The paper details its model for the graph structure and attributes, and it shows results for the different approaches they evaluated -- in the end, they propose

The paper details its model for the graph structure and attributes, and it shows results for the different approaches they evaluated -- in the end, they propose a novel approach of using rows/columns for adjacency lists and JSON blobs for properties/attributes. The paper references 3 relational DBs: DB2, Oracle, and PostgreSQL... Most modern relational databases such as DB2, Oracle or Postgresql have features to support both relational and nonrelational storage within the same database engine, making it possible to perform an empirical comparison of the utility of relational versus non-relational storage structures for property graphs. Presumably they used PostgreSQL since it's the only open-source RDBMS referenced, and it supports CET translations, which they used. The Berkeley DB reference is regarding the Titan backend datastore they used. Presumably they chose BDB over one of Titan's distributed backeds like Cassandra or HBase since BDB is single server so it's a more apt comparison to the other DBs.