Clarify memory rounding of HMC

[Andreas Scheuring]

The HMC implements memory rounding:

Docs form the "initial-memory" [1] attribute of a partition:

"[...]
Note: If the value input by the user does not fall on an
increment boundary, it is rounded up to the closest
increment boundary."

Also consider [2], page 100.

We need to clarify how this affects u
- can we end up with consuming more memory than specified in our max_memory property
- an there be a missmatch between used memory and the one specified in the flavor and quota?
- ...

[1] https://www-01.ibm.com/support/docview.wss?uid=isg2cb468b15654ca89b85257f7200746c16
[2] http://www-01.ibm.com/support/docview.wss?uid=isg22dec9fb95bdff17385257dc4007a164f

[sreeteja (sreeteja)]

I think this is duplicate of https://bugs.launchpad.net/nova-dpm/+bug/1689234

[Andreas Scheuring (andreas-scheuring)]

no, this one is about the dpm partition memory properties, while the bug you are referring to is about the max_memory openstack dpm property

[sreeteja (sreeteja)]

Yes I tried a scenario in which I gave the flavor as 1000MB ram with the ram_allocation_ratio as 1 and due to the HMC rounding property it rounded off this value to 1024MB and launched an instance...but the in the dashboard under the hypervisor stats it is showing as ram-used 1000MB and also ram_filter which is a scheduler filter com pares this condition
https://github.com/openstack/nova/blob/4f91ed3a547965ed96a22520edcfb783e7936e95/nova/scheduler/filters/ram_filter.py#L39
even this parameters like memory_mb are all taken from the flavor so...there are two problems in this

1.With the existing code if we allocate the max_memory as 2000MB to be used by a specific host due to HMC rounding property it consumes 2048MB (because even if flavor is provided as 1000 it rounds to 1024 only)...OVER CONSUMPTION OF RAM

2. This also shows wrong hypervisor stats and mislead scheduler ram_filter...

[Andreas Scheuring (andreas-scheuring)]

nice elaboration!

[Andreas Scheuring (andreas-scheuring)]

I think the solution is to deny instance creation if the flavor has a ram size that requires rounding (is not a multiple of 256) with an appropriate error message in the log. We should also add this piece of information to the docs.

The alternative would be to live with the fact and just print warnings...

Andy, Marco, what do you think?

[Sreeram Vancheeswaran (sreeram-vancheeswaran)]

I'd vote for the second option, i.e, to live with this and print warning messages, as it is rounding off to the nearest multiple of 256, and according to me this is ok to live with.

[Marco Pavone (pavone)]

we should not silently accept the rounding taking place.
My preferred way is to check for the actual memory consumption after the create partition has been triggered, and report this back to scheduler (i.e. update in nova database). However, I didn't check if this is possible from both current zhmcclient and nova(-spawn) implementations...

[Andreas Scheuring (andreas-scheuring)]

@Marco: But that doesn't solve the issue 2 issues sreeteja laid out.

#1 More memory than the max_memory parameter might get consumed
#2 This is one of the reason for #1. The scheduler makes assumptions about the requested memory along the flavor - he does not know anything about the memory rounding taking place).

We're already reporting back the right values IMO [1]. The problem is issue #1 and #2.

[1] https://github.com/openstack/nova-dpm/blob/cb0074b90e432ffcb3c971576cfda65d84d855dc/nova_dpm/virt/dpm/host.py#L109,L115

[sreeteja (sreeteja)]

Yes as andreas told we are reporting the right value by [1] https://github.com/openstack/nova-dpm/blob/cb0074b90e432ffcb3c971576cfda65d84d855dc/nova_dpm/virt/dpm/host.py#L109,L115 but the hypervisor metrics are shown wrong as they are looking at the memory value consumed as the flavor RAM value that is provided ...I would like to attach the screenshot of the hypervisor stats tried for the above usecase

[sreeteja (sreeteja)]

This also solves this bug https://bugs.launchpad.net/nova-dpm/+bug/1691039

[Andreas Scheuring (andreas-scheuring)]

There are a copule of other bugs related to memory. I guess this is an area that needs some love

* https://bugs.launchpad.net/nova-dpm/+bug/1691039
* https://bugs.launchpad.net/nova-dpm/+bug/1689234
* https://bugs.launchpad.net/nova-dpm/+bug/1700974
* https://bugs.launchpad.net/nova-dpm/+bug/1701235