Large-scale RDF graph databases stored in shared-nothing clusters require query processing engine that can effectively exploit highly parallel computation environment. We propose algebra of RDF graphs and its physical counterpart, physical algebra of RDF graphs, designed to implement queries as distributed dataflow programs that run on cluster of servers. Operations of algebra reflect the characteristic features of RDF graph data model while they are tied to the technology provided by relational query execution systems. Algebra of RDF graphs allows for the expression of pipelined and partitioned parallelism. Preliminary experimental results show that proposed algebra and architecture of query execution system scale well with large clusters of data servers.