Artificial Intelligence Computing Leadership from NVIDIA
VMotion behaviour
Hi there, Hopefully a quick one for you all... With a standard vSphere GRID deployment, where two ESXi hosts are identical, including GPU, and vMotion functions as expected, can anyone provide details of whether any 'pre-flight' checks are done on the GPU side prior to the migration task? Imagine the scenario above, however all the GRID VMs on one host are assigned, say T4-2Q. On the target host, a VM is running T4-4Q. Will the migration task stop the migration because the profiles do not match or will the migration take place, and the error only occurs when that VM attempts to boot with the non-matching profile? Thanks in advance - I would run this up myself but our test environment is down for a few days! Neal
Hi there,

Hopefully a quick one for you all...

With a standard vSphere GRID deployment, where two ESXi hosts are identical, including GPU, and vMotion functions as expected, can anyone provide details of whether any 'pre-flight' checks are done on the GPU side prior to the migration task?

Imagine the scenario above, however all the GRID VMs on one host are assigned, say T4-2Q. On the target host, a VM is running T4-4Q. Will the migration task stop the migration because the profiles do not match or will the migration take place, and the error only occurs when that VM attempts to boot with the non-matching profile?

Thanks in advance - I would run this up myself but our test environment is down for a few days!

Neal

#1
Posted 01/17/2020 01:01 PM   
Hi If the 4Q VM you’re migrating is powered on, the migration will fail due to insufficient resources. If the 4Q VM is powered off, then the migration will succeed, but it won’t be able to power on until a GPU with suitable Profile support is available. The issue becomes more interesting in environments that rely on DRS, as DRS is not GPU aware at this time. Regards MG
Hi

If the 4Q VM you’re migrating is powered on, the migration will fail due to insufficient resources. If the 4Q VM is powered off, then the migration will succeed, but it won’t be able to power on until a GPU with suitable Profile support is available.

The issue becomes more interesting in environments that rely on DRS, as DRS is not GPU aware at this time.

Regards

MG

#2
Posted 01/18/2020 09:46 AM   
Scroll To Top

Add Reply