We have implemented Pangaea as a user-space NFS (version 3) server using the SFS toolkit [19]. Our prototype implements all the features described in the paper, except that support for recovery from permanent failures (Section 6) is still fragmentary. Pangaea currently consists of 30,000 lines of C++ code.
A Pangaea server maintains three types of files on the local file system: data files, the metadata file, and the intention-log file. A data file is created for each replica of a file or directory. The node-wide metadata file keeps the extended attributes of all replicas stored on the server, including graph edges and version vectors. Data files for directories and the metadata file are both implemented using the Berkeley DB library [30] that maintains a hash table in a file. The intention-log file is also implemented using the Berkeley DB to record update operations that must survive a node crash. All the Berkeley DB files are managed using its ``environments'' feature that supports transactions through low-level logging. This architecture allows metadata changes to multiple files to be flushed with a sequential write to the low-level log.