The runtime optimization of federated SPARQL query engines is of central importance to ensure the usability of the Web of Data in real-world applications. The efficient selection of sources (SPARQL endpoints in our case) as well as the generation of optimized query plans belong to the most important optimization steps in this respect. This paper presents CostFed, an index-assisted federation engine for federated SPARQL query processing over multiple SPARQL endpoints. CostFed makes use of statistical information collected from endpoints to perform efficient source selection and cost-based query planning. In contrast to the state of the art, it relies on a non-linear model for the estimation of the selectivity of joins. Therewith, it is able to generate better plans than the state-of-the-art federation engines. In an experimental evaluation based on FedBench benchmark, we show that CostFed is 3 to 121 times faster than the state of the art SPARQL endpoint federation engines.