Message-Driven File System Experiments

As parallel machines continue to scale, the data sets that are generated continue to grow in proportion. Additionally, large production runs make take several week or months, so it is important to save intermediate results periodically to guard against unexpected shutdowns and mechanical failures. Thus, frequent storage of large quantities of information on backing store is required. Conventional file storage systems, while adequate for small numbers of computers, do not take advantage of the bandwidth opportunities provided by three-dimensional, mesh-connected architectures such as the CRAY T3D and MIT J and M-machines.

A scalable, concurrent file system has been developed in which disks are attached via a two-dimensional plane of edge connections at the periphery of the machine. There are two levels of concurrency in the file system. Compute nodes may send messages to arbitrary locations within the file system plane, and data blocks are striped across physical disks. The implementation is "stateless" and thus consistent with the industry standard Network File System (NFS). Thus the file system appears over the internet as a conventional Unix file system.

The file system is an enabling technology that is necessary for applications experiments. The system has provided feedback to hardware engineers on how to organize disk hardware so as to improve throughput. This feedback has been incorporated into the M-machine design.

Zadik, "The Message Driven File System: A Network Accessible File System for Fine-Grain Message Passing Multicomputers", Masters Thesis, Department of Computer Science, California Institute of Technology. 1995