Graph-Driven Insights for Heterogeneous Data Center Networks: Leveraging Converters With a Novel Graph Transformer Approach

Data centers extensively employ topologies such as CLOS, BCube, and Xpander to manage their large networks of devices. Data center networks (DCN), which rely on traditional topologies that are inherently static and inflexible to traffic pattern variations, face increasingly the adoption of dynamic topologies facilitated by multi-mode converters, providing a more adaptable solution without the need to expand the existing communication resources. However, fully utilizing the heterogeneous information of DCNs and achieving a network-wide joint optimization of converters remain significant challenges in algorithm design. This paper marks the first foray into using graph representation learning tools for DCN topology optimization and introduces MorpheusClos as an innovative solution. We present the novel DyTa block, which integrates graph transformers with causal convolution for extracting heterogeneous DCN graph representations, enabling joint optimization of network-wide converters from a graphical perspective, thereby improving network adaptability and performance under diverse traffic conditions. Extensive experiments and simulations reveal that MorpheusClos significantly outperforms existing topologies, reducing average path lengths by 29% and congestion probabilities by 57%.