A new group has been set up at the Atlas Centre to provide specialist scientific support to users of the JRCSU Cray X-MP, 48. The group, known as the Advanced Computational Science Group (ACSG), was set up initially under the guidance of Professor P G Burke of Queen's University Belfast during his period of attachment to the Atlas Centre. Nominally the group should consist of four staff members from the Rutherford Appleton Laboratory, one for each of the SERC Board areas plus a computing specialist on attachment from the Central Computing Division. Currently the membership of the group consists of:
At present there is no representative from the Astronomy and Planetary Science Board but it is hoped to appoint someone fairly soon.
The purpose of this group is to provide specialist support beyond that normally available from the User Support group; the members of ACSG have come from within RAL, from Laser Division, High Energy Physics and from Central Computing. Our expertise includes plasma physics, computational fluid dynamics, non-linear optics, Lattice Gauge methods and linear algebra, but we are all anxious to help as wide a user community as possible. If time permits, we hope to carry out some research programmes ourselves, in collaboration with universities and industry. We are currently pursuing contacts in research areas which might benefit from use of the Cray X-MP, but which have not yet ventured into supercomputing. In Engineering Board two areas of current activity are Computational Fluid Dynamics and structural analysis and we hope soon to pursue semiconductor device modelling and semiconductor physics. Within Science Board contacts are being made with the DNA database community and in Astrophysics with radio astronomers.
We realise that all areas cannot be covered at once and although we are few in number we are always willing to talk to potential supercomputer users and offer advice. Users are welcome to visit the Atlas Centre or if appropriate we will travel to you.
Outside of SERC, it is hoped that both NERC and MRC will soon appoint people to work at the Atlas Centre in support of their Councils' programmes. Users from ESRC are welcome to approach members of ACSG as they see appropriate, but our relevant experience may be somewhat limited!
It has been possible to calculate the physical properties of solids using parameter free total energy methods for several years. The calculated values of interatomic distances and bulk moduli are generally within a few percent of their experimental values. Some calculations have predicted results that were later verified experimentally. The calculations allow us to examine the structures of surfaces and defects that affect the electrical and mechanical properties of materials. As the capability of computers increases the method can be applied to increasingly complex systems.
When the total energy method was first developed, direct matrix diagonalisation was used to solve the equations for the electronic states. As the number of atoms in the unit cell increases and more basis states are employed in the calculation, standard direct matrix diagonalisation techniques require a prohibitively long computational time. This problem is currently being overcome by the use of iterative diagonalisation techniques. The calculations use up to 10,000 basis states and the matrices are not sparse.
The Cray X-MP is ideally suited to these calculations. Its large memory and rapid execution speed are beneficial to all users. The particular features of the Atlas Centre's Cray that are valuable in this study are the Solid State Storage Device (SSD) and the large backing storage. The calculations have to be repeated at a number of points to obtain an average over the system. The SSD greatly reduces the amount of time required to move each set of wavefunctions into and out of main memory. At the end of a set of iterations the wavefunctions have to be stored so that the calculation can be continued later if necessary or so that these wavefunctions can be used as the starting point for a new calculation. The Masstor M860 Cartridge Storage allows these datasets to be stored as though online and removes the necessity for continuous transfer of datasets between storage and magnetic tapes.
The programs are written to exploit the vector computational mode of the Cray wherever possible. However, the scatter gather hardware allows several further parts of the program to execute in vector mode. Users familiar with other vector machines will probably have appreciated the ease of getting vector code running efficiently on the Cray without the need to call special subroutines.
Access to the Cray X-MP from Cambridge is by means of the Joint Academic Network (JANET). At either end, there are gateways to and from JANET. It has been quite common to experience breaks in the communication connections, especially during normal working hours. The exact reasons for these are being investigated. Nevertheless, these events outside a user's control are extremely frustrating. Remote users of the Cray X-MP facility would welcome greater certainty that a connection, once established, would remain uninterrupted for longer periods.
As well as being one of the popular weapons of the "Star Wars" scenario. X-ray lasers are a potentially useful weapon in many areas of fundamental research and are being studied both theoretically and experimentally on a worldwide scale. The essential feature of a research X-ray laser would not be particularly intense beams, but rather a coherent beam of X-rays with only enough intensity to activate detectors in a reasonably short time. Most of the currently envisaged applications are based on 3D imaging using X-rays to penetrate either living cell tissue or solid samples. To be of use in biological research the X-ray wavelength should lie in the 'water window' between about 8nm and 2nm where the differential absorption of carbon (protein) and oxygen (water) gives good contrast in live cells.
Progress to shorter wavelength lasers is hindered by the fact that the power required to "pump" the laser increases roughly as the inverse fourth power of the wavelength, and the current lack of high reflectivity X-ray mirrors requires a high single pass gain. In practice X-ray lasers need a large optical laser to provide the intense pump power and this limits the number of laboratories able to carry out X-ray laser research.
In the UK the large optical laser is provided by the SERC Central Laser facility and is used by university research groups from Hull, Imperial College and Queens University Belfast. The experiments are supported by theoretical work at Hull, QUB, RAL and St Andrews.
X-ray laser action occurs in a variety of ionised atomic species under specific histories of density and temperature. As the transitions progress to shorter wavelengths the trend is to move into higher Z ions and at about Z = 35 relativistic effects in the ions become important and the electron-electron coupling effects become increasingly complex. Ab initio atomic structure calculations can be compared with observed spectra and serve to give confidence in calculations of radiative and collisional rates. Full R-matrix calculations of rates in Neon-like Selenium and Yttrium are beyond the memory limits of the Cray-1 and require the memory and SSD capabilities of the X-MP to handle the coupling matrices.
Calculating the intrinsic atomic properties is only the first part of the job since it is then necessary to simulate the thermal and hydrodynamic behaviour of real laser targets in order to guide the experimenters into the optimal parameter range. These calculations require some degree of symmetry to be forced on the problem in order to be tractable, one dimensional calculations are feasible on the IBM 3081 while 2D simulations require the full Cray performance. The fluid codes are largely vectorisable and some work in this area has been done at Cray Research. The full X-ray laser problem requires solving a set of identical atomic physics rate equations in each of the cells of the fluid model, this is exactly the requirement to make optimal use of the multitasking features of the X-MP/48 since apart from slight variations due to rate of convergence the individual cell calculations take the same time to execute. With the current charging algorithm this is not particularly advantageous.
The atomic physics and hydrodynamic models currently exist independently and grant approval to construct and use the coupled model has just been obtained.
Output to videotape appears a natural way to deal with the massive amounts of data processed and generated by the Cray X-MP/48 at RAL. RAL Graphics Group are producing plans for a video output facility, similar in function to that of an animation film recorder. While plans for this are not yet complete (and funds for the project are not yet certain), many users have shown great interest in the idea. This article indicates the structure of the proposed video facility and describes some ways in which it is expected to be used.
The aim of the system will be to allow the production of sequences of video frames on videotape, in the same sort of way as a film recorder assembles frames on film. The output will be recorded on a high-band U-matic recorder; these are extensively used for broadcast work and out-perform all domestic formats.
The output distributed to users can be this master tape or may be a copy onto whichever format is required - low-band U-matic and VHS (HQ) would certainly be provided; Beta and video-8 could be added if there was sufficient demand.
The possible content of the videotape would be entirely up to the user. All the current graphics systems (GKS, GINO-F, GHOST-80, NAG) would have routes to the video facility. Other packages (like Movie-BYU) are also being considered, since the animation aspect of video output can only be used effectively with good software support. The recording facility itself would provide for frames to be repeated onto tape, allowing programmed slow-motion.
The recording system would be connected to a graphics controller which, in most cases, would be responsible for converting graphics commands into pictures in a number of frame stores. This controller would in turn be connected to the Cray, via the Hyperchannel, micro VAX or possibly (in the future) directly to a high speed channel. An animation controller would control the operation of the master recorder, the slave recorders and the switching between frame stores.
The output will be full colour: the framestores will have 24 bits per pixel, providing 8 bits each for red, green and blue components.
The throughput of the system is determined by the mechanical requirements of the recorder. To record a small group of frames, the recorder must rewind a little, start to scan the tape in READ mode, switch to WRITE mode for the required number of frames at the required point and then stop graciously. On the form of recorder being considered, this sequence takes something like 15 seconds. Video output is displayed at 25 frames/second, so 1 second of output would take 25 × 15 seconds (or 6 minutes) to record if only one frame was recorded on each burst. This can be improved if multiple frames are recorded each time. One way to achieve this is by providing multiple framestores; it also occurs when the user wants a frame repeated, as in the "slow-motion" mentioned above.
In most cases, the graphics controller in the video system will be responsible for generating the graphics. In this case, the amount of graphical data sent from the Cray to the video system will typically not be large and the only delay will be m writing the videotape. In some cases, the majority of the graphical work may be done on the Cray and in this case the data sent from the Cray to the video system may be quite large (768 × 576 × 24 bits = 1.26 Mbytes). Transmission delays through the links between the Cray and the video system may then lengthen the recording time per frame by a factor of 2 or 3.
The system being proposed will allow the production of video output of very high quality, easily and conveniently. If four framestores are included in the system and average jobs repeat each frame once (a common animation technique), a one minute video will take 3/4 hour to record.
To improve on this performance, more framestores can be added but this would also require an increase in the complexity of the animation controller. If higher volumes of video output were required, the master recorder would have to be replaced by a much more expensive version (£80,000 instead of £9,000). It would also be likely that a direct connection to the Cray would be essential if the recorder was not to be starved of pictures.
