• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 1
  • Tagged with
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Train Re-scheduling : A Massively Parallel Approach Using CUDA

Petersson, Anton January 2015 (has links)
Context. Train re-scheduling during disturbances is a time-consuming task. Modified schedules need to be found, so that trains can meet in suitable locations and delays minimized. Domino effects are difficult to manage. Commercial optimization software has been shown to find optimal solutions, but modied schedules need to be found quickly. Therefore, greedy depth-first algorithms have been designed to find solutions within a limited time-frame. Modern GPUs have a high computational capacity, and have become easier to use for computations unrelated to computer graphics with the development of technologies such as CUDA and OpenCL. Objectives. We explore the feasibility of running a re-scheduling algorithm developed specifically for this problem on a GPU using the CUDA toolkit. The main objective is to find a way of exploiting the computational capacity of modern GPUs to find better re-scheduling solutions within a limited time-frame. Methods. We develop and adapt a sequential algorithm for use on a GPU and run multiple experiments using 16 disturbance scenarios on the single-tracked iron ore line in northern Sweden. Results. Our implementation succeeds in finding re-scheduling solutions without conflicts for all 16 scenarios. The algorithm visits on average 7 times more nodes per time unit than the sequential CPU algorithm when branching at depth 50, and 4 times more when branching at depth 200. Conclusions. The computational performance of our parallel algorithm is promising but the approach is not complete. Our experiments only show that multiple solution branches can be explored fast in parallel, but not how to construct a high level algorithm that systematically searches for better schedules within a certain time limit. Further research is needed for that. We also find that multiple threads explore redundant solutions in our approach.

Page generated in 0.0783 seconds