Evaluating Service Curves via Measurement

In recent decades, on-chip networks have emerged as crucial platforms for System-on-Chip (SoC) communications, owing to their versatility and superior performance over traditional bus-based models. As the demand for high computational performance with real-time guarantees continues to grow, especially in industrial contexts, it has become essential to analyze and optimize NoC architectures to improve latency or delay, ensuring Quality of Service (QoS), and overall architectural design. This thesis explores the application of Network Calculus (NC), an algebraic theory utilized to determine worst-case delay bounds in networks, where traffic with service guarantees is depicted as curves on a Cartesian plane, as a tool for performance analysis and optimization of NoCs. NC leverages transformation operations, such as min-plus convolution and deconvolution, to model traffic profiles and estimate the minimum service guarantees for flows traversing network nodes. To support the complex task of designing such systems with both performance and QoS requirements, frameworks such as ARM MPAM envisage systems with in-hardware support to resource partitioning and the observation of its effects. This opens the path to new venues of application for traditional QoS techniques. By examining traces of sampled traffic, this work estimates service curves, validates QoS requirements, worst-case delay and backlog bound analysis, bandwidth estimation, and detects congestion and contention within SoC architectures. The research specifically focuses on wormhole NoCs, investigating architectural development and utilizing NC to estimate the service curves and provide guarantees of QoS. The proposed approach offers unique features for simulating heterogeneous NoCs with monitoring facilities, varying numbers of virtual channels for each unidirectional port, and flexible link capacities. Also, it supports realistic (internal delay is not zero and data-rate is finite) routers, and ideal (internal delay is zero and data-rate is infinite) routers. Moreover, supports a comprehensive set of analyses at both packet and flit levels, including throughput, e2e delay, transfer delay, link utilization, traffic packet success and loss rate, etc. Furthermore, a priority-driven scheduling mechanism is proposed for Virtual Channels (VCs) in wormhole NoCs. This mechanism supports the allocation of VCs to different traffic with varying priorities, facilitating VC sharing and enhancing traffic management. The findings underscore the potential of Network Calculus not only in theoretical design scenarios but also in practical, real-time applications where systems operate under unpredictable conditions. By integrating NC with reactive reconfiguration schemes, this research offers a robust framework for dynamic NoC management, ensuring both high performance and reliable real-time operation.