| Since the beginning of the 21 st century,multi-core processors have been developing rapidly,the GPU core has a natural computing advantage compare to the CPU core.To improve the computing performance,heterogeneous multi-core processors which integrated CPU cores and GPU cores have emerged.Most of the current processor chips access shared resources through the Network on Chip(NoC).With the increasing number of integrated components,the probability of failure increases.The configuration of the NoC has an important impact on its performance.Heterogeneous Network on Chip(HNoC)integrates other components,such as Last Level Cache(LLC)and Memory Controller(MC),in addition to the integrated CPU core and GPU core.These various components work together on one chip.The quality of communication between the components becomes critical.The emergence of heterogeneity has brought new challenges to the design of NoC.First,GPU has a large number of threads that can process a large amount of computing resources in parallel.This feature makes the communication between the GPU and the LLC and the memory controller more frequent.The communication of each link in the traditional NoC is relatively average.But congestion in the vicinity of the GPU node in the HNoC is serious,which decreases system's performance.Second,when a faulty node or a faulty link occurs on the NoC,the traditional fault-tolerant routing algorithm bypasses the faulty node and transmits the data packet to the downstream router to complete the communication.However,when the traffic of the downstream node is already dense,the local traffic of the node is too large.If the dense traffic cannot be effectively balanced,network congestion will be caused and system performance be degraded.In order to solve the two mentioned problems above,this paper first analyzes the traffic between CPU,GPU and memory controller in heterogeneous systems,explores the characteristics of communication between different nodes,obtains the heterogeneous system traffic model,and proposes a suitable network topology for heterogeneous systems.The topology is designed to reduce the distance of communication-intensive nodes,thereby reducing access latency.Secondly,a congestion awareness fault-tolerant routing algorithm based on HNoC is proposed.Our experience partitioned the whole HNoC into four area,then added a local congestion monitor for each area.The algorithm runs based on the congestion current network traffic provided by the monitor,heuristically.The turn model is applied one by one to the all nodes on the HNoC.The algorithm uses turn models in order to avoid deadlocks in wormhole-based networks.In order to evaluate the performance of the proposed method,this paper selects the Noxim a NoC simulator as the experimental platform.Our experiment compares the proposed algorithm with the traditional fault-tolerant algorithm.In our purposed HNoC topology when the failure rates are 5%,10%,15% and 20%,the average latency reduce 35%,22%,25% and 24 respectively,throughput attenuation decrease 12.5%,41.7%,23.7% and 26.3%,respectively.In conclusion,the proposed method has a significant improvement of performance. |
