The network file system protocol (NFS) has been the leading distributed file system for workstations since it was first introduced by Sun Microsystems in 1986. The geographical scale of NFS has been limited to the local area due to its relatively low performance on the wide area Internet. However, with the advent of high bandwidth wide area networks such as ATM, NFS over WANs may become more promising. In this paper, the performance of NFS over various sizes of WAN is studied The effects of ATM flow-control and queueing strategies on NFS are discussed, as are the performance of TCP and UDP as NFS transport protocols. The primary conclusion is that standard NFS over UDP works well over ATM WANs as long as ATM-level flow control keeps the cell loss rate under one percent. In some cases, NFS over TCP works badly with small packets due to unfortunate interactions with TCP's congestion window.

In credit-based flow control for ATM networks, a buffer is first allocated to each VC (virtual circuit) and then credit control is applied to the VC for avoiding possible buffer overflow. Receiver-oriented, adaptive buffer allocation allows a receiver to allocate its buffer dynamically, to VCs from multiple upstream nodes based on their bandwidth usage. The paper describes, in detail, such an adaptive algorithm capable of supporting a wide range of link speeds and propagation delays, and also packing multiple allocation and credit records in a single message. Analysis and simulation results show that even under highly bursty traffic, the adaptive scheme guarantees no cell loss due to congestion, and achieves excellent performance in utilization, fairness, ramp-up and packing, while requiring only relatively small node memory and bandwidth overhead. The required memory need only be 4*RTT+2*N, where RTT is the link round-trip time in cell cycles and N is the number of VCs.