Dynamic Load Adaptation in Network on Chip for Reconfigurable Systems
Anders Rångevall (D98)
Much research point in the direction that future integrated circuits will have a built-in network to connect the different regions of the chip.
With ever increasing clock speeds the effort needed to distribute the clock and the problem of keeping everything synchronized becomes much harder. One suggestion to alleviate this is the so-called GALS technique, Globally Asynchronous Locally Synchronous. Here the chip is divided into separate regions each with its own clock. The NoC enables parts from different clock domains to communicate. Another issue with increasing clock speeds is power efficiency. NoC offers the possibility of reduced power consumption by shortening interconnection wire length. It could also have the ability to power down unused parts of the network. NoC are also attractive because they have the ability to offer higher data throughput compared to many other interconnection techniques.
One issue with packet switched NoC that need to be addressed is the possibility of the network being congested if too many packets are injected into it at the same time. There has been previous work done on techniques to avoid network congestion but that has mostly been targeted towards networks for parallel computers. Much can be adopted from that work but NoC will most likely need novel techniques. In this particular case one reason is the high level of dynamism that comes with a reconfigurable platform, the units connected to the network, the network load and traffic patterns will all change during run-time.
The objective is to find a way to avoid network congestion by limiting packet injection. The Injection Rate Control mechanism should be able to automatically adapt to different traffic patterns and perform well under both low and high loads. The mechanism should be fairly scalable to fit both small and large networks but it should primarily be suited to fit a specific network that is currently available. For scalability reasons it would be good if the Injection Rate Control used only local knowledge of network traffic. This autonomous Injection Rate Control mechanism can be seen as forming one part of a distributed operating system.
The available Network on Chip is part of the Gecko system. This is a demo platform to explore different types of reconfigurable computing and heterogeneous platform run-time management. The Gecko-system presently has a global mechanism to regulate packet injection in a fairly crude and fashion.
The work will include the following steps:
1. Find a suitable Injection Rate Control algorithm.
2. Modify a network simulator to test the algorithm.
3. Implement the automatic Injection Rate Control mechanism on the Gecko system.
Advisor: Lambert Spaanenburg (EIT)