Numerous studies have focused on the performance and cache consistency of network file-access protocols [4,8,11,13]. In particular, the benefits of meta-data caching in a distributed file system for a decade old workload were evaluated in [13].
The VISA architecture was notable for using the concept of SCSI over IP[6]. Around the same time, a parallel effort from CMU also proposed two innovative architectures for exposing block storage devices over a network for scalability and performance [3].
Several studies have focused in the performance of the iSCSI protocol from the perspective of on data path overheads and latency[1,5,12]. With the exception of [5], which compares iSCSI to SMB, most of these efforts focus solely on iSCSI performance. Our focus is different in that we examine the suitability of block- and file-level abstractions for designing IP-networked storage. Consequently, we compare iSCSI and NFS along several dimensions such as protocol interactions, network latency and sensitivity to different application workloads. A recent white paper [14] compares a commercial iSCSI target implementation and NFS using meta-data intensive benchmarks. While their conclusions are similar to ours for these workloads, our study is broader in its scope and more detailed.
A comparison of block- and file-access protocols was first carried out in the late eighties [10]. This study predated both NFS and iSCSI and used analytical modeling to compare the two protocols for DEC's VAX systems. Their models correctly predicted higher server CPU utilizations for file access protocols as well as the need for data and meta-data caching in the client for both protocols. Our experimental study complements and corroborates these analytical results for modern storage systems.