摘要： 为满足大规模脉冲神经网络（SNN）的计算需求，类脑计算系统通常需要采用大规模并行计算平台。然而随着节点数量的增多，通信在仿真中所占比例大幅增加，导致计算效率下降。类脑模拟器开源软件NEST采用缓冲区大小相等的策略，有效缩短了通信时间，但是由于缓冲区互相无交流，使得通信数据量持续增加，因此其在能耗方面表现较差。分析NEST集群的负载特性，针对其中的通信问题进行稀疏性优化，提出基于SNN子图跨节点优化的神经元重分布算法ReLOC。通过优化SNN子图的跨节点分布减少每一轮神经元到进程的数量，从而减少跨节点脉冲，使进程间通信更加稀疏，达到缩减每一轮通信进程的目的。在此基础上，以稀疏交换的思想对NEST本身的通信机制进行改进，使有脉冲交换的进程进行数据交换，从而在连接稀疏的情况下提升通信效率。以包含28个Xilinx PYNQ节点的计算集群作为实验平台，运行皮质微电路SNN模型和平衡随机网络模型，验证ReLOC算法的有效性。实验结果表明，相比循环分布算法，重分布算法能够使通信的平均稀疏性提高20%，同时配合稀疏交换最多可使通信能耗减少98.63%。

关键词: 脉冲神经网络, 神经元重分布, PYNQ集群, NEST仿真器, 稀疏交换

Abstract: To meet the computational requirements of large-scale Spiking Neural Network(SNN), brain like computing systems generally require large-scale parallel computing platforms.However, the proportion of communication steps required to conduct a simulation increases greatly with increasing numbers of nodes, resulting in reduced computational efficiency.Open-source brain like simulation software NEST uses the strategy of equal buffer size, which considerably reduces the required communication time.However, due to the lack of communication between buffers, the amount of communication data continues to increase.Therefore, the performance of such methods is poor in terms of energy consumption.To solve this problem, this paper analyzes the load characteristics of a NEST cluster and carries out a sparsity optimization of the communication system.Furthermore, it proposes a neuron relocation algorithm called ReLOC based on SNN subgraph cross-node optimization.Optimizing the cross-node distribution of the SNN subgraph reduces the number of neurons to be processed in each round.This reduces the cross node pulse, increases the sparsity of the inter-process communication, and reduces the length of each round of the communication process.On this basis, the communication mechanism of NEST is improved based on the idea of sparse exchange, enabling the process to exchange data via a pulse exchange mechanism and improving communication efficiency under conditions of sparse connection.Using a computing cluster containing 28 Xilinx PYNQ nodes as an experimental platform, this paper implements a cortical microcircuit SNN model and a balanced random network model to verify the effectiveness of the proposed ReLOC algorithm.Experimental results show that ReLOC algorithm can improve the average communication sparsity by 20%, and reduce communication energy consumption by up to 98.63%.

Key words: Spiking Neural Network(SNN), neuron relocation, PYNQ cluster, NEST simulator, sparse exchange

中图分类号: 

• TP391