Centralized traffic monitoring for online-resizable clusters in Networks-on-Chip

Abstract

Runtime-based traffic monitoring is one of the most essential system services for modern many-core platforms. It ensures self-awareness of the current system load and enables other runtime mechanisms, like application mapping and adaptive routing, to optimize workload operations. While Networks-on-Chip introduced a scalable and massively parallel communication infrastructure for the growing number of on-chip resources, scalable traffic monitoring becomes a key concern. A high degree of runtime adaptivity at different system layers comes with the costs that monitored system states need to be available at different locations, scopes and resolutions. Furthermore, many-core systems with hundreds of resources will be less single purpose and run workloads composed of various application domains, traffic and timing characteristics. Thus, the monitoring operations should be adaptive to consider this variability and provide customizable behavior. In the work at hand, a runtime configurable and cluster-based traffic monitoring is proposed. Our experiments show that in maximum 2% hardware overhead per tile in an 8×8 NoC is needed to enable online resizable clustering of up to 64 IP cores, centralized traffic monitoring of all path and link loads inside these clusters at resolutions of 1%, and configurable monitoring timing adjustment in a range of 103 to 105 clock cycles.