The packet sampling algorithm based on mask matching is a practical distributed traffic sampling algorithm. But the algorithm is ineffective in measuring an important network traffic characteristic, which is the distribution of packet arrival time interval. First of all, the reason leading to the problem is analyzed according to the error theory. Then, an improved program of double sampling is introduced in the initial algorithm in order to reduce the system error of measurement. As the improved algorithm will bring extra burden to the measurement system, a solution which is to increase the sampling mask bits is put forward and its feasibility is demonstrated. Finally, the improved algorithm is tested with the actual network traffic data. The experiment results show that the measurement of the distribution of packet arrival time interval is in line with the real situation. And it has little effect on the measurement accuracy of other network performances.