Because we don't want to use 20 IPv4 addresses for the cluster of 20 nodes, when we only have so much addresses assigned to our institute. We could have gone the NAT route, but then we'd need to have some router. And if we designate the head node as router, all traffic would not go through the switch directly, but first through the head node and then out. This would mean that the nodes are less independent, as they have this one additional choke-point. Our university gave us a /64 for this network, so we just used that and it worked flawlessly, also for university-internal distro-package fetching and host-cluster connections.

Also, research software is usually working nicely with IPv6. If we encounter something that would really need IPv4, we could update the thing and give it some local subnet - but currently we were lucky.