Live migration (host flag) between AMD EPYC

[harmonyp]

Wondering has anyone tested if live migration with host flag between different AMD EPYC (7001 series).

It would be good to know what works and what doesn't if anyone has tests.

 

[shanreich]

I haven't personally, but one thing to keep in mind when attempting this is, that the 5.15 kernel is known to cause problems with migrations and AMD EPYC CPUs. So if you try this and encounter any problems, please try our opt-in 5.19 kernel as well in order to rule out any problems related to the kernel.

 

[fiona]

Hi,

Wondering has anyone tested if live migration with host flag between different AMD EPYC (7001 series).

It would be good to know what works and what doesn't if anyone has tests.

in general, migrations between CPUs of the same vendor are fine. That said, there are sometimes kernel bugs (but there can even be such bugs when using the same model), so I'd say the closer the model the better.

I haven't personally, but one thing to keep in mind when attempting this is, that the 5.15 kernel is known to cause problems with migrations and AMD EPYC CPUs. So if you try this and encounter any problems, please try our opt-in 5.19 kernel as well in order to rule out any problems related to the kernel.

Are you referring to the TSC-related issue? A fix for that is contained in kernels >= pve-kernel-5.15.74-1-pve. Or is there some other issue still?

 

[shanreich]

Are you referring to the TSC-related issue? A fix for that is contained in kernels >= pve-kernel-5.15.74-1-pve. Or is there some other issue still?

Yes, I think that's the one I meant. I wasn't aware that there was already a fix available for 5.15 - thanks for pointing it out!

 

[shanreich]

I just looked it up again: Just yesterday someone encountered an issue on 5.15.74-1-pve with migration and AMD 7xxx EPYC CPUs. It has been resolved with upgrading to 5.19 - so maybe there is another issue with 5.15 after all.

 

[bzb-rs]

Our cluster is on 2nd Gen EPYC with HA enabled, found no issues with live migrations even after latest upgrade to 7.3-3