[PVE-User] memory management of a vm in HA

[PVE-User] memory management of a vm in HA

[marcos negrini]

Hello:I am administrator of a proxmox cluster and I have been testing High Availability, and I am very satisfied with the performance; but I wanted to understand a little more in depth how the memory management works. I did the tests in a cluster of 3 servers with a SAN storage, I tried to cut the power to a physical server and almost instantly the vm's that were configured with HA went to the next node, my question is, how do you manage the memory of each vm's? do you pre-share it in the other physical servers so that the memory status of each one is not lost? how do you manage the loss of the information that was not copied? is there any technical document of this implementation in proxmox?Regardspd: sorry for my english level, I hope my doubt is interpreted.

Re: [PVE-User] memory management of a vm in HA

[Fabian Grünbichler]

On June 17, 2021 4:47 pm, marcos negrini wrote:
> Hello:I am administrator of a proxmox cluster and I have been testing High Availability, and I am very satisfied with the performance; but I wanted to understand a little more in depth how the memory management works. I did the tests in a cluster of 3 servers with a SAN storage, I tried to cut the power to a physical server and almost instantly the vm's that were configured with HA went to the next node, my question is, how do you manage the memory of each vm's? do you pre-share it in the other physical servers so that the memory status of each one is not lost? how do you manage the loss of the information that was not copied? is there any technical document of this implementation in proxmox?Regardspd: sorry for my english level, I hope my doubt is interpreted.

I'd suggest reading [1] as a starting point. To answer your questions:

- guest memory is not replicated or shared between nodes, HA just tries 
  to ensure the guest is running "somewhere" according to the HA 
  configuration
- ideally your guests' volumes are on shared storage, but if you can 
  live with losing data since the last replication, ZFS with replication 
  can also be an option
- if a node disappears/crashes/loses quorum/.. it gets fenced, the still 
  quorate part of the cluster will notice and "steal" the affected HA 
  resources
-- if the fenced node is still responsive, it's watchdog timer will 
   expire and it will shutdown (stopping all running guests in the process)
-- the stealing node will wait a certain amount of  time to give the 
   fenced node time to be completely fenced, then it will take over the 
   guest configs and start the guest
- additionally, you can configure what should happen to HA resources on 
  (orderly) node shutdown/reboot (see "Node Maintenance" in the admin 
  guide) - here one of the options is to migrate them to other nodes, 
  which is possibly what you triggered in your test?

1: https://pve.proxmox.com/pve-docs/pve-admin-guide.html#chapter_ha_manager