Clustered Networks-on-Chip: Simulation and Performance Evaluation

Abstract

In many-core Networks-on-Chip (NoC) systems, two topics have recently been researched; NoC clustering and NoC simulation. NoC clustering investigates grouping processing elements (PEs) based on common characteristics between them, like spatial locality or communication patterns. Research showed that NoC clustering achieved better performance and load balancing. Furthermore, it allows scalability of the NoC grid. The other topic, NoC simulation, provides tools for early evaluation of NoC systems. This allowed easy investigation of NoC systems with large number of PEs. Plenty of research has been done in both topics separately. In this paper, we try to fill the gap. We extend an existing NoC simulation tool by adding support for simulating NoC clusters and inter-NoC traffic. The presented NoC clustering simulator supports the modular design technique, which in turn offers the flexibility in configuring cluster parameters. Examples of configurable parameters that are supported by our modified simulator are: adding interconnection topologies to configure how the NoC entities are connected with each other, adding a cluster manager that maps tasks to NoC entities based on the inter-NoC communication pattern, adding latency factors of cluster links in the case of inter-NoC traffic, and adapting the routing algorithm for both inter-cluster and intra-cluster traffic. A clustered NoC simulation case study shows the effectiveness of the modifications made to the simulator. For example, a certain NoC setup is simulated twice, initially non-clustered then clustered, using the added features. Collected data shows that the average clustered NoC routers' energy consumption is 50% lower than the consumption in the non-clustered case.