The current broker implementation is optimized for delivery of messages. As a result current paging implementation is aggressively caching paged messages and it works well with a small number of paged queues. But in use cases where there are thousands of paging queues the JVM memory requirements become unsustainable. For a example for 1000 paging queues the current implementation requires about 60GB of memory just for caching.
The broker should offer ways of configuring/limiting the size of those caches.