-
Notifications
You must be signed in to change notification settings - Fork 47
Description
With google-batch as provider it seems that certain machine types such as C3 (but probably including others such as M3 based on their disk requirements) that require pd-ssd rather than pd-standard are unable to launch despite specifying --disk-type "pd-ssd" in dsub invocation, because the bootDisk type is not also changed from pd-standard. For example, the same dsub invocation fails at Job state is set from SCHEDULED_PENDING_FAILED to FAILED if
--machine-type "n1-standard-8"
is switched to
--machine-type "c3-standard-8"
--disk-type "pd-ssd"
where gcloud batch jobs describe contains the following:
allocationPolicy:
instances:
- policy:
bootDisk:
sizeGb: '50'
type: pd-standard
disks:
- deviceName: datadisk
newDisk:
sizeGb: '200'
type: pd-ssd
machineType: c3-standard-8
provisioningModel: SPOT
These dsub flags work fine on google-cls-v2 to request C3 machines, so presumably the bootDisk type is being set correctly (or is propagated from datadisk type).
This is related to this issue (and will probably need to also be resolved for N4 and newer machine types to allow for hyperdisk-balanced to also be propagated to bootDisk type after passing through as --disk-type "hyperdisk-balanced" ), but also impacts machine types that in theory should be supported by existing --disk-type options.