DRS overloads Host memory when activating maintenance mode

[marcn86]

Hi everyone,

as described above I have an issue with the placement of VMs during an automatic DRS evacuation action when putting a host in maintenance mode.

I have a cluster constisting of 8 ESXi 6.0 hosts á 32pCPU / 512GB RAM. There are some really big VMs (>256GB RAM) running in the cluster.

My problem starts when putting one of the hosts in MM. DRS sometimes migrates 2 of the big VMs on one of the other hosts, which results in an overload of the hosts physical memory, host starts to swap, VMs become barely responsive and so on...

Does'nt DRS check the resources of target hosts before migrating a VM during an evauation action?

I hope someone can help me

Regards,

Marc

[Finikiez]

Hi!

DRS makes deceisions on VM's active memory from ESXi perspective.

To avoid placement "big" VMs on the same host I recommend to make VM anti-affinity rules in your DRS cluster. This will help you.

VM-VM Affinity Rules

[depping]

DRS looks at active memory indeed, you could change DRS to look at consumed and coincidentally Frank wrote about that yesterday: http://frankdenneman.nl/2018/01/15/vsphere-6-5-drs-memory-balancing-non-overcommitted-clusters/ Of course a VM to VM anti-affinity rule would also work, especially when it is only 2 VMs that may be a better solution. Either way, you have two options now 🙂

[marcn86]

Thanks for your replies so far!

We're using the DRS advanced option "PercentIdleMBInMemDemand = 100" for some time, I think it's the expected behavior is the same as with the new additional option "Memory Metric for Load Balancing", isn't it?

I had two VMs (VM1 with 300GB RAM and VM2 with 256GB RAM) on the same host after activating MM on another host in the cluster. This host ran at "99% consumed memory" and the two VMs stopped working properly.

I didn't expected DRS two put these two VMs together with these DRS settings.

[Finikiez]

Do you have resource pools or per-vm memory reservations configured in your cluster?

[marcn86]

No resource pools and no reservations configured.

[Finikiez]

Hmm..

That might be a bug or maybe advanced parameter hasn't been assigned correctly.

You can check how your DRS cluster works with Fling utility DRS Lens