Make minCpuPlatform a variable to enable launching Cascade Lake instances on Google Compute Platform

[christopheredsall]

I haven't yet been able to successfully launch a "Cascade Lake" instance. The result coming back from GCP is

[provisioner@mgmt ~]$ sudo tail -f /var/log/slurm/elastic.log 
2019-10-20 21:54:28,338 startnode  INFO     Starting c2-standard-4-0001 in citc-test-ce europe-west4-c
2019-10-20 21:54:30,947 startnode  ERROR     problem launching instance: <HttpError 400 when requesting https://compute.googleapis.com/compute/v1/projects/citc-test-ce/zones/europe-west4-c/instances?alt=json returned "The selected machine type 'c2-standard-4' has a required CPU platform of 'cascadelake'. The minimum CPU platform must match this, but was 'skylake'.">

[christopheredsall]

Google documentation

Specifying a Minimum CPU Platform for VM Instances

[christopheredsall]

This is currently set in create_node_config in citc.py

https://github.com/ACRC/slurm-ansible-playbook/blob/b2b687434cbae5a5855399637db130c107d03cac/roles/slurm/files/citc_gcp.py#L105

[christopheredsall]

As a quick YOLO test changing the hard coded value

--- /opt/cloud_sdk/lib/python3.6/site-packages/citc_cloud.py.orig	2019-10-20 22:18:43.864878497 +0000
+++ /opt/cloud_sdk/lib/python3.6/site-packages/citc_cloud.py	2019-10-20 22:19:12.751080593 +0000
@@ -102,7 +102,7 @@
                 ]
             }
         ],
-        'minCpuPlatform': 'Intel Skylake',
+        'minCpuPlatform': 'Intel Cascade Lake',
         'metadata': {
             'items': [{
                 'key': 'startup-script',

works, the node provisions and boots:

2019-10-20 22:20:06,577 startnode  INFO     Starting c2-standard-4-0001 in citc-test-ce europe-west4-c
2019-10-20 22:20:09,602 startnode  INFO     c2-standard-4-0001:  No VNIC attachment yet. Waiting...
2019-10-20 22:20:15,237 startnode  INFO       Private IP 10.1.0.3
2019-10-20 22:20:15,246 startnode  INFO      Started c2-standard-4-0001

[christopheredsall]

Another approach that works is to just remove that attribute:

--- /opt/cloud_sdk/lib64/python3.6/site-packages/citc_cloud.py.orig	2019-10-21 13:45:37.285506506 +0000
+++ /opt/cloud_sdk/lib64/python3.6/site-packages/citc_cloud.py	2019-10-21 15:11:17.601853276 +0000
@@ -102,7 +102,6 @@
                 ]
             }
         ],
-        'minCpuPlatform': 'Intel Skylake',
         'metadata': {
             'items': [{
                 'key': 'startup-script',

[christopheredsall]

What would be the downsides to not specifying minCpuPlatform?

[geoffnewell]

Extend citc_gcp.py: create_node_config() to cater for Slurm Features to allow for specifying minCPUPlatform in the SBatch Slurm job config. For example:

#! /bin/bash

#SBATCH --job-name=test
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=1
#SBATCH --time=10:00
#SBATCH --constraint="minCPUPlatform=Intel Cascade Lake"

srun -l hostname