Computation hierarchy for in-network processing

In this paper we explore the network level architecture of distributed sensor systems that perform in-network processing. We propose a system with heterogeneous nodes that organizes into a hierarchical structure dictated by the computational capabilities. The presence of high-performance nodes amongst a sea of resource-constrained nodes exposes new tradeoffs for the efficient implementation of network-wide applications. Our experiments show that even for a low relative density of resource-constrained nodes to high-performance nodes there are certain gains in performance for a heterogeneous and hierarchical network over a homogeneous one. The introduction of hierarchy enables partitioning of the application into sub-tasks that can be mapped onto the heterogeneous nodes in the network in multiple ways. We analyze the tradeoffs between the execution time of the application, accuracy of the output produced and the overall energy consumption of the network for the different mapping of the sub-tasks onto the heterogeneous nodes. We evaluate the performance and energy consumption of a typical sensor network application of target tracking via beamforming and line of bearing (LOB) calculations on the different nodes. Our experiments also include the study of the overall performance and energy consumption of the LOB calculation using two different types of resource constrained sensor nodes (MICA and MICA2 nodes) and show how these metrics are affected by changes in the node architecture and operation. Our results indicate that when using MICA motes as resource-constrained nodes, 85% of the time on average the hierarchical network outperforms a homogeneous network for approximately the same energy budget. When using MICA2 motes as resource-constrained nodes, 54% of the time the hierarchical network performs better than a homogeneous network with approximately the same energy budget.