Analysis and Management of Communication in On-Chip Networks

• Main Author: Jafari, Fahimeh
• Format: Doctoral Thesis
• Language: English
• Published: KTH, Elektronik och Inbyggda System 2015
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-161178; http://nbn-resolving.de/urn:isbn:978-91-7595-458-5

Regarding the needs of low-power, high-performance embedded systems and the growing computation-intensive applications, the number of computing resources in a single chip has enormously increased. The current VLSI technology is able to support such an integration of transistors and add many computing resources such as CPU, DSP, specific IPs, etc to build a Systemon- Chip (SoC). However, interconnection between resources becomes another challenging issue which can be raised by using an on-chip interconnection network or Network-on-Chip (NoC). NoC-based communication which allows pipelined concurrent transmissions of transactions is becoming a dominate infrastructure for many core computing platforms. This thesis analyzes and manages both Best Effort (BE) and Guaranteed Service (GS) communications using analytical performance approaches. As the first step, the present thesis focuses on the flow control for BE traffic in NoC. It models BE source rates as the solution to a utility-based optimization problem which is constrained with link capacities while preserving GS traffic requirements at the desired level. Towards this, several utility functions including proportionally-fair, rate-sum, and max-min fair scenarios are investigated. Moreover, it is worth looking into a scenario in which BE source rates are determined in favor of minimizing the delay of such traffics. The presented flow control algorithms solve the proposed optimization problems determining injection rate in each BE source node. In the next step, real-time systems with guaranteed service are considered. Real-time applications require performance guarantees even under worst-case conditions, i.e. Quality of Service (QoS). Using network calculus, we present and prove the required theorems for deriving performance metrics and then apply them to propose formal approaches for the worst-case performance analysis. The proposed analytical model is used to minimize total cost in the networks in terms of buffer and delay. To this end, we address several optimization problems and solve them to consider the impact of various objective functions. We also develop a tool which derives performance metrics for a given NoC, formulates and solves the considerable optimization problems to provide an invaluable insight for NoC designers. === <p>QC 20150310</p>