Traffic simulations are widely used for long- and medium-term forecasting of traffic. Now-a-days, with the growing problem of queues during rush hours, the demand arises for dynamic traffic management and, therefore, short-term forecasting. Apart from the need for new, adapted dynamic assignment models the second important part in this new development is the required computational power. Most commercially available computers are unable to produce an accurate forecasting for the next 5 to 30 min. Within the desired time and budget. Analysis of existing assignment models and their most time consuming part -- shortest path finding -- has shown that the main structure of the models can be parallelized. The use of parallelism thus seems apparent. Several general purpose parallel computers, such as n-cubes, are commercially available. However, apart from being more expensive they loose a large part of their expected performance by the amount of necessary interprocessor communications. Additionally, the programming of such computers has turned out to be more difficult than expected. A simple linear array of typically 16 processors, the so called linear processor array (LPA) is proposed. This one-dimensional parallel computer with high-speed buffered interconnections between each pair of neighbouring processor boards, parallel accessible by both a control board and a general-purpose host computer, forms a transparent concept for the programmer. The optimally configured boards together with the high speed intercommunication allow a cost/performance improvement factor of 100 compared with a minisupercomputer like the convex c1. LPA should be a powerful tool for future developments of online traffic control, route guidance, ramp metering, among other things. This paper appears in Transportation Research Record No. 1306, In-vehicle Information Systems: modelling Traffic Networks and behavioural Considerations 1991.
Abstract