A statistic-based benchmark was proposed for performance assessment and monitoring of model predictive control; the benchmark was straightforward and achievable by recording a set of output data only when the control performance was good according to the user's selection. Principal component model was built and an auto-regressive moving average filter was identified to monitor the performance; an improved T-2 statistic was selected as the performance monitor index. When performance changes were detected, diagnosis was done by model validation using recursive analysis and generalized likelihood ratio (GLR) method. This distinguished the fact that the performance change was due to plant model mismatch or due to disturbance term. Simulation was done about a heavy oil fractionator system and good results were obtained. The diagnosis result was helpful for the operator to improve the system performance.