This paper reports a protein-simulation grid that uses grid remote procedure calls (GridR-PCs) to a special-purpose cluster machine for molecular dynamics simulations. The grid was implemented using Ninf-G, Torque, LAM, and the Globus Toolkit. To avoid the inefficiency of a single GridRPC session using all the nodes of the cluster, we designed the grid so that it works efficiently when multiple GridRPC sessions share the cluster. This was done by putting the dedicated nodes (PCs with special computation boards) under the management of the Torque system, thus enabling the manager to dynamically configure a cluster with the requested number of dedicated nodes. In addition, a new job type was added to the Globus toolkit and new backend procedure was added to Ninf-G. The Ninf-G stub was separated from processes that actually perform the force evaluation on the dedicated nodes. Simulations for two proteins gave promising results. Simulations performed using a four-node cluster and a 100-Mbps LAN for GridRPC sessions were 4.6-17.0 times faster than the same simulation performed on the local client PC, while their communication overhead was less than 20% of total execution time. Even when the the four-node cluster machine was shared between two distinct simulations of proteins, the two GridRPC communications did not interfere with each other. This showed the efficacy of multiple GridRPC sessions.