Tournament based task allocation in a parallel MIS algorithm

Abstract

This paper explores the use of a tournament data structure for task allocation and dependency tracking in a parallel Maximal Independent Set (MIS) algorithm as a way to reduce contention in counter updates and improve runtime. The tournament data structure adds a noticeable overhead to the algorithm that causes T time of the algorithm to increase but then there is a steady improvement in performance as we increase the number of worker threads. Running the tournament algorithm with 12 threads in our experimental setup, we are able to get an average speedup of 1.13 for our test suite of 7 real-world graphs.