At present, virtualization technology has been widely applied in data centers. However, VCPU (virtual CPU) scheduling strategy still faces intolerable I/O delay, especially for I/O-latency sensitive VMs which suffer from significant performance degradation when competing with CPU-intensive VMs. This paper presents a flexible and efficient VCPU scheduling algorithm FLMS (flexible I/O latency and multi-processor sensitive scheduler) which utilizes VM classification, VCPU binding and flexible slicer to reduce VM response delay. The work also redesigns the load balancing strategy to ensure optimal VCPU migration. FLMS is suitable for the current mainstream virtualization solutions. It has a 30% improvement comparing with the latest software virtualization. With hardware-assisted virtualization, FLMS makes it possible for VMs to achieve near bare-metal performance and ensures the fairness of the whole system