A Trace-Based Structural Observability Framework for Network-on-Chip Routing Evaluation

Abstract

Traditional evaluation of Networks-on-Chip is based on aggregate metrics. Most of these metrics (latency, throughput, hop count, packet loss, energy consumption, buffer occupancy, thermal or reliability summaries) compress detailed execution traces into a few scalar values. This dimensionality reduction hides spatial, temporal, and resource-level differences between simulations. This paper introduces an Entropic Structural Observability framework for routing-independent post-simulation analysis of NoC traces. The main idea of the framework is to convert typed packet-level events into probability distributions over used resources, including routers, directed links, time windows, and critical resources. As a final report, the method builds a structural signature. This signature includes normalized entropy, effective support, concentration, spatiotemporal mutual information, distribution drift, topology-aware spatial statistics, buffer-pressure measures, and classical imbalance indicators. The analysis does not modify simulation traces; it operates as a diagnostic layer. It reveals activity distribution, temporal dependence, spatial evolution, resource pressure, and structural imbalance. The framework is evaluated on 256-router 2D and 3D mesh topologies using deterministic and adaptive routing policies under diverse traffic patterns and injection rates, enabling its structural signatures to be examined across different network dimensionalities. The results reveal distinct structural regimes: deterministic dimension-order routing shows broad spatial and temporal dispersion, DyAD exhibits stronger time-space coupling and drift, and Fully-Adaptive presents an intermediate profile with high dispersion but moderate temporal variation.