Open Stack

On 09/07/2017 02:27 PM, Sahid Orentino Ferdjaoui wrote:
> On Wed, Sep 06, 2017 at 11:57:25PM -0400, Jay Pipes wrote:
>> Sahid, Stephen, what are your thoughts on this?
>>
>> On 09/06/2017 10:17 PM, Yaguang Tang wrote:
>>> I think the fact that RamFilter can't deal with huge pages is a bug ,
>>> duo to this limit, we have to set a balance  between normal memory and
>>> huge pages to use RamFilter and NUMATopologyFilter. what do you think
>>> Jay?
>
> Huge Pages has been built on top of the NUMA Topology
> implementation. You have to consider isolate all hosts which are going
> to handle NUMA instances to a specific host aggregate.
>
> We don't want RAMFilter to handle any NUMA related feature (in Nova
> world: hugepages, pinning, realtime...) but we don't want it blocking
> scheduling and that should not be the case.
>
> I'm surprised to see this bug I think libvirt is reporting the total
> amount of physical RAM available on the host and that is not depending
> of the size of the pages.
>
> So If that is true libvirt is reporting only the amount of small pages
> memory available we will probably have to fix that point instead of
> the RAMFilter.

 From what I can see, in Mitaka the RAMFilter uses "free_ram_mb", which is 
calculated from "self.compute_node.memory_mb" and 
"self.compute_node.memory_mb_used".

memory_mb is calculated for libvirt as self._host.get_memory_mb_total(), which 
just calls libvirtmod.virNodeGetInfo() to get the memory available.  As far as I 
can tell, this is the total amount of memory available on the system, without 
paying any attention to hugepages.  (I ran "virsh nodeinfo" to check.)

memory_mb_used() is kind of messed up, it's calculated by running 
get_memory_mb_total() and subtracting the memory showed as available in 
/proc/meminfo.

So from what I can see, RAMFilter should be working with all the memory and 
ignoring hugepages.

(There is some additional complexity involving limits and hypervisor overhead 
that I've skipped over, that might possibly affect what's going on in the 
original problem scenario.)

Chris