As I understand from here
http://static.springsource.org/spring-batch/reference/html/scalability.html
the better approach would be to have just one poller and than distribute the job to the cluster through one of the mechanisms provided by spring Batch (I think the one named Remote Chunks is the best choice here).
I don't think you should worry about the clustering strategy as this is handled either by Spring Batch or by other clustering distribution mechanisms.