Hardware Supported Adaptive Data Collection for Networks on Chip

Abstract

Managing future many-core architectures with hundreds of cores, running multiple applications in parallel, is very challenging. One of the major reasons is the communication overhead required to handle such a large system. Distributed management is proposed to reduce this overhead. The architecture is divided into regions which are managed separately. The instance managing the region and the applications running within the regions need to collect data for various reasons from time to time, e.g., to collect data for proper mapping decision, to synchronize tasks or to aggregate computation results. In this work, we propose and investigate different strategies for adaptive data collection in meshed Networks on Chip. The mechanisms can be used to collect data within regions, which are defined during run-time in respect of size and position. The mechanisms are investigated while considering delay, NoC utilization and implementation costs. The results show that the selection of the used mechanism depends on the requirements. Synthesis results compare area overhead, timing impact and energy consumption.