Towards a Novel Congestion Notification Algorithm for a Software-Defined Data Center Networks

Nowadays, besides the evident benefits of Data Centers (DC) in terms of information storage and processing, that posed also a few problems needed to be addressed. The computer network research community has been dedicated to resolve one of the severe problems of Data Center Networks (DCN): the congestion notification problem. The quality of connection links is the decisive factor for the success of a DC in providing higher bandwidth and low interconnection latency for high performance cluster computing. Therefore, the congestion detection and reaction mechanisms are vital to the performance of a DC. There are divers congestion notification algorithms, such as Backward Congestion Notification (BCN), Quantized Congestion Notification (QCN), Forward Explicit Congestion Notification (FECN,) and E-FECN. However, the common disadvantages of those algorithms is the lack of ability to predict congestion. In this paper, we propose an AI-based congestion notification algorithm in applying the Long short-term memory (LSTM) algorithm and its variants, such as GRU, Skip-LSTM, and Bi-Directional LSTM, into the original BCN algorithm, called IBCN, in order to increase the capacity of predicting link congestion of the original one. We also implement IBCN in a Software-defined DCN that is a combination of the DCN and the Software-defined Network (SDN) architecture. The obtained experimental results showed that the IBCN outperforms the original BCN at any network load levels (from 10% to 90%) with the average improvement of 50% of throughput. In addition, the proposed solution gives also better results than other traditional linear prediction algorithms.