[SOLVED] "Pull the plug" HA without downtime?

[Julen]

I was wondering...
The HA feature in proxmox ensures that a VM (or LXC) will start on another node in case a node goes down, but the machine will shutdown and will have to reboot on the next node.
Would there be any way to ensure that the machine will instantly appear on the other node once its hypervisor is down without losing its memory?

That would probably require a really fast network and storage to use as "shared RAM". It would certainly have an impact on the performance, as it would be sort of fully swapped, but it would have the advantage of being able to guarantee 100% uptime.
This would be specially useful for the kind of VMs (M$ VM$...) which don't handle "pull the cable" situations in such a clean way.

Any thoughts?

 

[Denny]

The only virtualization platform that I am aware of that has that feature is Xen (with Remus extensions) Is it worth taking on all the challenges associated with xen for that feature? Maybe engineering something at the network or app level would better serve (load balancing or CARP)?

 

[wolfgang]

Hi,

what you talking is not HA it called Fault tolerant, hot-standby or COLO.
The problem with this technology is you need a powerful network min 40GBit dedicated per VM.
The reason is you have sync every Memory bit and Network package life to both hosts.

Qemu can do this but it not stable yet and so not implemented in PVE.

 

[janos]

Hi,

what you talking is not HA it called Fault tolerant, hot-standby or COLO.
The problem with this technology is you need a powerful network min 40GBit dedicated per VM.
The reason is you have sync every Memory bit and Network package life to both hosts.

Qemu can do this but it not stable yet and so not implemented in PVE.

Hi,

We have vmware cluster, and this technology (at least with vmware) working well over 10GBit or 1 GBit (optical) network. Yes, offcourse, 40gbit network, or 100gbit will be better, but better equipment cause only better performance, its not a requirements.

 

[Julen]

That's very interesting...
I didn't know about Remus and COLO before, but as Wolfgang says it must be quite demanding from a hardware point of view.
I still think that it would be a "nice to have" option, even if the VMs run much slower.

Deny mentioned the load balancing option, which I totally agree with. Except that the problematic (Windows) VMs (the ones which don't handle shutdowns very well), happen to be the ones that cost a lot per machine (in licenses) and are not very "replication friendly". So, I thought that it could be an option to run those ones significantly slower, in order to avoid any possible shutdowns.