The announcement of the Software Distribution Service (see below) is the direct result of the need expressed at the EASE Workshop on Numerical Methods at Warwick last year. The Community is encouraged to use this facility and comment on its usefulness.
Two publications instead of one, both free! We welcome the appearance of the Graphics Newsletter, which represents an amalgamation of the IGWP and the AGOCG newsletters. It will be published and distributed each month with this Newsletter. Editorial enquiries should be addressed to Rae Earnshaw as detailed therein.
Following recommendations made at the EASE Working Party on Numerical and Mathematical Software at Warwick University in February 1988, the University of Kent has been funded for two years to provide a pilot numerical software distribution service, which will formally start on 1 January 1990.
The software will be distributed over JANET via electronic mail using the netlib system currently providing a similar services in the United States. Examples of software which will be available from 1 January 1990 are:
A full list of the available libraries will be obtainable by sending a mail message to netlib@uk.ac.ukc whose body contains the single line
send index
The object of this service is to provide researchers with rapid access to specific numerical routines. For example, assume a user wishes to solve a system of linear equations, she consults her local numerical expert who advises that the Linpack routines dgeco and dgesl are appropriate. She then types the following:
mail netlib@uk.ac.ukc end dgeco dgesl from linpack
In a short time (hopefully less than an hour) mail comes back from netlib containing the double-precision Fortran subroutines dgeco and dgesl along with any other Linpack routines that these specified routines call.
There are a number of variants of the send command which control the amount of dependent software sent out by the system. There are also facilities available for retrieving the names of routines in the collection related to specified keywords and for retrieving electronic mail addresses. Further details are given in the index file and in the article by Dongarra and Grosse in CACM Vol 30, 1987 pp. 403-407.
We are very keen to expand the software available and we welcome suggestions for further libraries/ packages that researchers would find useful. We would also like to expand the directory of e-mail addresses. Both suggestions and addresses may be e-mailed to:
netlib-suggest@uk.ac.ukc
Note: netlib-suggest is already available; netlib itself will not officially come on-line until 1 January 1990.
At Sheffield University, from the point of view of computing, the 1988/89 session will be remembered as the Year of the PC. As a consequence of the ever increasing power of these machines we have witnessed the transfer of applications, previously associated with mainframe facilities, onto PCs. Users who have gained experience on PCs due to the availability of a wealth of PC-based office software are now looking to mount serious scientific and engineering applications on the same platform. Although it is not advisable to make predictions into the future, it seems likely that this trend will continue during the coming academic year. You may very likely find that next year (if not sooner) your favourite mainframe software is available on the IBM-Compatible machines running MS-DOS (or PC-DOS). Admittedly some large mainframe packages cannot yet be fully implemented on these machines and are available only as cut-down versions. This is, however, changing with the increased availability of Transputer based accelerator cards with appropriate software, and the advent of 80386/7 machines (such as the IBM MODEL 80) with compilers that support their use in virtual mode giving a very large, linear address space, even under MS-DOS.
Given this background and the steadily increasing demand on our already overloaded mainframes, Computing Services at Sheffield have recently investigated the possibilities of porting some FORTRAN programs from our main computers (an IBM 3083 and two Prime 9955) onto a PC, in a cost-effective manner.
The programs we were looking to port are quite self-contained, requiring only a FORTRAN compiler and a library giving access to simple graphics primitives. They had been widely ported in the past to a large number of machines where it could not be assumed that standard graphics (GKS) and common numerical facilities (like NAG) would be available. There is also a need to be able to use large arrays.
Even these simple requirements eliminated a number of systems. So after a preliminary study, we concentrated on two promising options:
Initial work involved looking at published material comparing the performance of the two engines. A useful article by Stephen S Fried in the 1987 Byte extra edition 'Inside IBM PCs' gave a single precision performance of 3700K Whetstones for a T800 clocked at 20 MHz while the performance of the 80386/7 combination clocked at 16MHz was quoted as 1354K Whetstones. These results were achieved with different compilers from the ones we were proposing to use, indeed the T800 benchmark was done with Occam. We repeated the measurements using the FOPR12 FORTRAN benchmark. In addition we did some quick measurements on our local mainframes to give us some feel for the results.
K Whetstones
IBM 3083 5737
Prime 9955-1 2620
Model 80 1403
Transputer 2690
The PC-based workstation figures are calculated from elapsed times, whereas the mainframe figures are from system clocks. The latter, hopefully, give a reasonable indication of the single-user, elapsed time performance of the mainframes.
These figures broadly confirm those published for 80386/7 and T800. Further, although the indicated performance of these engines is lower than the mainframes, they look attractive for computationally intensive programs when their dedicated performance is compared with the small fraction of a mainframe that a user might normally expect.
As a tough example of the software to be ported we selected FLUENT. This is a computational fluid dynamics code of about 36000 lines which runs on our mainframes. It is processor and memory intensive, requiring a minimum of 6Mbyte of storage even for small problems and many overnight mainframe batch runs before a typical problem converges. It has a built-in interactive graphics facility which relies on a small set of graphics primitives being made available. Beyond the initial loading of the program, little use is made of disc files except for a dump and restart facility.
After some effort the software was made to work on both platforms. A very simple port of the code was done, with no attempt being made to parallelise it for more than one Transputer or even to optimise it. The Salford Compiler on the 80386/7 was significantly easier to work with. It was much quicker for both compiling and loading, and gave better diagnostics. It also comes with an excellent integral debugger. Many users will love or hate the compiler from acquaintance with the earlier version available on Prime equipment. The performance of the completed FLUENT implementation on the T800 was about 15% quicker than on the 80386/7. This was disappointing, being much lower than expectations based on the relative Whetstone figures. The mainframes both looked slightly better than might be indicated by their Whetstone performance relative to the 80386/7, but not wildly so. We tentatively attribute the relatively poor performance of the T800 (still better than 80386/7) on a real problem to the static allocation of its fast memory (an ineffective use of this memory for a large application) as opposed to the dynamic fast caches possessed by the other machines.
It has been shown that porting a large mainframe program onto PC-based work-stations is perfectly feasible. PC-based workstations can handle problems as large as have hitherto been possible only on the mainframe computers, certainly up to 16Mbyte and in some cases beyond using virtual memory on the 80386/7 and FTN77/386. From the point of view of turn-round times, a dedicated PC can be considerably faster than a shared, heavily-loaded mainframe.
Based on our experience the Salford Compiler-80386/7 combination is the more satisfactory alternative where:
The advantages of a Transputer board solution outweigh an 80386/7 solution where absolute performance is important or where the software can make use of multiple transputers through the application of parallel processing techniques.
Depending on how the parallelisation has been achieved, performance of the software may be enhanced by simply adding more transputers (up to the limit of your budget and the capacity of the PC power supply). The price/performance of both systems is very attractive. For example the current price of a sensibly configured IBM compatible 80386/7 machine (20MHz, 40Mbyte disc, VGA colour display and 4Mbytes memory with MS-DOS) is now under £3500 incl. VAT. Memory remains expensive, however, and expanding the machine to 16Mbyte would cost a further £3000. A site license for FINn/386 for the UK Higher Education community is about to be available through CHEST (someone has to pay the site charge).
The market place continues to change: new Transputers are promised and the 80486 processor is available (though buying a machine with one in is a matter for 1990 rather than 1989). These promise an increase in performance by a factor of between two and four. More powerful Weitek floating point chip-sets are available for the 80386, supported by some compilers.
It would also be interesting to look at the price/performance of these combinations, compared to a modem workstation, a SPARCStation say. Any conclusions are bound to be of only temporary value as the manufacturers continue to leap-frog each other's price/performance.
Note: The views expressed in this article are those of the authors and do not of necessity reflect those of the EASE policy.
The last remaining Engineering Computing Facility Prime Computer offering a user service to the Engineering Board supported community will be switched off on 31 March 1990.
Before this happens a significant amount of work will have to be completed to ensure that all the data still required in the Prime Archive is transferred onto other systems.
Efforts have been made to contact all existing users of the system who have files currently archived, but some may have been missed. If you arc still, or have been, a Prime user and have used the archive for data that IS still needed, it is essential that action IS taken immediately to recover it.
At UMIST, we have been gathering together information and software which will facilitate the transfer of software from a SERC Prime (past or present) to a Sun. The ECF team at UMIST formerly supported the SERC Primes. A collected set of hints, suggestions, and look-up tables has been collated into a booklet entitled, Prime to Sun - Conversion Guide.
This guide is obtainable by e-mailing ease@uk.ac.umist with your full postal address.
The UMIST team will continue to offer a consultancy service for users making the transition from Primes to Suns. A small collection of conversion tools has been assembled at UMIST and these are detailed in the Guide. If you have already transferred software from a SERC Prime and to a Sun and have any advice or tools to offer, please contact via email ease@uk.ac.umist.
This is the last article in a series of four, the first three of which were in Issues 14 (Data Exchange at RAL), 15 (The STEP Standard for Engineering Data) and 16 (Data Modelling Language Express).
The primary aim of STEP (STandard for Exchange of Produce data) project is to define a standardised method of exchanging engineering data. The way in which this is being done involves a standard model of the data which is then mapped onto a physical file format to produce a neutral file for exchange. The data modelling language being used is Express which was described in the third article.
Once a standard data model for an application area has been formulated in Express, this may be used to automatically generate a schema for a database. It should then be possible for many applications programs to access the database with only the knowledge of the data model and the mapping algorithm. In this way, the standard data model forms the basis for integration.
In the Database Section of the Application Integration Group a project is underway to determine the feasibility of these ideas. There are three areas of work which are being investigated and each of these will be discussed in more detail.
RAL have developed a compiler which takes an Express model as input, parses and checks the model and then outputs one of a number of formats. One output format which is in the process of being developed is an SQL file for creating a database schema. A mapping is being developed to automatically create database tables using SQL from the Express definitions.
An example is given below:
ENTITY person;
first_name : STRING(10);
last_name : STRING(10);
middle_name : OPTIONAL STRING(10);
birth_year : INTEGER;
END_ENTITY;
ENTITY car;
make : STRING(10);
serial_no : INTEGER;
engine_size : REAL;
colour : STRING(8);
owner : person;
prev _owners : LIST[0:#] OF person;
UNIQUE
serial_number;
WHERE
engine_size > 0.0;
END ENTITY;
This is an express model which describes the attributes and relationships of two entities, person and car. It would map into SQL in the following way:
CREATE TABLE PERSON
(PERSON_ID SMALLINT NOT NULL,
FIRST_NAME CHAR(lO),
LAST_NAME CHAR(lO),
MIDDLE_NAME CHAR(10),
BIRTH_YEAR INTEGER,
UNIQUE (PERSON_ID) )
CREATE TABLE CAR
(CAR_ID SMALLINT NOT NULL,
MAKE CHAR(10),
SERIAL_NO INTEGER NOT NULL,
ENGINE_SIZE REAL,
COLOUR CHAR(8),
OWNER SMALLINT,
UNIQUE (SERIAL_NO, CAR_ID) )
CREATE TABLE CAR_PREV_OWNERS_LIl
(CAR_ID SMALLINT,
LIST_INDEX SMALLINT,
PREV_OWNERS SMALLINT)
It can be seen that a table has been created for each entity in Express and that the table takes the same name as that entity. Each table has one column for each of the attributes of base types in the Express with the column type matching the attribute type. In addition each table is given an extra column for the purposes of assigning a unique id to each instance of an entity. This ensures uniqueness of rows in the table and provides a uniform method of referencing rows. In the CAR table the last column contains the identifier of a person already existing in the PERSON table, this is a foreign key. The integer in this column must match an integer in the person_id column in the person table. Unfortunately it is not possible to enforce this referential integrity in SQL at present.
The entity car includes an attribute prev _owners which is a list of pointers to person entities. In a relational database each column must be single valued so it is not possible to represent a list of items in a single column. Therefore, an additional table is created to support this construct. The table is given a name consisting of the original entity name, the name of the attribute in that entity and a subscript to denote the level. This subscript allows for multiple tables to be created if an attribute is, for example, a multi-dimension array.
In Express, entity and attribute names are allowed to be of unlimited length. Therefore, the names generated by the Express to SQL mapping described above would not be suitable for implementation into an SQL database as SQL restricts name length to 18 characters. In fact, many commercial SQL databases restrict this further down to as low as 12 for column names and 9 for table names.
A mapping has been developed which takes the Express names and generates table names of only 9 characters and column names of 12 characters. The table names and column names are guaranteed to be unique within the whole schema.
In addition to building the database schema, standard access software is required to enable each application to access the data in a common way. An application should not need to know whether it is accessing data from a database or from a neutral file. Therefore, standard access software is being developed that will keep the application independent of the data storage medium. The first implementation will be a C language binding using Dynamic SQL.
The area of work is to implement a demonstrator system to evaluate the usefulness of the tools developed. An Express model has been developed based on the RALBIC neutral file (RALBIC - A Simple Neutral File for Finite Element Data, Emson, Greenough, Diserens & Duffey, RAL-87-102.), for electromagnetics and semi-conductor device modelling entities and attributes, and the CAD*I specification (Specification for Exchange of Product Analysis Data, Thomas, Van Maanen & Mead (eds), Springer-Verlag, 1989) for basic finite element entities. This model has been translated into SQL by hand using the mapping described above and is currently being built using INGRES.
The next stage of the project will be to adapt an existing semi-conductor device modelling software package, which at present obtains data from a RALBIC neutral file, to access the data directly from the database. This is the first step to using the STEP methodology and a relational database to achieve applications integration.
We have been asked by the Computing Facilities Technical Advisory Group (CFTAG) to produce an assessment of superworkstations for visualisation in engineering.
This assessment will be part of the EASE supported programme of work. We believe that there must be a significant pool of people who are wrestling with these problems in this area of visualisation but without the aid of superworkstations. It is to these people that we are addressing this appeal for information.
Can you tell us where the bottlenecks are in your current methods of working? At what point does your work cease to be interactive because of limitations in the hardware available to you? Are you being forced to make approximations which you would prefer to avoid? Can you provide us with examples of problems which stretch the capabilities of your workstations?
We need to set up benchmarks for these superworkstations and we believe that by supplying us with this information you will help us define benchmarks which are appropriate to the work of the community.
If you have plans for future work which you believe needs more powerful tools then please let us know. This assessment will be directed at the problem of visualisation of data, which may have been generated on a supercomputer elsewhere. However the most interesting problems will be those which need the increased computing power plus the graphics capability of the superworkstation and so allow us to assess it as an integrated tool.
I recently attended an SERC funded EASE seminar organised by RAL and UMIST. I learned a great deal about the new science of AA. For the uninitiated it can be differentiated from AI by the lack of intelligence. The frontiers are currently being pushed forwards by those interested in IT.
What Engineers have to understand is that BASICally C and FORTRAN and other HLLs are not high enough. What we need are FPLs and LPLs.
In an attempt to be objective, I will allow OOLs and OODBs to speak for themselves. Together they say OOPS! C++ looks promising on the grounds that no devotee of AA would miss the opportunity presented by C + OO = COO. But less of this SMALLTALK ...
UIMS are the stage beyond HCIs. However, do they go far enough? Should we not be developing UIDEs and ISDEs? Current advances in PDLs and DDLs may help this dialogue. CFC and CFTAG are still undecided, however it was nice to see that some engineers were achieving ECSTACY. DGT, SEAS, UoD.
As reported earlier in the Newsletter, the Computers in Teaching Initiative (CTI) Centre for Engineering was established to provide a service to engineering departments around the country. One of its objectives is to establish and maintain an information database containing details of software for use in engineering education.
Questionnaires from 72 departments have been analysed and some of the results for language teaching are summarised in the five accompanying graphs.
The category called Other consists of combined departments or those which did not fit into one of the other categories. Civil includes structural and building. Electrical includes electronics.
Although more departments are teaching Fortran, more student hours are spent learning Pascal. Basic is also popular, but not as many hours are spent on it, perhaps because it is a simpler, and usually interpreted, language. The miscellaneous category, Other, asked the respondent to fill in the name of the language. Four departments are teaching Occam or Occam2, three are teaching an assembler, one department has been teaching Smalltalk, and one, Pop11.
Most departments teach more than one language with the top two languages, Fortran and Basic being paired up most often.
The AI Support for Engineers project is being carried out by the Artificial Intelligence Applications Institute, University of Edinburgh. This activity is funded by the Computing Facilities Committee of the Engineering Board of SERC to provide specialised support for practising engineers who wish to investigate the application of AI in their work.
We are compiling a directory of AI related community to find out what has already been done or is being done. Forms for collecting information were distributed with issues 11 and 13 of the Newsletter. We have now received 105 replies.
Each project in the directory will have a synopsis, details of hardware and software used, a list of publications related to the project and a contact name and address. The directory will also have an index so that information can be found easily.
The first release of the directory will be ready by the end of November. It will be made available free of charge. If you would like to receive a copy please contact us.
As announced in the first issue of AI for Engineers, we are carrying out a series of evaluation of tools for building knowledge-based systems. The first report in the series is an in-depth evaluation of Crystal, a very popular expert system shell for the PC. This report is still available.
A new addition to the series is a survey of the interfacing capabilities of PC-based expert systems shells. These shells are:
All these products claim to access Lotus 1-2- 3, dBase III, and ASCII format files, although the relative importance attached to being able to interface with external data differs between them. In addition to this, these products access packages and files in different ways and require different levels of skill from the user in order to do so. If you would like to receive a copy of this survey please contact us.
The following short courses on AI are provided as part of the EASE programme and they are free of charge to suitable applicants who are working on projects funded by SERC grants. The course schedule in the first six months of 1990 are listed here. These courses are very popular, so early booking is advised.
We are establishing an AI for Engineers Community Club. The Club will provide a focus for the activities of our support programme. The open meeting will identify particular work areas and guide the choice of activity, such as surveyor workshop topic, or case study.
Members of the community can be involved with the Club at any of several different levels, such as receiving and reading the material that will become available as part of the work of the project, attending workshops, or being involved in the technical aspects of the chosen areas. The Club will have an advisory group of around eight people whose responsibility will be to represent the interests and needs of the community and specify work items that will meet those interests and needs. AIAI and the steering group will report back to the community by regular newsletters and at an annual open meeting. Feedback from the community will direct future work.
The Club will be launched on the 6th December at the one day seminar on Why Should Engineers Use AI Techniques, which is to be held at the Rutherford Appleton Laboratory. There will be a number of invited talks on the relevance of AI in, and the application of AI to, different areas of engineering.
Part of the time in the afternoon will be devoted to the launching of the Club and this will only be successful if the views of the community are clearly expressed.. The meeting will set objectives for the Club, ideas for work items and activities will be sought, and the steering group will be formed.
Parallel Computing - Hardware, Software and Techniques Conference at UMIST, 18-20 April 1990
Super Computers/Workstations, February
Computer Algebra for Engineers, March
Visualisation of Engineering Data (workshop),April
Engineering Applications using AI Techniques, date to be arranged
It is proposed to hold a Show in May 1990 at the Turing Institute in Glasgow to encourage and stimulate the robot community with novel and creative ideas.
Why Should Engineers Use AI Techniques, 6 December 1989
The Third Refinement Workshop, 9-11 January 1990.
IEE Vacation School, Software Engineering for Electronic Systems Designers 1-6 April 1990, University of Warwick
This School is intended primarily for engineers involved in the design of electronic systems containing embedded microprocessors and of other real-time systems which are dependent upon software for their functionality and performance.
Eighth British National Conference on Databases (BNCOD-8), Call for Papers 9-11 July 1990, University of York