In short - I would recommend using
SolrCloud in a model where you have a load balancing in place (either by the built-in Solr Load Balancing or by an external load balancer that is handling the load between the nodes). This should be the recommended approach for handling the shards optimally.
In that case you must connect to the load balancer url and pass
solrcloud=true in the connection string.
So your connection string should look like this:
If you need some inspiration you can refer to Setting up SolrCloud Documentation.
If you are not having a load-balancing in place (which I wouldn't recommend :)). You should just connect to the
Master server on all the servers that are doing
writes and use the slave for all the servers that are doing
reads as normal Solr servers, but you are going to lose the primary capabilities of scaling :).