Clock skews of devices on the Internet are viewed as one way delay noise, but their distribution is unknown. We explore the distribution of clock skews to see the conflict probability. In this paper, we introduce an accurate clock skew estimation algorithm to filter inaccurate clock skew estimation by comparing the results between linear programming method and least square fitting. Delay jitter and other noises affect the estimation result. When the difference of two methods is large, the estimation result is unstable and inaccurate, so the estimation result should be dropped. Based on this algorithm, we use traces of real Internet measurements to collect 1825 accurate clock skews of different devices to establish a fingerprint database. Furthermore, we show the distribution of clock skews and comparing conflict probability with different number of devices. The distribution shows that clock skews are diverse, and most of clock skews are in the region of [-100, 100] PPM. The results indicate that when the number of devices is small (<5), clock skews won’t be conflict with each other, so clock skews are good tools to detect faked devices or NAT; When the number of devices increases, the conflict probability increases linearly, so clock skews of different devices can not distinguish each devices effectively.