非均匀业务流下的片上网络缓冲区分配算法

Buffer allocation algorithm under non-uniform traffic in NoC

针对片上网络的非均匀业务流,提出一种基于模拟退火遗传算法的缓冲区资源分配算法,对系统的有限缓冲区资源的分配问题进行了研究。该算法建立在二维Mesh结构的片上网络通信模型基础上,根据各节点间的业务流特征,估计出节点中各输入通道的负载大小,再根据其负载情况采用模拟退火遗传算法进行缓冲区资源的分配,从而使整个网络的平均延时性能最优。实验中设置了不同的热点位置和热度,结果表明,该算法可以更合理地分配缓冲区资源,有效降低数据包的传输延时。在单热点通信流量下,热度为100%和300%时,可分别降低传输延时32.58%~65.29%和35.54%~70.38%;在双热点通信流量下,可降低传输延时52.02%~70.43%。同时,该算法具有良好的收敛性。

For non-uniform traffic in network-on-chip, this paper proposed a buffer allocation algorithm based on simulated annealing genetic algorithm and researched the allocating problem of limited buffer resources in a system. Based on the model of communication performance for 2D Mesh network-on-chip, the proposed algorithm firstly estimated the load on each input channel in different routers according to the traffic characteristics. Then,it used simulated annealing genetic algorithm to allocate the buffer resources according to the distribution of the load on all channels, minimizing the average delay of the system. In the sim- ulation, the algorithm set different hotspots and hot degrees. Simulation results show that the buffer allocation algorithm is more reasonable to allocate buffer resources and more effective to decrease average packet latency. For single hotspot traffic, the algo- rithm can decrease 32.58% to 65.29% and 35.54% to 70.38% packet latency when hot degree is 100% and 300% respectively. For dual hotspot traffic ,the algorithm can decrease packet latency 52.02% to 70.43%, is less than those by uniform algorithm. Meanwhile, the algorithm has good convergence.