Dynamic Memory is not working all the time

Question

We are in the process off moving our 2008R2 VM's from the 2008R2 HyperV servers to new Server 2012R2 Hosts.
We shut down the VM's copy the files and VHD's to the new CSV's en import the VM in the Hyperv Manager. Then we make them high available in the Failover Cluster Manager (Configure role - Virtual machine). We mount the integration tools and update the
VM to version 6.3.9600.16384
For a specific type of VM (mostly RDS Host servers) we always had Dynamic Memory configured (when they were hosted on de 2008R2 platform), so we are using the same settings on the 2012r2 platform. The memory settings were;
Startup memory: 1024 MB
Minimum memory: 1024 MB
Maximum memory: 12288 MB
These VM's reboot every morning, this is done for specific reasons. But now once in a while (once per week/2 weeks) we notice that the VM's are not using more memory then 1024 MB while the demand is much higher. Rebooting the server helps most of the times,
live migrating to another host also helps. In the VM we see that memory usage in the taskmanager is 99-100%, and after the move it immediately starts using more than the minimum configured amount.
Until the failover the memory usage was 1024 MB and it did not get any higher.
This happened several times. Last week we changed the Memory configuration to:
Startup memory : 2048 MB
Minimum memory: 2048 MB
Maximum memory: 12288 MB
But this morning we had a call about the performance of one of the VM's, We saw that it was only using 2 GB memory while the demand was much higher. After live migrating it to another host it started using more memory immediately.
The 2012R2 hosts are not overcommited, there is a lot of memory still available for the VM's. Those VM's never had this problem on the 2008R2 Hyperv platform.
Any idea why this happens?
Peter Camps

Answer 1

Peter,
I think this is a bug of some sort. I say that because the components that make up dynamic memory are as follows.
Memory Balancer(Host service, coordinates how memory changes are made.) This is also what shows the memory demand counter i believe.
Dynamic Memory Virtualization Service Provider (this is included your VMWP.exe proccess, one per VM. Essentially how it runs on the host. He listens to the Service Client for metrics)
Dynamic Memory Virtualization Service Client (this is inside the VM and reports to the Dynamic Memory Virtualizaton Service Provider.)
Since you live migrated the machine it made dynamic memory work on the other host. This means the Service Client is running in the client and shouldn't be an issue. The Memory Balancer is the server and shouldn't be the issue, so this means the "Dynamic
Memory Virtualization Service Provider" is in question. When you live migrate the machine its going to create a new VMWP.exe process on the clustered server. So now the question is it the host that couldn't listen to the service or the worker process
skipped a beat and has a bug.
Out of curiosity does it happen to both hosts? Also have you profiled the servers to see how much memory they really require on start-up? When you reboot the RDS servers, how many VM's do you reboot and is it a staggered process?