Accelerating Traffic Models Using GPU-Based Technology
Richard Bradley, Atkins, Roger Himlin, Highways England, Ian Wright, Atkins
Presenting research into large scale transport model speed improvement, tested in the SATURN modelling software. Findings show how GPU technology can be used in macro traffic assignment models & whether it can provide a step change in model runtime.
Highways England is responsible for the Strategic Road Network in England and is planning numerous major highway investments. Highways England needs to understand the individual and combined Return on Investment of schemes, and for this purpose is building five regional transport models that provide sufficient detail to appraise the economic impacts for complete journeys, including long distance movements of people and freight. Within the same model they also need to understand sufficient local detail to ensure the modelling is appropriately accurate to represent capacity restraints and local appraisal.
Highways England has chosen the SATURN highway modelling software as the most appropriate tool for this level of representation of the English highway network. The scale of the regions, combined with the outcome requirements, means that a single model run, including three time periods, is expected to take up to two days each. If for each region there are 20 schemes to be tested, and in two forecast years and three growth scenarios, that’s potentially 1,200 model run-days for one analysis iteration.
Highways England therefore needed to consider ways of reducing this runtime overhead. Modelling software has started to embrace parallelisation techniques in recent years with the introduction of Central Processing Unit (CPU) multi-threading, and SATURN has already implemented CPU multi-core parallel computing. However, whilst each CPU thread has significant ‘clock’ speed they are limited in number with a typical modelling budget PC having a maximum of 20 threads. With the concept of parallel computing within assignment models already established, Highway England turned to Graphics Processing Unit (GPU) technology as a potential way of creating a step-change in traffic model runtimes. GPU parallel computing has a number of differences to the equivalent CPU computing and, with a typical modelling budget PC, provides upwards of 6 TFLOPS (Terra Floating Point Operations Per Second) of theoretical performance in contrast with roughly 0.1 TFLOPS available in modern i7 processors.
Investigations into CPU multi-core execution have highlighted that the areas of the highway modelling process that take the more significant computer resources, in terms of memory and compute time, are associated with the assignment of traffic to the model network, and specifically focussed on tree building algorithms and subsequent skimming. The processes used in SATURN are typical of macroscopic traffic modelling software and therefore any solution is common across the modelling industry, with potential implementation of SATURN improvements also applicable to other packages.
Highways England therefore entered into a collaboration agreement with Atkins (as one of the co-owners of SATURN), the University of Sheffield, Transport Systems Catapult and The Hartree Centre, a department of the Science and Technology Facilities Council. This team is researching the introduction of GPU technology to macroscopic traffic assignment modelling with a view to providing the step change in runtimes desirable for the modelling industry as a whole and required for the regional transport models.
In its first practical implementation, the project is expected to provide a substantially faster assignment ‘engine’ for SATURN models of the future. This step-change in speed will provide more manageable runtimes for large scale models but it is hoped that faster runtimes can allow additional future year scenarios to be tested to enable a deeper exploration of investment uncertainty and network resilience. Furthermore, increased traveller choice introduced through Intelligent Mobility will place even greater emphasis on model performance and GPU technology has potential use in other areas of modelling software, such as demand models.
This paper will specifically present the findings of the current phase of the GPU development. It will provide insight into the development process and benchmark performance of introducing GPU technology to macroscopic traffic models. This process has included testing alternative tree building algorithms, more suited to parallelisation, and understanding how PC memory may be best managed within complex modelling assignment looping processes and PC hardware constraints. The project is also exploring the performance of the different types of GPU hardware available across a range of typical PC modelling budgets. The project will conclude as to whether GPU technology can practically deliver the step-change needed to enable the development of more ambitious large scale transport models, and how the scale achievable might influence the use of models in the future.
Association for European Transport