Performance evaluation of meta-data transfer and storage in clusters

International audienceAn expressive number of scientific application executed on cluster architectures need some form of permanent storage with high capacity. As a consequence, the performance of the storage system becomes crucial to such applications. Besides granting data I/O performance, it's important to provide efficient and scalable access to file's meta-data, allowing the system to scale and, yet, offer a good level of consistency in the shared file system. Also the file transfer between nodes of the cluster architectures is an important factor that affects the I/O performance. And this is important to know the behavior and influence in the shared file system. In this paper we present a low-overhead meta-data synchronization algorithm used in dNFSp, a distributed file system based on NFS that proposes changes exclusively on the server side, keeping compatibility with traditional NFS clients. With this synchronization schema, dNFSp can offer the same level of coherence of traditional NFS servers. We also evaluate the communication characteristics of the environment where the tests were executed in order to analyze possibilities of scalability of the system