TT-PECN: Temporal Transformer Based Predictive Explicit Congestion Notification For Data Center Networks

Abstract

Data Center Networks (DCNs) handle a large number of data flows, which can cause congestion, packet loss, and delay. Existing Explicit Congestion Notification (ECN) methods mainly react to the current queue condition and may not respond early enough to upcoming congestion. This research paper proposes TT-PECN (Temporal Transformer based Predictive Explicit Congestion Notification) to provide proactive congestion notification in DCNs. The proposed method uses a Temporal Transformer to learn network traffic patterns from time-series data and predict future queue conditions in terms of target queue length. The prediction is used to make ECN marking decision. The model is trained using data generated in NS-3.34 with DCTCP (Data Center TCP) and RED with ECN support. Different link bandwidths and TCP flow levels are used for evaluation. The performance of TT-PECN with DCTCP is evaluated with existing DCTCP and TCP NewReno. Average per-flow throughput and average per flow packet delivery ratio are used as performance metrics. It has been observed that under various distinct scenarios, TT-PECN with DCTCP performs better as compared to existing DCTCP and TCP NewReno.