Jump Over Left Menu
Issue 12: January 1991
Cray Gigaflop Awards for 1990
Supercomputer users from the UK were well represented at the 1990 Cray Gigaflop awards for the highest sustained computing performance from real application programs. The competition is open to all Cray customers, but it is a sad reflection on the investment in academic supercomputing in the UK that the UK competitors all had to use Cray Y-MP machines in the USA to achieve competitive performances.
The threshold for this year's entries was increased to 1.5 Gflops, which clearly implies that the entrants have achieved a high degree of vector and parallel processing efficiency in their codes. It is a requirement that the codes have to be self contained and have to tackle a complete application problem rather than just be a computational kernel from a larger package.
Stephen Wilson from RAL, Vic Saunders from SERC Daresbury Laboratory and David Moncrieff from Florida State University came in third overall with a performance of 2.229 Gflops (just 0.078 Gflops separated the first three places) with a development of their work on the Atlas Cray X-MP for the calculation of fourth order many body perturbation corrections to molecular energies. Their code makes good use of the library routines for matrix multiplication to get best vector performance and uses a dynamic allocation of work to CPUs to get the best load balancing and overall performance.
Alan Davies and Roman Grzonka of the NERC Proudman Laboratory achieved 1.768 Gflops for their three dimensional "shallow sea" oceanographic model running on an eight processor Y-MP, the same model running on the four processor X-MP at Atlas achieves 0.5 Gflops and was described in an earlier issue of FLAGSHIP. Their model uses a geometrical decomposition with vectorisation in one of the horizontal directions for best vector length and parallel computation of the vertical motions which are handled by a modal decomposition for best accuracy.
Walter Temmerman and Zdislawa Szotek of Daresbury, with collaborators from the USA, had their band structure calculations of Perovskite based high Tc superconductors running at 1.509 Gflops, also on an eight processor Y-MP. The problem is essentially one of inverting an enormous number of matrices, but each matrix is only about 80 by 80.
Outside the academic community, there were Gigaflop awards to the UK for theoretical chemistry from the Royal Armament Research and Development Establishment and for Computational Fluid Dynamics work from Nuclear Electric at the Berkeley Nuclear Laboratory. This year it was particularly noticeable that the proportion of entries from industry was much higher than in earlier contests.
Perhaps at some time in the future we will be able to achieve these excellent performance figures on machines in the UK rather than having to go the USA for the ultimate in hardware performance.
Roger Evans, Head of Advanced Research Computing Unit
For fifteen years communications on the RAL site and within the United Kingdom has been based on the familiar X.25 based Coloured Book services which have served us well. In the late 70s this gave us several years lead over other countries. So why are we looking at the so called IP communications services?
IP (Internet Protocol) is, like Coloured Books, shorthand for a whole set of protocols providing a variety of customer services. IP is actually the lowest level of protocol. Over it there are a number of protocols such as TCP (Transmission Control Protocol) which makes sure there is a reliable data path between the two ends. Over TCP runs FTP (File Transfer Protocol), Telnet for interactive use, SMTP (Simple Mail Transfer Protocol), and X Windows for advanced interactive use. NFS (Network File Server), which runs over a different transmission control protocol, provides file access. There are other protocols of less interest.
Although the objectives of Coloured Book and IP protocols are much the same, they work rather differently. With X.25 a call or virtual circuit is established end to end and data is sent across this reliable connection. This is known as a "connection network". With IP, each packet of data is launched into the network together with its destination and it may or may not arrive. Higher level protocols, such as TCP, sort out the mess and provide a reliable service. This is a "connectionless network". Pundits will argue the benefits and drawbacks of each technology but although the way things work is different, the arguments are much more to do with the popularity, quality of implementations, high level services, costs, and political issues.
A couple of years ago we put in a high speed local area network. This was very carefully designed and implemented. It has proved very reliable and traffic is growing. We want, over a few years, to get rid of the relatively slow network we have relied on and put as much traffic as we can on our new fast 10 Mbits/sec network.
We were expecting to run the current X.25 Coloured Book services over the Ethernet and the standards for this had been developed by the JNT. RAL, together with ULCC London, and DEC ran a successful project to check that all this worked. However, we decided at an early date that we would allow any sorts of traffic to use the network as long as it did not prejudice reliability and performance.
The results were interesting. It is difficult to measure traffic accurately, since the network consists of a number of sub-networks. However, we believe the biggest use, representing 75% of traffic, has turned out to be DECnet. The reason is that DEC has developed DECnet to utilise Ethernets and with the hundred or so VAXs on site the temptation to use its excellent quality between VAXs was overwhelming. The second most popular set of protocols has proved to be IP with 20%. The initial popularity has been caused by the UNIX machines for which IP is the natural choice. Most sites using UNIX worldwide use IP and as a result the implementations have become highly reliable and well matched to UNIX systems. In fact there are services such as NFS and X Windows which are unavailable with other protocols. Since then IBM PCs, IBM VM/CMS, and VAX/VMS systems have been blessed with good IP implementations.
Experience has shown that IP systems are very easy to mount and use, in contrast to such systems as SNA and to a lesser extent X.25 and Coloured Books where the systems are rather more complex. A further attraction is that these systems are very widely used and are therefore relatively bug-free. In addition, there are often competing products with attractive bells and whistles from which to choose. Even more attractive in these hard times, we find that some products are free. Most people will be aware that the IP protocols were a development from the USA ARPA project and have thus been well supported by the mighty Dollar.
Within the site, NFS is popular and the only way file access can be achieved. Also within the site X Windows provides an important service as high quality terminal access from desk top workstations becomes affordable. There is also a demand for X Windows from off site for access to the CRAY. These require an IP network. Although these services are freely available on site, only a few pilot connections have been made off site. There are hopes that further connections will be made. Once such a service is available then users may well wish to use other protocols, even where there are adequate Coloured Book ones, for reasons of convenience or functionality. Choice is a nice thing but it may well have a tendency to complicate networking and undermine some of the UK networking policies.
The principal network within the USA for academic traffic is now based on IP. IP has recently become popular within Europe for inter-site traffic and there are moves to try and co-ordinate this activity. With the emergence of such a large combined USA/Europe network there are great attractions for UK users to have good access to it. This can be achieved by gatewaying Coloured Book protocols to IP or by allowing IP to penetrate the UK. Gateways are well known to be a cause of loss of quality and frustration. Allowing IP to penetrate way well upset the networking strategy of the UK. A group has recently been set up to advise the JNT on this problem and they have now reported and recommended some use of IP across the JANET X.25 network.
With any network technology there are many ancillary problems: how to manage it, providing directory services; how to link one site to another; how to ensure security; and many other questions. In the Department we have been looking at these questions so that we are well poised to take advantage of any opportunities that arise in this area.
Paul Bryant,Head of Communications & Small Systems Group
RAL-CGM now available
The RAL-CGM system is now available to the UK academic community. It allows the conversion between any of the three standard encodings of Computer Graphics Metafile (CGM) and also the viewing and plotting of CGMs on a number of devices.
As a result of encouragement by AGOCG (the Advisory Group On Computer Graphics), tools for generating, converting, viewing and plotting Computer Graphics Metafiles (CGMs) have been developed. This system, called RAL-CGM, has been developed by a team at the Rutherford Appleton Laboratory from a system written by them over the last two years. RAL-CGM is now being made available to the UK academic community (Research Councils, Universities and Polytechnics) free of charge.
What RAL-CGM contains
The system, written entirely in C, consists of a main program and three groups of modules:
- input modules, one per CGM standard encoding;
- output modules, one per CGM standard encoding;
- interpreter modules for various devices, currently including PostScript and X Windows.
In addition, there are a number of utility modules that are used by the main modules.
The RAL-CGM system allows CGMs of any encoding to be converted to any other encoding. It also allows any CGM to be viewed or plotted on any supported device. For example, a CGM may be converted into PostScript format so that it can be plotted on a PostScript plotter. CGMs may be viewed on any X-windows device, the user being able to use buttons or the keyboard to view pictures in random order.
This first release (1.00) contains a limited number of interpreter modules, and others such as HPGL and Tektronix 42xx are being developed for the second release. RAL-CGM has been installed on many systems, and release 1.00 is configured for use on
- Sun 3 and 4 machines under SunOS/4
- VAX under VMS
- Cray under Unicos
- IBM under VM/CMS (IBM c/370 and
- Waterloo compilers)
- Silicon Graphics Iris 3130 under UNIX V.2
- Primagraphics Topaz under UNIX V.3 and GNU C compiler
Work is already in hand towards a PC (DOS 3, Microsoft C) version that will be part of the second release.
What RAL-CGM does
CGMs can be converted from one encoding to another, output as PostScript files or viewed under X Windows. The X Windows viewer is based entirely on Xlib and so is not restricted to any one system, such as DECwindows. RAL-CGM is distributed in source form, complete with installation and tailoring command files and help files for each system. It is part of the design aim of the system that people wishing to write interpreters for other devices will be able to do so; a skeleton interpreter is included. Enthusiastic implementers should however note that the interfaces to the utility modules - which provide emulation of filled areas, text and font facilities and the complex primitives - will not be stabilized until the second release.
RAL-CGM is being distributed by the following routes:
- UNIX systems
- via the Netlib software distribution system at the University of Kent;
- VAX/VMS systems
- from RAL as detailed below;
- IBM VM/CMS systems
- from RAL by contacting Roy Platon at RAL
Obtaining a copy for UNIX systems
RAL-CGM for UNIX can be obtained either on a tape or over the network. To obtain it over the network, send to firstname.lastname@example.org a mail message whose body contains a single line:
send encoded.tar from ralcgm
You should receive a number of big files and a single small file which contains a script and instructions to glue the separate files back into a large file (the size of which will be around 1.2 Mb). Name this, say, tarfile, then
should produce a file called tarfile.z. (Warning: some UNIX systems do not have the uudecode command. It is hoped, however, that every installation will have at least one machine - Sun for instance - that does).
zcat tarfile.z > ralcgm.tar
will generate the original tar file (the size of which will be around 3.5 Mb), which can be split into directories and files in the normal way, i.e.
- Create a directory with a suitable name, for instance "ralcgm".
- Copy the ralcgm tarfile to that directory.
- cd to that directory.
- Issue the following command:
tar xvof ralcgm.tar
- This will unravel the tar file into the correct directory structure. The top directory will contain a file called README which gives instructions on how to complete the installation.
To obtain RAL-CGM on a tape, send a QIC24 cartridge (any length is long enough) to:
Tim Hopkins, Computing Laboratory, University of Kent
To load the software:
- Create a directory with a suitable name, for instance "ralcgm".
- cd to that directory.
- Issue the following command (device names are system dependent.
This is for the Sun.
tar xvof /dev/rst8
- This will create the correct directory structure. The top directory will contain a file called README which gives instructions on how to complete the installation.
Obtaining a copy for VAX/VMS systems
RAL-CGM for VAX/VMS can be obtained over the network with the following VAX command:
$TRANSFER/CODE=FAST - UK.AC.RL.VE::RALCGMl.BCK - RALCGM1.BCK RALCGM DISTRIBUTION
This will fetch a copy of the VAX/VMS saveset called RALCGMl.BCK from a VAX at Rutherford.
To install RAL-CGM from the saveset:
- create a directory with a suitable name such as RALCGM;
- copy the RALCGMl.BCK file to that directory;
- set default into the directory;
- issue the following command:
$BACKUP/LOG RALCGMl.BCK/SAVE/ -SELECT=[RALCGM. ..]*.* [...]This will unravel the saveset into the correct directory structure. The top directory will contain a file called README, which gives instructions on how to complete the installation.
RAL-CGM version 1 uses the UNIX/C language command line interface so options are preceded with a minus sign, rather than in the normal VAX /option form. See the help library for more details of the options and the format to use.