Performance Analysis Of SWDM Systems For Datacenter Interconnects

This thesis focuses on the statistical analysis and prediction of fiber performance in data center environments, specifically for short links. The study analyzes the relationships between various bandwidth metrics, such as the -3dB, -5dB, -10dB bandwidths, and the equivalent bandwidth, in predicting the maximum achievable fiber length (Lmax) and Bit Error Rate (BER). The dataset used for the analysis comprised 8 lasers, 4 wavelengths, and 3766 OM4 fibers, with measurements taken across multiple short distances commonly found in data centers. Initially, a detailed correlation analysis was conducted to explore the relationships between these bandwidth metrics, Lmax, and BER. The equivalent bandwidth emerged as the most strongly correlated parameter with Lmax, particularly at shorter distances, such as 30m and 50m, where correlations reached as high as 96%. Additionally, the analysis was extended across a frequency range of 50 GHz to 300 GHz to better understand how frequency impacts this correlation. Despite these efforts, the correlation alone was insufficient to provide highly accurate predictions of Lmax. To address this challenge, a Machine Learning (ML) model was developed to predict Lmax using the four key bandwidth parameters. The model successfully captured the complex relationships that could not be fully explained through statistical correlation alone. With an accuracy of 99.98%, the ML model significantly outperformed traditional predictive methods based solely on correlation analysis. This work highlights the efficiency and practicality of using machine learning techniques in data center environments, where accurate and rapid predictions are essential for network performance evaluation and planning. The model developed in this thesis offers a highly effective alternative to time-consuming data extraction processes, enabling near-instant predictions of Lmax with high accuracy, thereby optimizing operational efficiency in data centers