Although client applications may initiate first contact with any node in a Swarm cluster - even choosing to access the same node every time - it is recommended that the node of first contact be evenly distributed around the cluster. A DNS server that supports round-robin resolution of a group of IP addresses associated with a single canonical name can help an administrator distribute the node of first contact (also called the Primary Access Node- PAN).
Any mechanism that virtualizes the IP addresses of Swarm nodes or tries to control the nodes to which a client connects will interfere with Swarm and will create communication errors. Following HTTP 301 redirects by the client application is a requirement. This requirement is ameliorated if using an SCSP Proxy or Content Gateway to connect to the storage cluster.
The placement of Swarm nodes behind an HTTP load balancer device is not a supported configuration.
Please contact support or your SE for architecture recommendations.