Due to a sizeable number of our servers still having only 1G NICs we run into issues with live-migrating some instances with a lot of RAM and high memory activity (i.e. that the services cause memory changes at a higher rate than it can be migrated off to a new node).
There are two qemu options which should help here and which we could evaluate:
- XBZRLE (Xor Based Zero Run Length Encoding):
https://github.com/qemu/qemu/blob/master/docs/xbzrle.txt
- auto-converge: