Use "Clusters" -> "Logical Networks" -> "Manage Networks" for assigning network for migrations. It's ovirtmgmt by default. Note that migration network has to have IPs on hosts.
Nice, do you think I should dedicate a link only for migration, for safety ?
Hosts have VDSM configuration option - [vars]/migration_max_bandwidth in /etc/vdsm/vdsm.conf. It's 52MB/s by default. So it looks like your network is already highly used.
It's not highly used, it's yet on a 10/100 Mb switch...But it's just temporary. Anyway it's interesting because it simulates a highly use network, which can happen.
Your VM migration fails because migration is so slow so it just can't migrate without downtime. You can adjust "Use custom migration downtime" VM option to make any VM migrated over any network. Please note it's a maximal value. Actually it starts with a value 10 times lower and increases it automatically. 5000 is good value to start in your case.
Thanks, but it doesn't seem to change anything : I added max_outgoing_migrations = 1 migration_max_bandwidth = 5 in /etc/vdsm/vdsm.conf on my concerned host, but it still goes to 98% use of network and migration fail...I restart vdsmd.service but it doesn't seem to change anything. Any idea ?