A Task Assignment Method for Phi Structure

Xeon Phi is a high performance co-processor launched by Intel in 2012. Though Phi is specifically designed for Exascale super computer, the task assignment for Phi is yet to be studied. Based on the special needs of task assignment for Phi, this paper presents an algorithm evolved from graph bisection algorithm: a graph is formed based on the memory dependence of tasks, by traversal the graph with an assuming cut point, iteratively finds out the groups of tasks with least dependence on tasks outside group, this algorithm can provide a task assignment solution between CPU and Phi aiming at memory optimization. Experiment reveals that this algorithm can significantly reduce the total memory usage of the job, also increase the efficiency of the execution. By reducing the memory usage, this algorithm can eliminates the memory bottleneck of Phi and expand the using range of Phi.