The Helium cluster has two scratch file systems. One is /nfsscratch which is an NFS volume. This should meet the bulk of people's needs as it is very reliable. The other is a distributed file system that communicates between servers over the Infiniband fabric. This file system is mounted on /scratch and uses the Fraunhofer Parallel Cluster File System. While this can be significantly faster then the NFS based /nfsscratch, there are some ideosyncracies that may need to worked around, such as no support for hard links. If your work load writes out very large files than FhGFS (/scratch) may be worth trying.