Atomic writes for data integrity and consistency in shared storage devices for clusters

The recent trend to use storage area networks in distributed and cluster systems has necessitated an improvement of integrity and consistency guarantees of stored data in the presence of node or network failures. Currently, the main method for preserving data integrity and consistency is logging, e.g. journaling. This paper presents a new, general method for preserving data consistency by atomic multi-block writes. The atomic writes method guarantees that either all the blocks in a write operation are written or no blocks are written at all. Its main advantage is that it does not require a recovery phase after a failure. The atomic writes method should be implemented in both the operating system and storage system levels. It is easy to use and implement. We present the method, its implementation and an example of its use for handling meta-data consistency in the Global File System (GFS).