AI and ML Applications Traffic: Designing Challenges for Performance Optimization of Optical Data Center Networks

Artificial intelligence (AI) and machine learning (ML) are vastly becoming key parts of today’s applications, supporting diverse human activities and decision making. New training models are used requiring an explosive traffic exchange among data centers (DCs) servers, putting tremendous pressure on DC network (DCN). Therefore, new DCN design challenges are required to efficiently handle the emerging AI/ML applications traffic. Contrary to existing works that do not account for these challenges, in this paper we introduce an optical intra-rack DCN architecture, as an optical local area network (LAN) paradigm, and an efficient wavelength division multiplexing access (WDMA) algorithm tailored to serve the AI/ML traffic. For the proposed WDMA scheme, we assume diverse AI/ML traffic classes according to the required service time and quality of service (QoS) delivery and we adopt appropriate access rules for each of them to achieve high DCN performance and access fairness. Simulation results show that our proposal reaches high bandwidth utilization ($90 \%$ under nominal load 400 Gbps) and low end-to-end latency (<300 ns across all loads for high/medium QoS traffic, and <100 $\mu \mathrm{s}$ under nominal load for low QoS traffic), contributing to the agenda of new design challenges for optical DCNs suitable for AI/ML traffic.