Multi-Layer Causal Graphs for Distributed System Performance: Modeling Cross-Service Dependencies in Microarchitectures

Modern distributed systems based on microservice architectures face unprecedented challenges in performance management and fault diagnosis due to their inherent complexity and dynamic nature. This paper presents a comprehensive framework for modeling cross-service dependencies using multi-layer causal graphs, enabling more accurate performance analysis and root cause localization in microservice environments. We propose a hierarchical approach that captures both inter-service and intra-service causal relationships across multiple abstraction layers, including the infrastructure layer, metric layer, and invocation layer. Our methodology integrates causal inference techniques with graph-based representations to construct dynamic dependency models that adapt to the evolving nature of microservice systems. Through systematic analysis of service invocation patterns, performance metrics, and infrastructure telemetry, we demonstrate how multi-layer causal graphs can effectively identify performance bottlenecks and trace anomaly propagation paths. The experimental evaluation on benchmark microservice applications reveals that our approach achieves superior accuracy in root cause localization compared to traditional single-layer methods, with an average precision improvement of 23% and recall enhancement of 18%. Furthermore, the proposed framework exhibits excellent scalability, maintaining consistent performance even as system complexity increases with additional services and dependencies.