Based on software defined optical network and machine learning, the flexible resource management mechanism (ML-FRM) is proposed, which meets the resource requirements of different services in the optical data center networks. The machine learning is integrated to the SDON controller, which accomplishes the resource allocation algorithms according to the classification and clustering results. ML-FRM firstly utilizes unsupervised learning K-means algorithm to cluster traffic flows, and uses supervised learning support vector machine (SVM) algorithm to realize hierarchical classification of channel qualities. Fragmentation-Function-Fit algorithm is proposed to reduce the blocking probability, the results show that it has the lower blocking probability than First-Fit and Exact-First-Fit algorithms. ML-FRM allocates the required resources through different algorithms based on different traffic flow clustering results, and uses different modulation methods for different channel qualities. The analysis results show that ML-FRM has lower blocking probability, acceptable complexity level, and higher spectrum resource utilization efficiency than other algorithms under different offered load level.

基于软件定义的光网络和机器学习，提出了一种灵活的资源管理机制（ML-FRM），可以满足光数据中心网络中不同服务的资源需求。机器学习被集成到SDON控制器中，SDON控制器根据分类和聚类结果完成资源分配算法。ML-FRM首先利用无监督学习K-means算法对交通流进行聚类，然后利用有监督学习支持向量机（SVM）算法实现信道质量的分级分类。提出了使用分片函数拟合算法来降低阻塞概率，结果表明该算法具有比First-Fit和Exact-First-Fit算法更低的阻塞概率。ML-FRM根据不同的业务流聚类结果，通过不同的算法分配所需的资源，并对不同的信道质量使用不同的调制方法。分析结果表明，在不同负载水平下，ML-FRM具有比其他算法更低的阻塞概率，可接受的复杂度和较高的频谱资源利用率。