MIOS is a scalable operating system designed for large scale CCNUMA system. It introduces multi-instance kernel structure. In MIOS, each instance of OS kernel executes the same code, but runs on a node of the CCNUMA machine and manages its resources respectively. The MIOS provides a single system image running environment for all nodes of CCNUMA system, supporting process and thread task model. Aiming at the features of CCNUMA system and the requirements of scientific computing applications, the MIOS provides several optimizations, including weak shared thread model, cascaded task scheduling, adaptive communication between tasks and register-based lock. We have implemented MIOS on our Galaxy parallel computer system, a large scale CCNUMA system including 2048 processors. The evaluations on Galaxy system, including micro-benchmarks and real parallel applications, show that MIOS can provide comparable performance with a conventional OS for MPI applications. For OMP applications, the MIOS also can provide a good performance speedup on the large scale CCNUMA system with 2048 processors. The structure of MIOS can also provide experiences for designing operating system on many-core processor.