Short VM freeze when migrating to another host

zmk

What could be the algorithm for copying the RAM of a running virtual machine to another host?

Copy the RAM of the running VM to another host.
While the copying was in progress, the RAM of the running VM has already changed.
Copy the changes.
While the copying was in progress, the RAM of the running VM has already changed.
Copy the changes.

Finally, we understand that this is an infinite loop.
Freeze the running virtual machine.
The RAM of the non-running virtual machine no longer changes.
Copy the changes RAM of the non-running virtual machine.
After copying the changes, the RAM of the non-running VM on the old host matches the RAM of the VM on the new host.
Unfreeze the VM on the new host.

The more uncopied changes at the time of freezing, the longer the freezing time.

Copying of uncopied changes after freezing cannot happen instantly.

rfx77

@zmk We only had the dmesg entris on Xen, not on VMWare and not on HyperV

nikade

@arc1 how much ram/cpu/disk does your VM's have?
Seems like something is taking too long in the last phase of the migration, when the original source and destination VM are syncronized.

zmk

The problem may be in the transfer speed between hosts.

nikade

@zmk yeah maybe, we're connected with 2x10G on each host to the network and while doing a migration (without storage migration) between 2 hosts in the pool I can see it spike at 6-7Gbit/s.

arc1

@nikade 4cpu, 16ram and roughly 200gb disk.
10ping downtime was on test enviroment with slower speeds between hosts, so this explains longer freeze.
But on production 2x25gb lacp is still noticable freeze on VMs with more sensitive software (keepalived/etcd).Nothing too terrible we were just curious if this is normal behaviour.

nikade

@arc1 so if you go to XOA and the console of the VM, what happends then?
Is the VM frozen for the amount of 10 pings? Open taskmanager to see if there is any CPU activity.

arc1

@nikade Yes, the MV is frozen without cpu activity.

nikade

@arc1 said in Short VM freeze when migrating to another host:

@nikade Yes, the MV is frozen without cpu activity.

So the VM is actually frozen in the console?
Because if it wasn't I'd suggest adjusting the mac-aging in your switches, since the VM's mac adress will be bound to the physical hosts switch-port for a period of time after migrating.

andrewperry

We're seeing this issue when trying to migrate a Debian VM with 16GB of RAM.

It is a worker node in a Kubernetes cluster so it is likely that the RAM changes a fair bit. It is not uncommon for the migration to fail due to the freeze hitting a 30 second time limit.

A Windows 10 Pro VM with 16GB of RAM migrates fine, because not much is changing in the RAM I expect.

Following along for recommendations! Our hosts sound very similar to @arc1 except our network speed is slower, which is one thing we are working on.

nikade

And you guys aren't using any kind of dynamic memory?
Can you post a screen dump of the Advanced tab where it shows the memory configuration?

We have VM's with 128Gb ram that migrates just fine, when migrating it between hosts the network shows peaks at 7,6Gbit/s and it is migrated in about ~20 seconds.
Smaller VM's with 8, 16 or even 32Gb ram is migrated almost instantly.

robyt

@nikade
in VM (linux) with a free i see 94 gb of total memory

andrewperry

@nikade thanks for the ideas of where to look!

In my case we're testing and just have a 1Gb link between these hosts, which is what I was putting it down to.

This particular VM is a freshly migrated PV from Debian Xen with:

Memory limits (min/max)
Static: 16 MiB/16 GiB
Dynamic: 8 GiB/16 GiB

Could that Dynamic setting be the problem because as I recall it reduces the VM to 8 on migrate, so when doing the migrate perhaps 8 isn't enough for the VM?

I will try changing it to 16/16 and see if that has any noticeable impact. Thanks!

olivierlambert

Yes that's very likely.

nikade

@andrewperry yeah try set it 16/16Gb instead, it will probably do some magic

robyt

@olivierlambert hi, today i've upgraded my host..
The big VM frozen for ~7 minutes, is a big vm (96 gbram and 32 cpu) but 7 minutes is a very long time (for customer!)
i've setting 96/06 in dynamic: is a normal time?

olivierlambert

Can you provide more details? 96/96 in dynamic? Just doing live migration or suspending the VM?

robyt

@olivierlambert live migration, the vm is very important (today, in christmas holyday, i've received some phone calls for 7 minutes of freeze..)

olivierlambert

The topology looks insane Also, a live migration shouldn't make the VM inaccessible for more than few seconds, except if there is a LOT of memory pages changes, at a pace that is close to the transfer speed.

robyt

@olivierlambert ops.. why the best topology?