Achieving higher reliability in open server computer systems with low cost has been an increasing interest recently. To satisfy this general demand, we propose a new fault recovery mechanism. We extended the recovery cache scheme to adapt to state-of-the-art multiprocessor server computer systems, and built a system level fault recovery mechanism. It enables the system to recover from most intermittent hardware errors without rebooting the system. Furthermore, faulty processors can be isolated dynamically, and not only hardware errors but also many of operating system panics caused by unanticipated software errors can be recovered. The fault recovery mechanism is implemented with the "add-on" hardware module and controlling software module and fully transparent to application programs. Thus no modification is required to the basic hardware and binary compatibility is maintained which is mandatory for open systems. System performance was evaluated using TPC-C benchmark. We also built an experimental system with prototype hardware.