With the development and popularization of virtualization technology, more and more enterprises will deploy their business-critical systems on virtualization platform. While reducing the company's hardware and management costs, virtualization also brings severe challenges for system reliability. While the runtime system state replication backup method can improve the failure recovery capabilities of system, it also introduces huge overhead. This paper presents a performance optimization method based on hidden Markov model for system failure recovery. It analyzes runtime states of the system, and calculates the probability of system running tendency. Business system optimization is achieved by dynamically adjusting resources allocation between the failure recovery function and normal business function to reduce the runtime overhead. Experimental results show that the presented approach can guarantee reliability of the system while effectively reducing performance overhead by up to 2/3.