Microservice decomposition typically emphasizes logical or domain-driven boundaries, often overlooking performance bottlenecks from low-level system interactions. We present a system call-aware decomposition method that identifies and separates functions likely to interfere at the kernel level. By defining a collision score based on system call frequency and type, and using a fine-tuned Large Language Model to statically predict syscall behavior, we construct a function interaction graph for clustering. Evaluation on Python-based monoliths shows up to 30% latency reduction and improved scalability compared to traditional approaches, demonstrating the value of kernel-informed microservice design.