Guest CPU does not support 1Gb hugepages with explicit models

[Tapio Tallgren]

The CPU flag pdpe1gb indicates that the CPU model supports 1 GB hugepages - without it, the Linux operating system refuses to allocate 1 GB huge pages (and other things might go wrong if it did).

Not all Intel CPU models support 1 GB huge pages, so the qemu options -cpu Haswell and -cpu Broadwell give you a vCPU that does not have the pdpe1gb flag. This is the correct thing to do, since the VM might be running on a Haswell that does not have 1GB huge pages.

Problem is that Nova flavor extra specs with the libvirt driver for qemu/kvm only allow to define the CPU model, either an explicit model or "host". The host option means that all CPU flags in the host CPU are passed to the vCPU. However, the host option complicates VM migration since the CPU would change after migration.

In conclusion, there is no good way to specify a CPU model that would imply the pdpe1gb flag.

Huge pages are used eg with dpdk. They improve the performance of the VM mainly by reducing tlb size.

[Tapio Tallgren (tapio-tallgren)]

Similar discussion earlier:

https://review.openstack.org/#/c/168982/

[Sean Dague (sdague)]

I'm not sure how this boils down into an actionable bug. Can you specify the reproduce scenario and the expected vs. actual result?

[Tapio Tallgren (tapio-tallgren)]

I have a host that has a Haswell processor and supports 1GB hugepages. I want to launch a VM with CPU model Haswell to support migration to other servers that may have Haswell or later processors. When I launch the VM and look at the CPU flags, it does not have the pdpe1gb flag and therefore does not support 1GB hugepages.

I can specify cpu model as "host" but then I cannot migrate the VM between processor families.

[Launchpad Janitor (janitor)]

[Expired for OpenStack Compute (nova) because there has been no activity for 60 days.]

[Chris Friesen (cbf123)]

In recent versions of qemu the "Skylake-Server" cpu model has the flag, but any earlier Intel processor models do not.

[Chris Friesen (cbf123)]

There is a spec in review right now to address this: https://review.openstack.org/#/c/534384/

[Chris Friesen (cbf123)]

The code at https://review.openstack.org/#/c/534384/ has been merged, and should allow the operator to explicitly add the pdpe1gb flag.

Marking as fixed.