Weblogic Cluster failover takes too long under load

 We have issues with a cluster failover, but only in certain conditions :

If node1 is no longer available (power outage), and the IP-adress is gone, apache will send everyone to node2, then it takes ages before node2 answers that it will take the session, as if node2 is trying to contact node1 to make sure node1 is dead (we see stuff on T3, but no clear info).

If node1 is no longer available and the IP-adress is created on another host, without WLS, apache will send everyone to node2 and failover will work just fine.

Of course, this only happens in PROD, under load (hundreds of concurrent sessions).