Co-optimizing application partitioning and network topology for a reconfigurable interconnect

To realize the full potential of a high-performance computing system with a reconfigurable interconnect, there is a need to design algorithms for computing a topology that will allow for a high-throughput load distribution, while simultaneously partitioning the computational task graph of the application for the computed topology. In this paper, we propose a new framework that exploits such reconfigurable interconnects to achieve these interdependent goals, i.e., to iteratively co-optimize the network topology configuration, application partitioning and network flow routing to maximize throughput for a given application. We also present a novel way of computing a high-throughput initial topology based on the structural properties of the application to seed our cooptimizing framework. We show the value of our approach on synthetic graphs that emulate the key characteristics of a class of stream computing applications that require high throughput. Our experiments show that the proposed technique is fast and computes high-quality partitions of such graphs for a broad range of hardware parameters that varies the bottleneck from computation to communication. Finally, we show how using a particular topology as a seed to our framework significantly reduces the time to compute the final topology.