If available (development sites only), one option might to be use the REPD code in c36a5; this would eliminate the need for many short runs to do replica exchange, which many ad hoc methods do. The initial files (RTF, PARAM, PSF, etc.) are read only once with the REPD code.
Otherwise, you best bet might be to create a subdir in /scratch or /tmp (site dependent) on each of the nodes in the job, which contains the files read repetitively. This would at least eliminate any overhead in reading data from a remote disk, e.g. an NFS file server.