The exponential growth of data has posed serious challenges to the data management and analysis. Join query is a common data analysis operation, and MapReduce is a programming model implemented for parallel processing on large-scale datasets. Therefore the research on MapReduce based join algorithms and its cost model has a certain academic significance and application value. This study believes that the I/O (including the network and the local I/O) cost is the main factor affecting the performance of MapReduce based join algorithm. Furthermore, as the I/O cost is determined by the feature of both datasets and join operation, the executed plan of multi-ways join could be optimized by evaluating the I/O cost of two-ways join. In the study, an I/O cost model of two-ways join is proposed and then formally defined as a simple extension to the existing MapReduce based join algorithms, resulting in six join algorithms and their I/O cost functions through write-box analysis. In addition, an selection algorithm to find the best executed plan of multi-ways join is presented. The correctness and accuracy of the I/O cost model are validated through a series of experiments. The experiment results suggest that the I/O cost can accurately reflect the algorithm performance.