Document Type

Article

Publication Date

8-2019

Publication Title

Journal of Supercomputing

Abstract

Modern graphics processing units (GPUs) offer much more computational power than recent CPUs by providing a vast number of simple, data-parallel, multithreaded cores. In this study, we focus on the use of a GPU to perform parallel discrete-event simulation. Our approach is to use a modified service time distribution function to allow more independent events to be processed in parallel. The implementation issues and alternative strategies will be discussed in detail. We describe and compare our experience and results in using Thrust and CUB, two open-source parallel algorithms libraries which resemble the C++ Standard Template Library, to build our tool. The experimental results show that our implementation can be two orders of magnitude faster than the sequential simulation for large-scale simulation models.

Comments

This research was supported in part by a Fulbright Senior Research Scholarship and by an allocation of computing time from the Ohio Supercomputer Center.

DOI

10.1007/s11227-018-2254-4

Version

Postprint

Volume

75

Issue

8

Share

COinS