Comparative analysis of selected containerization tools in terms of MCP

Abstract

Model Context Protocol (MCP) extends AI model capabilities in modern software architectures through standardized tool interfaces. In such systems, MCP servers are often deployed as containerized services, making container runtime behaviour a critical factor for startup responsiveness, execution performance, and infrastructure efficiency. Despite extensive research on container runtime performance and growing interest in MCP-based systems, their interaction has received limited research attention. This paper presents an  empirical evaluation of container runtimes executing an MCP server under controlled conditions. Metrics including Time To First Successful Response (TTFSR), CPU usage, memory footprint, and p95 latency were collected across multiple scenarios. Statistical analysis was conducted using non-parametric methods with multiple-comparison correction and effect size estimation. The results indicate that cold-start latency differences remain inconclusive, while measurable differences are observed in memory usage and CPU consumption. These findings suggest that runtime selection should be guided by workload characteristics rather than architectural assumptions alone.