Existing algorithms of distribution reduct, maximum distribution reduct and assignment reduct for inconsistent decision tables are inefficient, which are not suitable for large data sets. A measurement of attribute importance based on the relative discernibility degree was presented firstly, which overcomes the shortcoming of positive domain in measuring the importance of attributes. Then, in order to simplify the decision table, some kinds of simplified consistent decision tables were defined. In the end, an efficient attribute reduction algorithm was designed based on the relative discernibility degree. Theoretical analysis and experimental results show the effectiveness and practicalbility of this algorithm on the large inconsistent data sets.