Abstract:
Task mapping is the assignment of job tasks to nodes. Traditionally, the goal of task mapping is to maximize locality, reducing the number of network hops needed to deliver messages as a way of reducing bandwidth consumption. We show that on a Dragonfly topology, such a strategy can be counterproductive because, while traffic is reduced, it is also concentrated on only a few global links, creating hot spots. We formulate the balanced adjacency coloring problem to design mappings that evenly spread network traffic, give optimal algorithms to solve it for a number of cases, and use simulations to show that mappings based on these algorithms can reduce the communication time of a stencil job by up to 20%.