Utilizing the DFS server with a GPFS file system has the potential to vastly alter the underlying GPFS caching needs. The DFS server must hold GPFS resources based on the cumulative activity of all DFS clients accessing the server. For performance and scalability reasons, DFS does not aggressively release held resources. Therefore, it is suggested that a DFS exported node be configured with the maximum data cache and file open cache possible. In addition, the DFS caching and coherency model is biased towards read sharing of file attributes and data which will reduce the GPFS performance in a multi-node writer scenario with concurrent DFS access. It is suggested that heavy use be avoided at times when maximum GPFS parallel efficiency is desired.
The DFS server optimistically grants file coherency guarantees and advisory byte range locking guarantees to requesting DFS clients. The DFS cache model optimizes the single writer and multiple reader case. However, GPFS pessimistically grants file coherency guarantees in order to allow for maximum concurrency for multiple writers. Additionally, DFS clients do not proactively return locking or coherency guarantees.
For further information on cache usage for GPFS, see the IBM General Parallel File System for AIX: Concepts, Planning, and Installation Guide sections on Cache and GPFS Configuration Considerations.