Knowledge graphs are the main representation form of intelligent data. With the development of knowledge graphs, more and more intelligent data has been released in the form of the resource description framework (RDF). It is known that the semantics of SPARQL correspond to graph homomorphism which is an NP-complete problem. Therefore, how to efficiently answer SPARQL queries in parallel over big RDF graphs has been widely recognized as a challenging problem. There are some research works using the MapReduce computational model to process big RDF graph. However, SPARQL queries in these works are decomposed into the set of query clauses without considering any semantics and graph structure embedded in RDF graph, which leads to overmuch MapReduce iterations. This study first decomposes the SPARQL query graph into a set of stars by utilizing the semantic and structural information embedded RDF graphs as heuristics, which can be matched in fewer MapReduce iterations. Meanwhile, a matching order of these stars is given to reduce intermediate results in MapReduce iterations. During the matching phase, each round of MapReduce adds one star with the join operation. The extensive experiments on both synthetic dataset WatDiv, and real-world dataset DBpedia are carried out. The experiments results demonstrate that the proposed star decomposition-based method can answer SPARQL BGP queries efficiently, which outperforms SHARD and S2X by one order of magnitude. Finally, extensive experiments show that the performance of the optimization algorithms is improved by 49.63% and 78.71% than the basic algorithm over both synthetic and real datasets.