Improving Packet Predictability of Scalable Network-on-Chip Designs without Priority Pre-emptive Arbitration

The quest for improving processing power and efficiency is spawning research into many-core systems with hundreds or thousands of cores. With communication being forecast as the foremost performance bottleneck, Network-on-Chips are the favoured communication infrastructure in the context mainly due to reasons like scalability and power efficiency. However, contention between non-preemptive NoC packets can result in variation in packet latencies thus potentially limiting the overall utilisation of the many-core system. Typical latency predictability enhancement techniques like Virtual Channels or Time Division Multiplexing are usually hardware expensive or non-scalable or both. This research explores the use of dynamic and scalable techniques in Network-on-Chip routers to improve packet predictability by countering Head-of-line blocking (blocked low priority packet blocking a high priority packet) and tailbacking (low priority packet utilising the link that is required by a high priority packet) of non-preemptive packets.

The Priority forwarding and tunnelling technique introduced is designed to detect Head-of-line blocking situations so that its internal arbitration parameters can be altered (by forwarding packet parameters down the line) to resolve such issues. The Selective packet splitting technique presented allows resolution of tailbacking by emulating the effect of preemption of packets (by splitting packets) by using a low overhead alternative that manipulates packets. Finally, the thesis presents an architecture that allows the routers to have a notion of timeliness in data packets thus enabling packet arbitration based on application-supplied priority and timeliness thus improving the quality of service given to lower priority packets. Furthermore, the techniques presented in the thesis do not require additional hardware with the increase in size of the NoC. This enables the techniques to be scalable, as the size of the NoC or the number of packet priorities the NoC has to handle does not affect the functionality and operation of the techniques.