چكيده به لاتين
Nowadays Network-on-Chip (NoC) provides remarkable facilities in transmitting data between Chip Multiprocessors (CMPs). Routing algorithms have a key role in Network-on-Chip (NoC) to increase the received packets at the destination routers. Congestion is a known problem that affects performance in Network-on-Chip communication. As deterministic routings do not consider the large-scale network’s status, Congestion-aware routing algorithms have been used for packet transmission but they are faced with hardware overhead for transmitting congestion information. In this thesis, we present and evaluate a novel routing scheme which combines the advantages of both deterministic and adaptive routing for traffic management in a 2D mesh network. For this purpose, we proposed a partitioning model that divides the network into two inter and intra-partition regions. Furthermore, in order to limit hardware overhead, congestion information is propagated in the header flit of the packet which is divided into two separate sections for storing inter and intra values. The purposed method utilizes this congestion information for routing decisions and forward packet toward partitions with lower traffic. Congestion information will be transferred between inter-partition regions to locally evaluate the traffic conditions. Using this evaluation, it is possible to use a deterministic algorithm such as dimension-order routing, which is known as XY to decrease the number of virtual channels. In the proposed algorithm, the congestion deflected from the center of the network and distributed to the entire network and tries to improve the overall system performance. The simulation of the purposed algorithm has been carried out the Noxim simulator. Under different traffic patterns, simulation results show that our routing algorithm improves the saturation throughput and average packet latency in comparison with other algorithms. The results indicate that our algorithm has at least 15% and at most 64% improvement in latency and between 5% and 16% improvement in throughput. In addition, the method that we proposed for having an average in congestion information shows 80% overhead improvement for a network with 64 routers.
