VMs of instance type M are not schedulable due to lack of 2M pages

[DanielFroehlich]

Tried to schedule a VM of Instance Type "Memory Intensive - xlarge" (4 CPU, 32 Gi Memory=

Starting this VM results in ErrorUnschedulable.
Diagnostics shows:

0/13 nodes are available: 1 Insufficient memory, 10 Insufficient hugepages-2Mi, 3 node(s) had untolerated taint {node-role.kubernetes.io/master: }. preemption: 0/13 nodes are available: 10 No preemption victims found for incoming pod, 3 Preemption is not helpful for scheduling.

So it seems that M type VMs are pre-configure to consume 2Mi hugepages.
Looking at our nodes, those large pages are not configured:

% oc describe node/storm2
Capacity:
  ephemeral-storage:                      523506800Ki
  hugepages-1Gi:                          0
  hugepages-2Mi:                          0
  memory:                                 792486316Ki
  pods:                                   250

Simple workaround: go with U-type (general purpose VMs) - those work!

[DanielFroehlich]

@rbo:
Do we want configure hugepages? I think that could actually make sense and gain a bit performance. Maybe 25% - 50% of node memory as 2Mi page. Wdyt?

[rbo]

Sounds useful to configure hugepages. +1

[DanielFroehlich]

Fine, then I will configure huge pages according to the docs:
https://docs.redhat.com/en/documentation/openshift_container_platform/4.20/html/scalability_and_performance/what-huge-pages-do-and-how-they-are-consumed#configuring-huge-pages_huge-pages

Plan:
UCS nodes have 1TiB of memory and plenty of free memory available (requested is <<400 GiB on all three nodes) - lets configure 128G of 2M huge pages on those nodes. Then we should have plenty of capacity and can check how it is used.

I will configure this when I have time for it.

[DanielFroehlich]

Fixed