In dynamic and autonomic grid environment, monitoring of heterogonous resources is critical for performance analysis & tuning, job scheduling, fault detection & diagnosis and so on. Existing researches are always limited at either inter-operation or introducing too much intrusive overhead. GridEye is a novel service-oriented monitoring system with a flexible, scalable architecture. An open, extendable information schema is adopted to standardize data contents and Web Service interfaces are employed to provide inter-operative ability for different grid system. Furthermore, an improved forecasting algorithm is provided for precise performance prediction on the basic of traditional MA (k) and ExS(?) models. Experiments proved that GridEye introduces very little overhead while providing good capabilities. It is also indicated that the algorithm can promote the precision of prediction.