Modern high-end embedded systems require both predictable real-time scheduling and high-level abstraction interface to their OS kernels. Since these features are difficult to be balanced by a single OS, some methods that accommodate multiple different versions of OS kernels, typically real-time OS and general purpose OS, into a single device have been proposed. The hybrid kernel, one of those methods, executes a general purpose OS kernel as a task of real-time OS which can support those features with reasonable engineering effort. However when adapting the approach to various combinations of OS kernels, which is required in the real-world embedded system design, the engineering effort of modifying the kernel becomes not negligible. This article introduce a method called a composition kernel which uses a thin abstraction layer for accommodating kernels without making direct dependencies between them. The authors developed the abstraction layer on an SH-4A processor and executed kernels on top of it. The amount of modifications to the kernels was significantly smaller than that in related work, while introducing only negligible verhead to the performance of the kernels.