The current implementation assumes a network of workstations that all have the same order of magnitude of physical memory and are interconnected by a local area network. It would be interesting to explore the requirements that heterogeneous networks pose to the design of the remote pager. For example, on a wider area network the time it takes to transfer a page may not be identical for each server. In this case there may be more than three levels in the memory hierarchy (local memory, remote memory, disk), depending on the variance of the cost of communication among the hosts of the network. Connecting machines that have an enormous amount of memory (e.g. a supercomputer) to a network of workstations also poses some problems. When the supercomputer memory is idle, it may not always be easy to find enough free remote workstation memory in order to be able to use reliability policies. In this case, a no reliability policy can be used, since all remote memory will be provided by a single host (the supercomputer).