ANUSF
Fujitsu VPP300 Userguide

 

NQS Batch queues

Interactive user

Project Accounting

File Systems

Programming Languages

Programming Tools

Vectorization and Tuning Guide

HPC Programming Hints

 

 

Older doco that may still be relevant

VP2200 User Guide
 &  Vectorization

 

[Back to]
[VPP]

 

   

VPP300 File Systems

The filesystems listed below are in order of increasing IO performance but decreasing flexibility of use:

home file systems

VPP users have a "current work" home directory (initially empty) on the VPP primary PE, vpp.anu.edu.au, with a disk quota limit (initially set to 100MBs). It is envisaged that this home directory will be used for active VPP work. The command quota -v shows your limits and current usage in kbytes. There are a number of things users can do to minimize their disk use:

All VPP users have a home directory on the SUN UltraSPARC, covpp, which may be used for more permanent storage of VPP-related files. (Users should keep in mind that large, long term datasets may be stored on the silo.) The VPP home disks are NFS mounted on covpp and ~/vpphome in your covpp home directory is a symbolic link to your vpp home directory. Note that the VPP cannot access the rest of your covpp home directory; you must transfer files to ~/vpphome before using them on vpp. Small files can transferred by cp on covpp, large files should be ftp'ed across (much faster).


Short Term large capacity file systems

Short term extra file space is available under /short on request. Please email anusf@anu.edu.au to request this to be set-up for your jobs. Note that the default time that any file can remain unused on /short before automatic deletion is 7 days Send mail to anusf@anu.edu.au to request a variation of this default for your project needs. It is also important to realise that the files stored on /short are not backed up, hence it should never be used as the sole store of non reproducible data or source code.


VFL and scratch file systems

VFL (Very Fast and Large) file systems are intended for high-speed input and output of large amounts of data. VFL-FS is not suited (i.e. is very inefficient) for small, frequently accessed files. Please read the details of how to use the VFL-FS efficiently before doing so. Note that the VFL filesystems /vflvol0 and /vflvol1 are intended for large scratch files only. It is NOT backed up.

Two VFL file systems are available on the ANU VPP300: /vflvol0 (on IOPE, PE0) and /vflvol1 (on IOPE PE12). In addition /vfllocal (historically sometimes referred to as /vflvol) will get you to the system connected to the IOPE nearest to the PE which your job is executing. /vflvol0 or /vflvol1 should be used in all circumstances where you need to know where your files exist after the job has finished or if you have a non-failsafe task to be run at the end of your job (for example a copy over the network). The special link /vfllocal is primarily of use for jobs that use out-of-core solvers, such as Gaussian.


mrfs

It is now possible to configure a temporary memory resident filesystem (mrfs) for the duration of your batch job. IO bandwidth to an mrfs is the same as memory bandwidth (very fast) but size is limited to memory size and it is a truly temporary filesystem. Typical uses of such a filesystem would be for TMPDIR while doing large compilations or for Gaussian work files.

Specify the -lV size (where size is, for example, 50MB) to nqsub to reserve part of your batch memory to be used for your batch mrfs. Note that this memory is part of that specified in any -lM nqsub option or the per-request memory limit default if you do not specify -lM.

Specify the -cc option to select the mrfs "directory" as the current working directory of your batch request. This directory will be initially empty and all its contents will be deleted at the end of your batch request - it is the user's responsibility to include commands in the NQS script to do any necessary copying in and out of permanent files. Other than that, the mrfs is just like any other filesystem but file IO will be as fast as memory access.

Apart from using the -cc option to make the mrfs the batch jobs current working directory, the actual name of the directory is available through the environment varable MRFSDIR. This variable is set by NQS as part of initiating your job. You might notice that the mrfs directory is located as a subdirectory of /var/spool/nqs_mrfs/ - it is only accessible by your batch job.