Independent, parallel I/O is available. Users can access independent filesystems residing on different RAIDs. The available directories in which users may create their own subdirectories and deposit their own files are:
However, users may not see these files on all service nodes, a deficiency that will be addressed in a later OS release. Specifically, users need to be on service-16 to do all normal UNIX file operations such as ls, mkdir, rmdir, ftp. All of the /raid_xxi filesystems are NFS mounted there, and only there. For applications running on the compute nodes to access the files, the following calls with standard UNIX semantics are available:
All file accesses are independent. Every opened file has an associated, independent file offset pointer. It is allowable for multiple compute nodes to write to non-overlapping regions of a single file, but it would be at the users' own risk to write into overlapping regions. No file locking is provided.
Support of the independent file I/O is provided by fyod, a daemon process residing on the I/O nodes.
There is no filesystem that stripes across multiple RAIDs.
For normal use, it is strongly suggested that users continue to log onto system-assigned service nodes to launch applications, and log onto service-16 only to do necessary file managements.
I/O should be buffered through the use of an I/O library or the UNIX standard I/O library with buffer size enlarged to the MBytes range. Appropriately buffered I/O can achieve about 8 MB/s per individual file on a RAID. It has been observed that in an application (CTH), it takes significantly (a factor of 10) more time to buffer the data than to do the actual write.