...
| Type | Frequency | Systems (nodes) affected | How updates are applied | Impact on Users |
|---|---|---|---|---|
| Routine patching | 9 to 11 times a year | Border systems:
Internal systems:
| Updates are applied in a rolling fashion by power cycling all the nodes during the maintenance window, except for the compute nodes. Compute nodes are updated by first having the scheduler drain the nodes and then power cycle the node once it is idle. Upon restarting, the node is once again place into service. | Interactive jobs may be impacted, but batch jobs placed in non-preempting partitions should not be impacted. Momentary disruptions to storage system access via the border systems will occur as the border systems are rebooted for patching. |
| Core infrastructure updates | 1 to 3 times a year | Border systems:
Internal systems:
| All jobs regardless of progress are either halted and requeued or canceled if the job was submit as not being able to be requeued. Users access is denied and all servers are idled or powered down for the duration of the maintenance window. Once maintenance is complete, nodes are restarted and jobs that were requeued are released and begin to run once again. Jobs that do not save checkpoints, will have lost any progress they made prior to the downtime. | A maintenance reservation is put in for the day of the maintenance starting at 8 AM HST. Batch jobs that would run longer than the start of the maintenance reservation will be queue and wait until the reservation is done (the end of maintenance). before they would be allocated resources. Interactive jobs that request a time limit that would overlap the maintenance reservation will be denied by the scheduler. Jobs that do not overlap the maintenance reservation will be scheduled as normal. |
...