason "saving state before powerdown"
Capture the state of all nodes.
sat status | tee -a sat.status
Capture the list of disabled nodes.
sat status --filter Enabled=false | tee -a sat.status.disabled
Capture the list of nodes that are off.
sat status --filter State=Off | tee -a sat.status.off
Capture the state of nodes in the workload manager.
For example, if the system uses Slurm:
ssh uan01 sinfo | tee -a uan01.sinfo
ssh uan01 sinfo --list-reasons | tee -a sinfo.reasons
For example, if the system uses PBS Pro:
ssh uan01 pbsnodes -aS | tee -a pbsnodes.aS
Check Ceph status.
ceph -s | tee -a ceph.status
Check Kubernetes pod status for all pods.
kubectl get pods -o wide -A | tee -a k8s.pods
Additional Kubernetes status check examples:
kubectl get pods -o wide -A | egrep "CrashLoopBackOff" | tee -a k8s.pods.CLBO
kubectl get pods -o wide -A | egrep "ContainerCreating" | tee -a k8s.pods.CC
kubectl get pods -o wide -A | egrep -v "Run|Completed" | tee -a k8s.pods.errors
Check HSN status.
Run fmn-show-status in the slingshot-fabric-manager pod and save the output to a file.
kubectl exec -it -n services \
"$(kubectl get pod -l app.kubernetes.io/name=slingshot-fabric-manager \
-n services --no-headers | head -1 | awk '{print $1}')" \
-c slingshot-fabric-manager -- fmn-show-status --details \
| tee -a fmn-show-status-details.txt
Check management switches to verify they are reachable.
Note: The switch host names depend on the system configuration.
(ncn-mw#) Use CANU to confirm that all switches are reachable. Reachable switches have their
version information populated in the network version report.
Provide the password for the admin username on the management network switches.
canu report network version
Example output:
Password:
SWITCH CANU VERSION CSM VERSION
sw-spine-001 1.6.20 1.4
sw-spine-002 1.6.20 1.4
sw-leaf-bmc-001 1.6.20 1.4
sw-leaf-bmc-002 1.6.20 1.4
sw-cdu-001 1.6.20 1.4
sw-cdu-002 1.6.20 1.4
(Optional) (ncn-mw#) If CANU is not available, look in /etc/hosts for the management network
switches on this system. The names of all spine switches, leaf switches, leaf BMC
switches, and CDU switches need to be used in the next step.
grep sw- /etc/hosts
Example output:
10.254.0.2 sw-spine-001
10.254.0.3 sw-spine-002
10.254.0.4 sw-leaf-bmc-001
10.254.0.5 sw-leaf-bmc-002
10.100.0.2 sw-cdu-001
10.100.0.3 sw-cdu-002
(ncn-mw#) Ping the switches obtained in the previous step to determine if they are reachable.
for switch in $(awk '{print $2}' /etc/hosts | grep 'sw-'); do
echo -n "switch ${switch} is "
ping -c 1 -W 10 $switch > /dev/null && echo "up" || echo "not up"
done | tee -a switches
(ncn-mw#) Check Lustre server health. See Lustre documentation for other health commands to run.
ssh admin@cls01234n00.us.cray.com
cscli csinfo
cscli show_nodes
cscli fs_info
From a node which has the Lustre file system mounted.
lfs check servers
lfs df
(ncn-mw#) Check for running sessions.
sat bootsys shutdown --stage session-checks 2>&1 | tee -a sat.session-checks
Example output:
Checking for active BOS sessions.
Found no active BOS sessions.
Checking for active CFS sessions.
Found no active CFS sessions.
Checking for active CRUS upgrades.
Found no active CRUS upgrades.
Checking for active FAS actions.
Found no active FAS actions.
Checking for active NMD dumps.
Found no active NMD dumps.
Checking for active SDU sessions.
Found no active SDU sessions.
No active sessions exist. It is safe to proceed with the shutdown procedure.
If active sessions are running, either wait for them to complete or cancel the session. See the following step.
NOTE If the Compute Rolling Upgrade Service (CRUS) is not being used to manage compute rolling upgrades, then this error can be ignored.
If cray-crus pods have failed to initialize, the GET request will fail. A system using PBS Pro or Slurm in which the workload manager scheduler
is not running in a pod on a worker node will fail to initialize. CRUS will be removed in a future release.
INFO: Checking for active CRUS upgrades.
ERROR: Unable to get active CRUS upgrades: GET request to URL 'https://api-gw-service-nmn.local/apis/crus/session' failed with status code 503: Service Unavailable
NOTE If the System Diagnostic Utility (SDU) has not been configured on master nodes, message like this will appear for the master nodes
which are not configured for SDU. If the warning appears for all master nodes, then to enable this after the system has been powered up again,
see the Configure section of the HPE Cray EX with CSM System Diagnostic Utility (SDU) Installation Guide to configure SDU and the optional RDA.
WARNING: The cray-sdu-rda container is not running on ncn-m001.
WARNING: The cray-sdu-rda container is not running on ncn-m002.
WARNING: The cray-sdu-rda container is not running on ncn-m003.
(ncn-mw#) Cancel the running BOS sessions.
Identify the BOS Sessions to delete.
cray bos sessions list --format json
Delete each running BOS session.
cray bos sessions delete <session ID>
Example:
cray bos sessions delete 0216d2d9-b2bc-41b0-960d-165d2af7a742
Coordinate with the site system administrators to prevent new sessions from starting in the services listed.
There is no method to prevent new sessions from being created as long as the service APIs are accessible on the API gateway.
Notify users and operations staff about the upcoming full system power off.
The notification method will vary by system, but might be email, messaging applications, /etc/motd on UANs, wall commands on UANs, etc.
Follow the vendor workload manager documentation to drain processes running on compute nodes.
For Slurm, see the scontrol man page.
Below are examples of how to drain nodes using slurm. The list of nodes can be copy/pasted from the sinfo command for nodes in an idle state:
scontrol update NodeName=nid[001001-001003,001005] State=DRAIN Reason="Shutdown"
scontrol update NodeName=ALL State=DRAIN Reason="Shutdown"
For PBS Professional, see the qstat and qmgr man pages.
Below is an example to list the available queues, disable a specific queue named workq, and check
that the queue has been disabled:
qstat -q
qmgr -c 'set queue workq enabled = False'
qmgr -c 'list queue workq enabled'
Each system might have many different queue names. There is no default queue name.
Return to System Power Off Procedures and continue with next step.