Efficient Parallel Linked List Processing

OpenMP is a very popular and successful parallel programming API, but efficient parallel traversal of a list (of possibly unknown size) of items linked by pointers is a challenging task: solving the problem with OpenMP worksharing constructs requires either transforming the list into an array for the traversal or for all threads to traverse each of the elements and compete to execute them. Both techniques are inefficient. OpenMP 3.0 allows to addresses the problem using pointer chasing by a master thread and creating a task for each element of the list. These tasks can be processed by any thread in the team. In this study, we propose a more efficient cutoff-based linked list traversal using our task-based parallel programming model, GPRM. We compare the performance of this technique in both GPRM and OpenMP implementations with the conventional OpenMP implementation, which we call Task-Per-Element (TPE).