You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
To make things cheaper, we should support interruptible instance type. This model is available on AWS as spot instances and Google as preemptible instance.
On AWS, an instance started as a spot instance will have a metadata service endpoint at http://169.254.169.254/latest/meta-data/spot/termination-time if it is marked to be interrupted. This should be polled for every 5 seconds.
This can then either be passed through the monitoring system so that the management node can respond or perhaps the compute node itself can trigger the POWER_DOWN state in directly Slurm with an scontrol call.
On the other side, we need a mechanism for specifying that the user wants spot pricing. Either this could be done on a cluster--wide level, a instance type-level or perhaps a job-level.
The text was updated successfully, but these errors were encountered:
To make things cheaper, we should support interruptible instance type. This model is available on AWS as spot instances and Google as preemptible instance.
On AWS, an instance started as a spot instance will have a metadata service endpoint at
http://169.254.169.254/latest/meta-data/spot/termination-time
if it is marked to be interrupted. This should be polled for every 5 seconds.This can then either be passed through the monitoring system so that the management node can respond or perhaps the compute node itself can trigger the
POWER_DOWN
state in directly Slurm with anscontrol
call.On the other side, we need a mechanism for specifying that the user wants spot pricing. Either this could be done on a cluster--wide level, a instance type-level or perhaps a job-level.
The text was updated successfully, but these errors were encountered: