What should the default minimum target be on Gadi after we migrate to Spack v1?

[harshula]

Once we finish the migration to Spack v1.x, we need to have a discussion about the default minimum target on Gadi. For example, should it be increased from x86_64 to x86_64_v3?

[harshula]

The sequence of decision making will be:

1. Choose the type of granularity (Issue Target granularity setting #50). i.e. generic Vs microarchitecture.
2. Then choose the minimum target for the type of granularity we have chose.

Please note, these decisions are for the defaults. They can be over-ridden via an environment's spack.yaml.

[manodeep]

v4 please on Gadi - both the normal and the normalsr queues support v4 instructions

[harshula]

Hi @manodeep , Does Gadi have one node that does not support x86_64_v4?

[harshula]

Hi @manodeep , I asked @rxy900 regarding Gadi's nodes during the Model Build meeting. He recommended using a minimum x86_64_v3.

[manodeep]

@harshula Only the (ex-Raijin) broadwell nodes do not support v4 (only v3); however, I don’t think any of our models run on the broadwell queues - do they? The only thing I can think of is that the ARE sessions run on broadwell

[harshula]

Hi @manodeep , Spack is now used to build packages that are run on login nodes. e.g. https://github.com/ACCESS-NRI/system-tools/

[manodeep]

@harshula login nodes use cascadelake cores - those support v4

[harshula]

My understanding is there are login nodes with old CPUs.

[manodeep]

Ahh okay - didn’t know that. Then v3 would be the conservative/safe choice.

We explicitly specify target as v4 in most of the SPR’s right?