This one year project was concerned with distributed computer systems in which several computers are logically and physically interconnected; the machines cooperate under decentralised control to execute applications programs. Interest has been concentrated in MIMD systems which have two particular properties. Firstly, the system is homogeneous in that all sites support the same basic executive using the same primitive level interface; and secondly, the system as a whole appears as a black box to the user, who is unaware of its internal structure (there is a single system wide operating system).
A further feature is that inter-site communication is along so-called thin-wire (relatively low bandwidth) channels, in contrast to high bandwidth thick wire channels, where physically the link is a parallel bus. Thus the location of the machines is permitted to be geographically separate if required.
The project began with a review of actual and potential application areas for distributed systems. At one end of the spectrum are existing commercial DP distributed systems, almost all of which are based on a large central mainframe at head office. Varying degrees of local processing are found in regional sites, ranging from simple data validation to accessing local copies of data with the main machine undertaking archiving and gathering management report information.
The most ambitious realisation of this type of scheme is in distributed data base management systems. Few, if any, such systems exist; major problems of reliability, data consistency, integrity and data base administration remain.
At the other end of the spectrum, the review included several projects which are setting out to connect together mini or micro computers via high speed links.
Application areas are often in process control and real time areas and there is a considerable interest in programming those systems in high level concurrent languages such as Concurrent Pascal or Modula.
As a result of this review, interest focused on distributed systems in which the prime shared resource is not CPU power but data. Investigations were made of suitable architectural models for systems in which data management is of major concern.
The main result of this feasibility study has been to isolate a fruitful area for further research. A successful application was then made to the SRC for funding for a three year project on the design, modelling and implementation of a distributed filestore (see following project description). The remainder of the feasibility study was largely concerned with refining ideas on this topic [1].
The research associate funded by the grant resigned at a fairly early stage in the year This was particularly regretted as his special interest. was in simulation and modelling and he left. just as this topic was beginning to build up in the project.
In view of the circumstances, no immediate replacement was sought; however K. Lunn was recruited in May 1979 as an SRC research student under the Instant scheme.
Visits were made to a number of academic and industrial establishments. Nineteen internal reports and working documents were produced plus ten trip reports.
1. K.H.Bennett, A Feasibility Study of Loosely Linked Computers - Termination Report, March 1980.
Data is a crucial asset for many organisations, and the reliable storage of data in computer systems is hence of considerable economic and practical importance. The purpose of this project has been to study how the replicated resources that occur in distributed systems can be harnessed to provide highly reliable file storage in a way that is simple for the user to exploit.
The original goals were to study a distributed filestore, by implementing a design, and using simulation and modelling to provide quantitative analysis. Two major objectives were set: firstly, to aim at a naming scheme which would hide location dependence; and secondly, to provide highly reliable storage through a single name of a user file, with the degree of reliability indicated by an integer.
A working distributed filestore supporting reliability through multiple copy techniques has been demonstrated. The filestore is implemented across two LSI11/02 Machines and is accessed from Terak and Microengine workstations; the interconnection is via a Cambridge Ring. A feature of the filestore is the overlay mount concept, to avoid interdependence of disc drive availability.
The Cambridge Ring has simulated several sub-projects, including simulation studies and real devices. A DMA access logic has been designed and simulation work on it is being undertaken; a terminal concentrator was designed early in the project and has been taken up elsewhere at Keele.
More recent work on the project, on majority consensus algorithms (including a working implementation) and network partitioning problems has been moved to a PDP11/34 running V7 UNIX; the LSI11/02 based implementation has pushed the modest equipment beyond its capabilities.
Some six external papers have been published; several more are in the pipeline. About 80 internal reports and 45 trip reports have been written; two workshops were held at Keele and research groups from a number of institutions have visited the project.
The project was completed in December 1982.
Some of the local area network results are directly applicable in products; the terminal concentrator is being used by the Keele Computer Centre for a TS29 PAD implementation. Informal discussions with several industrial concerns have occurred, although no direct exploitation has been made. It is expected that the distributed filestore algorithms could be incorporated into a UNIX-like distributed operating system.
1. K. H. Bennett, O. P. Brereton, K. Lunn and P. Singleton, The Keele Distributed Filestore, Internal Report DCP/WD/80.
2. K. Lunn and K. H. Bennett, An Algorithm for Resource Location in a Distributed Computer Network, ACM Operating Systems Review, Vol. 15 No.2. April 1981.
3. K. Lunn and K. H. Bennett, A Highly Reliable Distributed Directory System, Proceedings of 2nd International Conference on Distributed Systems, Paris 1981 (IEEE).
The context of this project is broadly similar to that of the preceding research (q.v.). However, it is hoped to show that the ideas for a reliable filestore can be incorporated into an existing distributed operating system (UNIX United).
Firstly it was decided to undertake a study of filestores to examine possible future directions and problems. Secondly. work on the replicated files was to continue, using UNIX United as the research vehicle. It was planned to produce a working implementation of a majority consensus algorithm for managing replicated file consistency. Finally it was planned to start work on a collaborative venture with Prof. Randell's group at Newcastle, on network topology problems in UNIX United.
The study of research problems in filestores led to the submission to SERC for funding to support a project Structures for Reliable Filestores. This envisages filestores developing into reliable object stores which not only retain the state information but also procedures to manipulate that state. It is thought that this approach will provide a unified solution to the problems of replicated files, primitives for database implementation and mixed media files. Continuing the theme of reliability. it is also hoped to undertake work on the management of partitioning in distributed systems.
An implementation of a majority consensus on top of UNIX United is under way and is on target for completion by the end of the project.
A pre-release version of UNIX United was made available by Newcastle. and successfully mounted at Keele. Feedback to Newcastle has been provided.
Work on the network topology problem has started; a definition of the problem has been produced.
A number of papers have been produced for submission for external publication.
Implementation of the majority consensus algorithm is continuing. This will form the basis for future work on replicate copy management. The research on network topology problems is getting under way and it is hoped to complete this study be October 1984.
Further work on the object store is being undertaken, with the objective of producing a specific design proposal document. It is planned to produce a very simple demonstration of mixed media files in order to obtain a feel for problems and applications.
Initial discussions with industrial firms suggest that there is considerable interest in our approach to reliable media files.
The most useful references are those given for the earlier project.
