Red Hat Knowledgebase

GFS performs best in an environment with few, large files and little contention. The larger the reads/writes and the more independence each node in the cluster has, the better the performance. This is not to say that bad performance will ensue if this model is deviated from. However, if you have thousands of small files that are constantly changing from different nodes in the cluster, you can bet that performance is going to suffer, as the file system has to "ping-pong" locks from node to node.

Tuning a GFS file system for performance is like tuning an application for performance on an SMP box. Sometimes, barriers and locking can slow down overall throughput of the application. Through careful planning, the time spent on barriers and locking can be minimized and performance of the application can be far above what could be accomplished with a single processor. Tune your GFS file system so that nodes play in separate portions of the file system, if possible. Creating many files in the same directory from different nodes can hurt GFS performance because the directory lock needs to be passed around. Try breaking up this single directory into multiple directories. There are many ways to improve GFS performance. Some things can be done on the fly, while others should be planned out ahead of time. If you are in doubt about what to do and want the highest possible performance, think about using Red Hat's Professional Services to guide you.