Brian J. N. Wylie - Selected Abstracts

Recent work

Earlier work

• PARAMICS - Moving Vehicles on the Connection Machine, Gordon D. B. Cameron, Brian J. N. Wylie & David McArthur, in proceedings Supercomputing'94 (Washington DC, USA), pp. 291-300, IEEE Computer Society Press, November 1994. [ISSN: 1063-9535]
  
  
  PARAMICS is a parallel microscopic traffic simulator which is, to our knowledge, the most powerful of its type in the world. The simulator can model around 200,000 vehicles on around 7,000 roads (taken from real road traffic network data) at faster than `real-time' rates, making use of a 16k-processor TMC Connection Machine CM-200 for the simulation aspect. The project aims to make available to road network planners a new range of tools, and demonstrates that use of high-performance computing in real applications is possible and worthwhile, while yielding important and interesting research results.
  
  [SC94.ps.gz - 10 pages, 366kB (1150kB when uncompressed)]

  ---

• PARAMICS: Parallel Microscopic Traffic Simulator, Brian J. N. Wylie, Gordon D. B. Cameron, Matthew D. White, Mark A. Smith & David McArthur, Edinburgh Parallel Computing Centre Technical Report EPCC-TR94-01, January 1994. In collection Science on the Connection Machine System, J.-M. Alimi, A. Serna & H. Scholl (eds.), pp. 69-79, Thinking Machines Corporation (Cambridge, MA, USA), 1995.
  
  
  PARAMICS is a state-of-the-art, parallel `microscopic' traffic simulator, developed in a collaboration between EPCC and SIAS Ltd., for simulations of traffic congestion in large-scale road networks. Individual vehicle and driver characteristics are modelled, responding to perceived road and junction conditions. To allow traffic planners to predict and attempt to rectify potential congestion problems, within the trunk road network of central Scotland at peak times, requires the real-time simulation of some 200,000 vehicles. The simulator is being implemented in the data-parallel C* language on the CM-200 massively parallel computer at EPCC, which has the potential to deliver the substantial computing demands required. From a detailed description of the Scottish road network, and the traffic within it, vehicles' movements and flow summaries can be displayed graphically, highlighting the formation of areas of congestion and allowing the probable causes to be identified.
  
  [EPCC-TR94-01.ps.gz - 8 pages, 335kB (1039kB when uncompressed)]

  ---

• Meteorological Data Assimilation Migration to Massively Parallel Systems, Brian J. N. Wylie & Patrick D. Surry, Invited presentation at the Spring'93 Cray User Group Conference (Montreux, Switzerland), March 1993, also available as a short paper.
  
  
  The observation data assimilation portion of the UK Meteorological Office unified weather forecasting and climate modelling code was studied by EPCC with a view to parallelisation. In the first instance, the original Fortran-77 source code (containing vector optimisations for the Cray Y-MP) was transformed via the array-language subset of Fortran-90 to CM-Fortran for the CM-200. Detailed profiling and analyses suggested performance benefits were possible from conceptually simple -- yet otherwise sophisticated -- algorithmic modifications, which were subsequently implemented. Significantly improved performance and scalability were thereby achieved from efficient implementation of effective massively parallel schemes.
  
  [CUG93s.ps.gz - 6 pages, 208kB (1775kB when uncompressed)]

  ---

• High Performance Fortran: A Perspective, Brian J. N. Wylie, Michael G. Norman & Lyndon J. Clarke, Edinburgh Parallel Computing Centre Technical Note EPCC-TN92-05, May 1992. Abstracted in the Serlin Report on Parallel Processing, July 1992, and ACM SIGPLAN Fortran Forum, 11(3), September 1992.
  
  
  High Performance Fortran (HPF) is an international effort, with the enthusiastic support of a number of parallel computer vendors and other interested parties, to define standard extensions to Fortran. It promises an attractive abstraction, with high-performance potential, for Fortran programs on parallel (and massively-parallel) computers, along with transparent portability to `traditional' serial and vector computers. This report is intended to provide background information on the HPF initiative, the technical aspects thereof, as well as our impressions of why this activity is progressing with such apparent enthusiasm.
  The contention is that the array-parallel operations of Fortran-90, coupled with additional directives providing information about appropriate decompositions and distributions, can enable compilers with sophisticated dependence analysis and knowledge of the MIMD parallelisation techniques to produce native message-passing code for efficient and effective utilisation of distributed-memory parallel computers. Data distribution, semaphore and message-passing code would automatically be generated where required.
  In addition to providing a perspective of the developing HPF situation in the United States, European activities and involvement are also addressed. While the HPF Forum meetings have been held in the US (mostly hosted by Rice University, Texas), a major part of the definition and discussion process has been achieved through electronic mailing lists accessible to all interested parties. There have also been a number of workshops in Europe to address current and potential HPF developments.
  Early versions of HPF are, however, likely to have only limited application areas, since they will have very restricted MIMD support, and neither unstructured data objects nor sophisticated load-balancing support. Such developments are likely in the future as continuing HPF language extensions. A subset of Fortran-90 has been chosen as the base for HPF development, and while Fortran-77 (also being a subset of Fortran-90) will be partially supported, `high performance' support for Fortran-77 is unlikely due to largely insurmountable problems with the way the language has been abused in practice to date. There are also worries that as yet no consideration has been made of providing HPF semantics in C.
  The draft language specification is due for presentation in November, with the official release due for early '93. Vendors have announced that they will have compilers providing partial HPF support available Q1'93, with full support (for the initial proposal) later in the year. It seems clear that the considerable momentum which HPF possesses, coupled with its attractiveness to customers of parallel computing, will make it an increasing force to be reckoned with in the numerical and scientific computing market.
  
  [EPCC-TN92-05.ps.gz - 23 pages, 60kB (193kB when uncompressed)]

  ---

• Parallelisation of the Unified Model Data Assimilation Scheme, Kenneth A. Hawick, R. Stuart Bell, Alan Dickinson, Patrick D. Surry & Brian J. N. Wylie, in proceedings 5th Workshop on the Use of Parallel Processors in Meteorology (ECMWF, Reading), November 1992.
  
  
  We describe the data assimilation scheme employed in the UK Meteorological Office's Unified Model and a number of schemes for its efficient implementation on massively parallel computer systems. A simple port to the Connection Machine CM-200 system was carried out and for which detailed timing and profiling figures are given. A number of alternative implementation algorithms were investigated and are also described.
  An algorithmic inversion of the vectorised implementation allowed the processors to work on the parallel update of the influences of observation data on all the model grid points together rather than working sequentially through a vector of model grid points. A further improvement to the implementation required the packing of multiple observations into vectors of appropriate length to the number of processors. Finally, the model grid was statically divided into latitudinal sections to allow a better balance of processor work load.
  Implementations of these schemes were carried out using High Performance Fortran style directives and Fortran-90 array constructs on the Connection Machine CM-200 system.
  Profiling measurements were made for typical runs at both weather prediction and climate resolutions. A number of potentially better implementation schemes are suggested for other parallel computer systems. We report on implementations in progress for other parallel processing programming paradigms.

  ---

• High Performance Fortran Migration via CM-Fortran, Patrick D. Surry & Brian J. N. Wylie, Edinburgh Parallel Computing Centre Technical Note EPCC-TN93-01, January 1993.
  
  
  Lists the CMF features which should be used since they have the same functionality as HPF (albeit with a slightly different syntax), along with a list of features to avoid (or reproduce with source code) since they will not be widely supported in initial HPF compiler releases.
  
  [EPCC-TN93-01.ps.gz - 10 pages, 31kB (94kB when uncompressed)]

  ---

• TaX - Transputer Activity Tachometer: Experience with a Simple Multicomputer Software Monitoring and Concurrent Visualization System, Brian J. N. Wylie, Edinburgh Concurrent Supercomputer Project Technical Note ECSP-TN-46, August 1990.
  
  
  It has generally been difficult to determine exactly which processes within a particular distributed computation are responsible for slowing down the whole computation through being overloaded with work. There are also typically extremely complex inter-process interactions. One would also like to know precisely the state of each process and how far it has progressed with its particular task. The ``Pulsars'' demonstration program is an exception to these restrictions, but it may be unique in that all the relevant information is visible in the program display -- in general, this will not be possible.
  When monitoring an application, it is desirable that the subject should not be unduly perturbed by the actions of the monitor. Hardware techniques for interference-free monitoring are an obvious solution, and several are being developed. They do tend to be inherently expensive and inflexible, however. Presentation and interpretation of the copious data from fine monitors additionally remains a challenging activity.
  In the absence of such a tool, the needs of the users of the transputer-based Computing Surfaces at Edinburgh were great, such that any monitoring tool may well prove beneficial. There were also a considerable body of skeptics who believed that a fully-software implementation would have such a perturbing effect to render it useless.
  A simple CPU load monitor and run-time visualisation system was developed, using the tools to hand, and from preliminary experiments, on actual applications, it has been demonstrated that the benefits far outweigh the drawbacks. Useful information is achievable, and can be made readily available for every processor (and possibly process) involved in a computation. In the absence of effective tools, education (and experience) are invaluable when programming distributed systems, and the dramatic real-time display of processor activity, in conjunction with the subject application, is very powerful in this regard.
  
  [ECSP-TN-46.ps.gz - 9 pages, 17kB (47kB when uncompressed)]

  ---

• PULSARS: Dynamic Process Contention Modelling & Monitoring, Brian J. N. Wylie, Edinburgh Concurrent Supercomputer Project Technical Note ECSP-TN-36, January 1990.
  
  
  PULSARS is a demonstration of dynamic processes, which start executing a task, may be halted through contention for resource allocated to other processes (or even have to abort and restart), and which terminate on completion of that task. The progress of each process simultaneously executing is continually graphically displayed, and a (necessarily incomplete) monitor of the state of execution/suspension of each process also. This model is analogous to the problems of contention and possible subsequent deadlocking of processes in shared-memory multicomputers.
  
  [ECSP-TN-36.ps.gz - 10 pages, 21kB (61kB when uncompressed)]

  ---

• Application of Two-Dimensional Cellular Automaton Lattice-Gas Models to the Simulation of Hydrodynamics, Brian J. N. Wylie, PhD thesis, University of Edinburgh, October 1990.
  
  
  Hydrodynamic equations are notoriously difficult to solve, both analytically and computationally, therefore less complicated methods, drawing on the power of cellular automata and lattice-gases are introduced. Their combined ability to capture the fundamental properties of fluid dynamics in an inherently simple manner is discussed. The FHP7 cellular automaton lattice-gas model of Frisch, Hasslacher and Pomeau, which will form the basis for the subsequent simulations, is described in detail, with a more general covering of the associated models.
  The scalable and flexible computational power of the transputer-based ECS multicomputer, and how this may be applied to the lattice-gas simulations at hand is addressed. The distributed multiprocessor architecture provides unique challenges, such that the implementation might achieve its potential. It is found that a straightforward one-dimensional geometric decomposition of the lattice, in conjunction with the loosely-synchronous nature of the distributed update, provides a natural load-balancing, and highly scalable efficiency.
  Visualisation of the development of the hydrodynamic features captured by the simulations, such that their content may be clearly extracted is also addressed. Many interesting transient and dynamic features, often occuring on time-scales which make their analysis by other methods difficult, are easily identified. Those occasionally found to be the result of artifacts, perhaps in the initialisation of the simulation are quickly identified, such that the simulations may be refined.
  Elementary static systems and flows are designed, such that the ability of the FHP7 lattice-gas to model incompressible hydrodynamics, and its multicomputer implementation, are verified against the theoretically and experimentally expected behaviour. Subsequently, more complex flow configurations involving obstructions and jets, generally beyond the limits of current analytic techniques are constructed, and found to qualitatively match experimental visualisations.
  No lattice-gases are currently known to accurately model compressible fluid dynamics completely, and the ultimate cause of this limitation still requires clarification. The behaviour of the FHP7 lattice-gas, in regimes where compressibility effects are expected to be important, is investigated with the aim of identifying those aspects of its microdynamics which cause breakdown of its macrodynamics.
  
  [PhD_thesis.ps.gz - 190 pages, 750kB (3394kB when uncompressed); PhD_thesis.pdf (1563kB);
    PhD_plates.pdf (1096kB); plates_gif.tar - GIF-format screen images in a Unix tar file, 1232kB]

  ---

• Cellular Automaton Lattice-Gas Hydrodynamics on a Parallel Supercomputer, Brian J. N. Wylie, in Proceedings 8th Technical Meeting of the occam User Group (Sheffield, England), pp. 205-214, IOS Publishers (Amsterdam), March 1988. [ISBN: 90-5199-0020-4]
  
  
  Computational fluid dynamics, one of the major consumers of supercomputing resources, embraces a wide range of disciplines from the relatively simple systems which can be solved theoretically to the highly complex systems currently required for industrial experimentation. Cellular automata and lattice-gas models of fluid systems, are greatly simplified versions of real systems, yet within certain limitations they have been shown to accurately model expected behaviour, while being inherently parallelisable. Preliminary investigations of cellular automaton lattice-gas hydrodynamics, at the University of Edinburgh on a MIMD Meiko Computing Surface (coded in occam), have shown some of the particular aptitudes of the model.