Gathering information from the distributed Web of Data is commonly carried out by using SPARQL query federation approaches. However, the fitness of current SPARQL query federation approaches for real applications is difficult to evaluate with current benchmarks as they are either synthetic, too small in size and complexity or do not provide means for a fine-grained evaluation. We propose LargeRDFBench, a billion-triple benchmark for SPARQL query federation which encompasses real data as well as real queries pertaining to real bio-medical use cases. We evaluate state-of-the-art SPARQL endpoint federation approaches on this benchmark with respect to their query runtime, triple pattern-wise source selection, number of endpoints requests, and result completeness and correctness. Our evaluation results suggest that the performance of current SPARQL query federation systems on simple queries (in terms of total triple patterns, query result set sizes, execution time, use of SPARQL features etc.) does not reflect the systems' performance on more complex queries. Moreover, current federation systems seem unable to deal with real queries that involve processing large intermediate result sets or lead to large result sets.