Radon - Getting Started

The Radon cluster is composed of desktop PCs recycled from instructional computing labs. Radon is currently entirely 64-bit Dell systems with Intel Pentium4 or Xeon processors of various speeds and with memory configurations between 2 and 4 GB of RAM. Nodes are connected with either 100 MB or Gigabit Ethernet. The machines reclaimed from instructional labs are older, slower, and lack high-speed interconnects, so high-communication or sizeable multithreaded programs are not a good fit. Still, there are a fair number of machines, and some codes may be able to take advantage of these effectively.

Detailed Hardware Specification

Radon is currently divided into three different sub-clusters, each with a different combination of CPU speed, memory, and interconnect. Subcluster "a" nodes have 3.6 GHz single-core Intel Pentium4 CPUs, 2 GB RAM, and Gigabit Ethernet; subcluster "b" nodes, 3.2 GHz dual-core Intel Xeon CPUs, 2 GB RAM, and Gigabit Ethernet; and subcluster "c" nodes, 3.2 GHz dual-core Intel Xeon CPUs, 4 GB RAM, and Gigabit Ethernet.

Sub-Cluster	Number of Nodes	Processor	Cores per Node	Memory per Node	Interconnect	Theoretical Peak TeraFLOPS
radon-a[001-144]	144	3.2 GHz Intel Pentium4	1	2 GB	Gigabit Ethernet	1.03
radon-b[001-048]	48	3.2 GHz Intel Xeon	2	2 GB	Gigabit Ethernet	0.61
radon-c[001-048]	48	3.2 GHz Intel Xeon	2	4 GB	Gigabit Ethernet	0.61

All Radon nodes run Red Hat Enterprise Linux 5 and use PBSPro 9.x for resource and job management. Operating system patches are applied monthly or as security needs dictate. All nodes have been configured to allow for unlimited stack usage, as well as unlimited core dump size (though disk space and server quotas may still be a limiting factor).

Obtaining an Account

Radon is a cluster operated by RCAC. Purdue faculty, staff, and students with the approval of their advisor may request access to Radon using the online Research Computing Account Request Form.

Login / SSH

To issue jobs on Radon, users may log on to the front-end host radon.rcac.purdue.edu via SSH.

SSH Client Software

All access to the RCAC systems must be through secure (encrypted) connections. Standard telnet and FTP are not supported. SSH, SCP, and SFTP may be used instead.

Secure Shell or SSH is a way of establishing a secure channel between a local and a remote computer. It uses public-key cryptography to authenticate the remote computer and (optionally) to allow the remote computer to authenticate the user. It is usually used to log in to a remote machine and execute commands similar to telnet, but it also supports tunneling and forwarding of X11 or arbitrary TCP connections. The associated SFTP and SCP protocols may be used to transfer files. There are many SSH clients available, depending on the operating system you use.

Linux / Solaris / AIX / HP-UX / Unix:

"ssh", "sftp", and "scp" are pre-installed. Log in using ssh myusername@servername.

Microsoft Windows:

PuTTY is an extremely small download and includes free SSH, SCP, and SFTP clients.
Secure CRT (SSH client) and Secure FX (SFTP client) are commercial clients which are available free to Purdue students, faculty, and staff with a Purdue career account.

Mac OS X:

"ssh", "sftp", and "scp" are pre-installed. You may start a local terminal window from "Applications->Utilities". Log in using ssh myusername@servername.
MacSSH and MacSFTP
NiftyTelnet 1.1 SSH

SSH Keys

SSH can be used in conjunction with many different means of authentication. One popular authentication method is called Public Key Authentication (PKA). PKA is a method of establishing your identity to a remote computer using related sets of encryption data called keys. PKA is a more secure alternative to traditional password-based authentication with which you are probably familiar.

To employ PKA via SSH, you manually generate a keypair (also called SSH keys) in the location from where you wish to initiate a connection to a remote machine. This keypair consists of two text files, one which is called a private key and one which is called a public key. You keep the private key file confidential on your local machine or local home directory (hence the name "private" key). You then login to a remote machine (if possible) and append the corresponding public key text to the end of a specific file, or have a system administrator do so on your behalf. In future login attempts, the public and private keys are compared to verify your identity, which then grants you access to the remote machine.

As a user, you can create, maintain, and employ as many keypairs as you wish. If you connect to a computational resource from your work laptop, your work desktop, and your home desktop, you can create and employ keypairs on each. You can also create multiple keypairs on a single local machine to serve different purposes, such as establishing access to different remote machines, or establishing different types of access to a single remote machine. In short, PKA via SSH offers a secure but flexible means of identifying yourself to all kinds computational resources.

Passphrases and SSH Keys

When a you create a keypair, you are prompted to provide a passphrase for the private key. This passphrase is different than a password in a number of ways. First, a passphrase is, as the name implies, a phrase. It can include most types of characters, including spaces, and has no limits on length. Second, this passphrase is not transmitted to the remote machine for verification. It is used only to allow the use of your local private key and is specific to a specific local private key.

Perhaps you are wondering why you would need a private key passphrase at all when using PKA. If the private key is kept secure, why the need for a passphrase just to use it? Indeed, if the location of your private keys were always completely secure, a passphrase might not be needed. In reality, a number of situations could arise in which someone may improperly gain access to your private key files. In these situations, a passphrase offers another level of security for you, the user who created the keypair.

Think of the private key/passphrase combination as being analogous to your ATM card/PIN combination. The ATM card itself is the object that grants access to your important accounts, and as such, should be kept secure at all times—just as a private key should. But if you ever lose your wallet or your ATM card is stolen, you are glad that your PIN exists to offer you another level of protection. The same is true for a private key passphrase.

When you create a keypair, you should always provide a corresponding private key passphrase. For security purposes, avoid using phrases that would be guessed by automated programs (e.g. phrases that consist solely of words in English-language dictionaries). This passphrase can never be recovered if forgotten, so make note of it. There are only limited situations when the use of a non-passphrase-protected private key is warranted—conducting automated file backups is one such situation. If you need to use a non-passphrase-protected private key to conduct automated backups to Fortress, see the No-Passphrase SSH Keys section.

Passwords

If you have received a default password as part of the process of obtaining your account, you should change it immediately when you log on for the first time. This can be done from any terminal/SSH session with the command "passwd". You will have the same password on all RCAC systems. If you change your password on any one RCAC system, it will change on all RCAC systems.

If you already have a Purdue career account, then you will initially be given the same userid and password as your career account. There is no need to change your career account password because you have received an account on RCAC systems.

There is not currently any requirement regarding how often you must change your password within RCAC, but for security reasons changing a password every six months, preferably every three months, is good practice.

All passwords should:

Be something you have never used as a password before, on this or any other system.
Be easy for you to remember and difficult for others to guess.
Be at least eight characters long.
Be a combination of upper and lowercase letters, numbers, and symbols.
TIP: Abbreviate a sentence or song lyric: "The dog Samson ate 4 new slippers!" = "TdSa4ns!"

Never share your password with another user or make your password known to anyone else. Systems staff will NEVER ask for your password, by email or otherwise.

Storage Options

File storage options on RCAC systems include home directories, scratch file systems, /tmp, and long-term or permanent storage. Each of these have different performance and intended uses, and some vary from system to system as well. Home directories and long-term storage are backed up nightly, but scratch and /tmp are not and may be occasionally purged without warning. Below is more detail about each of these storage options.

Home Directories

Your home directory is the default directory you are placed in when you log in.

You should use this space for storing files you want to keep long term such as source code, scripts, input data sets, etc. It should also be used for files you want to keep and which you use often. The home directory will physically reside on the BlueArc NFS Server. You can find the path to your home directory by logging in, and typing pwd:

$ pwd
/home/ba01/u103/myusername

The second component of the reply indicates the name of the host where your home directory physically resides. In this example, the home directory is on the RCAC home directory file server named "ba01" under area "u103". This will vary from person to person. Remember, you can always check where your home directory is located by doing a pwd command in your home directory.

Regardless of its physical location, your home directory and its contents are available on almost all the RCAC front-end hosts and their nodes via the Network File System (NFS). The only exception is Black.

Note that your home directory has a quota capping the size and/or number of files you may store within. For more information, refer to the Storage Quotas / Limits Section.

Lost Home Directory File Recovery

Only files which have been backed up overnight can be recovered. If you lose a file the same day you created it, it can NOT be recovered.

Files Lost Less than Seven Days Ago

For files lost less than seven days ago, RCAC has implemented self-service file recovery. Backups of all your files are made at midnight daily and you may access these directly.

To recover files lost after midnight today (same day as loss):

$ set BACKUP=`echo $HOME | sed "s,/u1,_snap/backup_snap/u1,;s,/home/,/autohome/,"`
$ cd $BACKUP

  (now locate the file or directory you wish to recover within here)

$ cp mylostfile $HOME
  (or)
$ cp -r mylostdir $HOME

To recover files lost prior to today, but in last week (2-7 day loss):

  (set this to the date you lost the files: 4-digit year, 2-digit month, 2-digit day)
$ set DATE=YYYYMMDD

$ set BACKUP=`echo $HOME | sed "s,/u1,_snap/backup_snap_$DATE*/u1,;s,/home/,/autohome/,"`
$ cd $BACKUP

  (now locate the file or directory you wish to recover within here)

$ cp mylostfile $HOME
  (or)
$ cp -r mylostdir $HOME

Files Lost More than Seven Days Ago

For files lost more than seven days ago, you will need to request RCAC recover your files from backup tapes. Please do so using the flost command from the front-end host of an RCAC resource:

$ flost

Scratch Directories

Scratch directories are provided by RCAC and are intended for short-term file storage only.

Backups are not performed on scratch directories. In the event of a disk crash or file purge, files in scratch directories can not be recovered. Please be sure to copy any important files to more permanent storage.

All files stored in RCAC scratch directories older than 90 days will be automatically removed (purged). Owners of these files will be notified one week before removal via email. For more information, please refer to our Scratch File Purging Policy.

RCAC scratch directories are provided by a central BlueArc server and are accessible from most RCAC systems. There are two primary scratch file systems: scratch95 and scratch96. A scratch directory already exists for all Radon users. Your RCAC scratch directory is located under scratch95 or scratch96 within a subdirectory by the first letter of your username.

To find the path to your RCAC scratch directory, run myscratch:

	$ myscratch
	/scratch/scratch96/m/myusername

The variable $RCAC_SCRATCH is also set to your RCAC scratch directory path. Use this variable in any scripts. Your actual scratch directory path may change without warning, but this variable will remain current.

$ echo $RCAC_SCRATCH
/scratch/scratch96/m/myusername

To find the path to someone else's RCAC scratch directory, use the command findscratch:

$ findscratch someuser
/scratch/scratch95/s/someuser

Note that your RCAC scratch directory has a quota capping the size and/or number of files you may store within. For more information, refer to the Storage Quotas / Limits Section.

/tmp Directory

The /tmp directory is intended for temporary files that are used during the execution of a process or job or while you examine files created by your jobs. Used properly, /tmp may provide faster local storage to an active process than any other storage option. However, do not use it for longer-term storage or critical results.

Files stored in /tmp are not backed up and are removed whenever space is low or whenever the system is rebooted. In the event of a loss, files in /tmp can not be recovered, so use it only for files that can be recreated relatively easily.

Long-Term Storage

Long-term Storage or Permanent Storage is available to RCAC users on the DXUL/UniTree archival storage system, commonly referred to as "Fortress". DXUL (DiskXtender for Unix and Linux) and UniTree are a software package that manages a hierarchical storage system. Program files, data files and any other files which are not used often, but which must be saved, can be put in permanent storage. Fortress currently has a 1.2 PB capacity. However, since two copies are retained for every file, the usable capacity is only 600 TB.

Recently used files smaller than 0.5 MB have their primary copy stored on low-cost disks, but the second copy is on tape or optical disks. This provides a rapid restore time to the disk cache. However, the large latency to access a larger file (usually involving a copy from a tape cartridge) makes it unsuitable for use as active storage.

In addition to poor performance, these two uses can cause severe problems with the system itself:

DO NOT store any actively used files on Fortress.
DO NOT store large collections of small files on Fortress.

Do not use Fortress as a second home directory. Instead, use tar or some similar archive tool to combine all the smaller files you wish to store into a single large file first.

For active data storage you should use either local storage or a scratch file system. You may then copy any results you wish to archive to Fortress when computation is complete.

Fortress is directly accessible (via FTP, SSH, SCP, SFTP, and NFS) from all RCAC systems, as well as most systems in ECN and CS and from several other major servers on campus. To access Fortress in any way other than NFS, you must login to fortress.rcac.purdue.edu. RCAC has more information about Fortress, including how to obtain a Fortress account and how to access your files on Fortress.

Environment Variables

There are many environment variables related to storage locations and paths which are automatically set for you upon log-in, or may be changed if necessary. In addition, many more environment variables are set for specific applications, such as compilers, when "modules" for these applications are loaded. (See the module command section for more information.)

Use environment variables instead of actual paths whenever possible to avoid problems if the specific paths to any of these change. Some of the environment variables you should have are:

$USER: your username
$HOME: path to your home directory
$PWD: path to your current directory
$RCAC_SCRATCH: path to scratch filesystem
$PATH: all directories searched for commands/applications
$HOSTNAME: name of the machine you are on
$SHELL: your current shell (bash, tcsh, csh, ksh)
$SSH_CLIENT: your local client's IP address
$TERM: type of terminal or terminal emulator being used
$OMP_NUM_THREADS: number of OpenMP threads

All environment variables begin with the dollar sign ($) and are all uppercase. They may be used on the command line or in any scripts in place of and in combination with hard-coded values:

$ ls $HOME
...

$ ls $RCAC_SCRATCH/myproject
...

$ ls $RCAC_SCRATCH/myproject/$HOSTNAME_data
...

You may find the value of any environment variable by using the echo command:

$ echo $RCAC_SCRATCH
/scratch/scratch95/m/myusername

$ echo $SHELL
/usr/local/bin/tcsh

You may list the values of all environment variables using the env command:

$ env
USER=myusername
HOME=/home/ba01/u101/myusername
RCAC_SCRATCH=/scratch/scratch95/m/myusername
SHELL=/usr/local/bin/tcsh
...

You may create or overwrite an environment variable using either export or setenv, depending on your shell:

  (for bash and sh)
$ export VARIABLE=value

  (for tcsh and csh)
% setenv VARIABLE value

Storage Quotas / Limits

Your disk usage is limited on RCAC systems. However, each filesystem (scratch, home directory, etc.) may have a different limit. If you exceed the soft limit or quota, you will see warnings whenever writing to the disk that you are over quota, but the write will still succeed. If you exceed the hard limit or limit, your write will fail until you either remove other files or your quota is increased. Generally, RCAC systems do not impose a soft limit—only a hard limit.

Checking Quota Usage

You may find out what your current quota is by using the quota command:

$ quota
Disk quotas for user myusername (uid 12345): 
     Filesystem  blocks   quota   limit   grace   files   quota   limit   grace
     ba01:/u103 2346272       0 5000000           17508       0   65535

The columns are as follows:

Filesystem: This indicates the line is for the user's files on /u103/, which doing echo $HOME confirms is the user's home directory filesystem.
Blocks: This shows how many 1 KB blocks the user's files take up. In this case, 2346272 KB / 1024 = 2291 MB, or 2291 MB / 1024 = 2.24 GB.
Quota: This shows that soft limits are not being imposed (0).
Limit: This shows how many 1 KB blocks the user's hard limit is. In this case, (5000000 KB / 1024) / 1024 = 4.77 GB.
Grace: This would show the grace period (in days) for any soft limit (none in this case).
Files: This shows how many file pointers (inodes) the user is currently using. This is based more on the number of files and directories and not the size.
Quota: This shows that soft limits are not being imposed for file pointers (0).
Limit: This shows the user's file pointer hard limit. It is possible, though unlikely, to hit this and not the size limit if you create a large number of very small files.
Grace: This would show the grace period (in days) for any file pointer soft limit (none in this case).

You may also see the disk usage of any given directory by using du:

$ du -hs
1.1G    .

$ du -hs $HOME
138M    /home/ba01/u103/myusername

This can be very helpful in figuring out where your largest files or directories are, so that you may clean out unneeded large files and avoid hitting your quota.

Requesting Quota Increase

If you find you need additional disk space on an RCAC account, please first consider archiving and compressing old files and moving them to long-term storage. If this option does not resolve the issue, you may send an email to rcac-help@purdue.edu and request additional space.

Archive and Compression

There are several options for archiving and compressing groups of files or directories on RCAC systems. All of the following tools are provided:

zip (more information)
Simple compression and file packaging utility.
Examples:

  (compress file somefile.c)
$ zip somefile.zip somefile.c

  (extract contents of somefile.zip)
$ unzip somefile.zip

  (compress all files in a directory into one archive file)
$ zip -r somefile.zip somedirectory/

  (compress all ".c" files in current directory into one archive file)
$ zip -r somefile.zip . -i \*.c

tar (more information)
Saves many files together into a single archive file, and restores individual files from the archive. Includes automatic archive compression/decompression options and special features that allow tar to be used for incremental and full backups.
Examples:

  (archive file somefile.c)
$ tar cvf somefile.tar somefile.c

  (archive and compress file somefile.c)
$ tar czvf somefile.tar.gz somefile.c

  (list contents of archive somefile.tar)
$ tar tvf somefile.tar

  (extract contents of somefile.tar)
$ tar xvf somefile.tar

  (extract contents of gzipped archive somefile.tar.gz)
$ tar xzvf somefile.tar.gz

  (archive and compress all files in a directory into one archive file)
$ tar czvf somefile.tar.gz somedirectory/

  (archive and compress all ".c" files in current directory into one archive file)
$ tar czvf somefile.tar.gz *.c

gzip (more information)
Compression utility designed as a replacement for compress, with much better compression and no patented algorithms. The standard compression system for all GNU software.
Examples:
```
  (compress file somefile - also removes uncompressed file)
$ gzip somefile

  (uncompress file somefile.gz - also removes compressed file)
$ gunzip somefile.gz
```

bzip2 (more information)
Strong, lossless data compressor based on the Burrows-Wheeler transform. Also available as a library.
Examples:

  (compress file somefile - also removes uncompressed file)
$ bzip2 somefile

  (uncompress file somefile.bz2 - also removes compressed file)
$ bunzip2 somefile.bz2

compress (more information)
Adaptive Lempel-Ziv compressor. Not often used today.

Windows users can work with these same formats using some of the following software:

7-Zip
Free Windows software package that can handle all the above formats.
WinZip
Commercial Windows software package that can handle all the above formats.
WinRAR
Commercial Windows software package that can handle all the above formats.

File Transfer

There are a variety of ways to transfer data to and from RCAC systems. Which you should use depends on several factors, including the ease of use for you personally, connection speed and bandwidth, the size and number of files to be transferred. For more details on file transfer methods and applications, refer to the Radon Complete User Guide.

Provided Applications

The third-party software on three commonly used RCAC systems is shown in the following table. Additional software may be available on other RCAC systems, and the software on a specific system can be seen by running the command "module avail" on that system. Please contact rcac-help@purdue.edu if you are interested in the availability of software not shown in this list.

	Radon	Steele	Julius/Caesar
R	✓	✓
AcGrace		✓
Amber	✓	✓
ANSYS	✓	✓
ATLAS	✓
BinUtils		✓
Boost	✓
ClustalX	✓
COMSOL	✓	✓	✓
CPLEX	✓	✓	✓
CUDA		✓
DX		✓
Ferret		✓
FFTW	✓	✓
FLUENT	✓	✓
GAMESS	✓	✓
GAMS	✓	✓
Gaussian	✓	✓	✓
GCC Compiler (C, C++, Fortran)	✓	✓	✓
GCC IA64 Cross-Compiler (xgcc-ia64)		✓
GMP	✓
GMT	✓	✓
GrADS		✓
GROMACS		✓
GhostScript		✓
GSL		✓	✓
HDF4 (Compiled for Intel, GNU, PGI)		✓
HDF5 (Compiled for Intel, GNU, PGI)		✓
ImageMagick		✓
IMSL	✓	✓	✓
Intel Compiler (C, C++, Fortran)	✓	✓	✓
Jasper	✓	✓
Java	✓	✓
LAM			✓
LAMMPS	✓	✓
LSTC			✓
Maple	✓	✓	✓
Mathematica	✓	✓	✓
MATLAB	✓	✓
Mitrionics FPGA Tools (mitrion)		✓
MPFR	✓	✓	✓
MPICH	✓	✓	✓
MPICH2		✓	✓
MPIExec		✓
MrBayes	✓
MUMPS		✓
MVAPICH (for Intel, PGI compilers)		✓
MVAPICH2 (for Intel, GNU, PGI compilers)		✓
MWRank	✓
NCBI	✓
NCL	✓	✓
NCO		✓
NetCDF (for Intel, GNU, PGI compilers)		✓
NTL	✓
NWChem	✓	✓
Octave		✓
PGI Compiler (C, C++, Fortran)	✓	✓
PKG-Config		✓
Python	✓	✓
RASC		✓	✓
SAS	✓
ScaLAPACK	✓	✓	✓
Stata			✓
Subversion	✓	✓	✓
Tau		✓
TecPlot		✓
TotalView	✓	✓	✓
UDUNITS		✓
VASP			✓
Vis5D		✓

Environment Management with the Module Command

RCAC uses the module command as the preferred method for a user to manage the processing environment. With this command, a user may load libraries and paths for using specific applications or compilers. These are organized into packages which may be loaded and unloaded as needed. Please use the module command and do not manually configure your environment, as RCAC staff will frequently make changes to the specifics of various packages. If you use the module command to manage your environment, these changes will not be noticeable.

Below follows a short introduction to the module command. You can see more in the man page for module. Typing module at the command line will give you a brief usage report.

List Available Modules

To see what modules are available on this system, use the "module avail" command:

$ module avail

------------------------ /apps/host/modules/versions -------------------------
3.1.6

-------------------- /apps/host/modules3.1.6/modulefiles ---------------------
dot         module-cvs  module-info modules     null        use.own

----------------------- /apps/host/modules/modulefiles -----------------------
R/2.6.2
R/2.7.0
amber/10
ansys/11.0
dx/4.4.4
fftw/2.1.5
fftw/3.1.2
fluent/6.3.26
gamess/24.MAR.2007.R3(default)
gaussian/D.01
gaussian/E.01(default)
gaussian03/D.01
gaussian03/E.01(default)
gcc/4.3.0
     ...

Load / Unload a Module

You should note that all modules consist of both a name and a version number. When loading a module, you may use only the name to load the default version, or specify which version you wish to load:

$ module load intel
  (load default Intel compiler)

$ module load intel/9.1.045
  (load version 9.1.045 of the Intel compiler)

Note that you will need to load any relevant modules within job submission scripts that use those applications. Loading the module before submitting your job is not sufficient. Also, if you use bash or ksh as your login shell, you will also need to add a line in any submission script to source /etc/profile before invoking "module". Users of csh and tcsh do not need to do this.

     ...
. /etc/profile
module load intel
     ...

To unload a module, use the “module unload” command. It will attempt to undo the changes to your environment, made by that module:

$ module unload intel
  (unload default Intel compiler)

$ module unload intel/9.1.045
  (unload version 9.1.045 of the Intel compiler)

List Currently Loaded Modules

To see what modules you have currently loaded, use "module list":

$ module list
Currently Loaded Modulefiles:
  1) intel/9.1.045

$ module unload intel
$ module list
No Modulefiles Currently Loaded.

Provided Compilers

Compilers are available on Radon for Fortran 77, Fortran 90, Fortran 95, C, and C++. The compilers can produce general-purpose and architecture-specific optimizations to improve performance. These include loop-level optimizations, inter-procedural analysis and cache optimizations. The compilers support automatic and user-directed parallelization of Fortran, C, and C++ applications for multiprocessing execution. More detailed documentation on each compiler set available on Radon follows.

Intel Compiler Set

To use the Intel compiler set (compilers and associated libraries) on Radon, load the "intel" module, using the "module" command.

Here are some examples for the Intel compilers:

Language	Serial Program	MPI Program	OpenMP Program
Fortran77	$ module load intel $ ifort myprogram.f -o myprogram	$ module load mpich2-intel $ mpif77 myprogram.f -o myprogram	$ module load intel $ ifort -openmp myprogram.f -o myprogram
Fortran90	$ module load intel $ ifort myprogram.f90 -o myprogram	$ module load mpich2-intel $ mpif90 myprogram.f90 -o myprogram	$ module load intel $ ifort -openmp myprogram.f90 -o myprogram
Fortran95	(not available)	(not available)	(not available)
C	$ module load intel $ icc myprogram.c -o myprogram	$ module load mpich2-intel $ mpicc myprogram.c -o myprogram	$ module load intel $ icc -openmp myprogram.c -o myprogram
C++	$ module load intel $ icc myprogram.cpp -o myprogram	$ module load mpich2-intel $ mpiCC myprogram.cpp -o myprogram	$ module load intel $ icc -openmp myprogram.cpp -o myprogram

Other versions of the Intel compiler and/or libraries may also be available. To see which versions are currently installed, use the command "module avail".

More information on compiler options can be found in the official man pages, which can be accessed if the appropriate module is loaded using the "man" command, or online here:

Here is some more documentation from other sources on the Intel compilers:

GNU Compiler Set

The official name of the GNU compilers is 'GNU Compiler Collection' or 'GCC'. To use the GNU compiler set (compilers and associated libraries) on Radon, load the "gcc" module, using the "module" command.

An older version of the GNU compiler will be in your path by default. Do NOT use this version. Instead, load the newer version using the "module" command.

Here are some examples for the GNU compilers:

Language	Serial Program	MPI Program	OpenMP Program
Fortran77	$ module load gcc $ gfortran myprogram.f -o myprogram	$ module load mpich2-gcc $ mpif77 myprogram.f -o myprogram	$ module load gcc $ gfortran -fopenmp myprogram.f -o myprogram
Fortran90	$ module load gcc $ gfortran myprogram.f90 -o myprogram	$ module load mpich2-gcc $ mpif90 myprogram.f90 -o myprogram	$ module load gcc $ gfortran -fopenmp myprogram.f90 -o myprogram
Fortran95	$ module load gcc $ gfortran myprogram.f95 -o myprogram	(not available)	$ module load gcc $ gfortran -fopenmp myprogram.f95 -o myprogram
C	$ module load gcc $ gcc myprogram.c -o myprogram	$ module load mpich2-gcc $ mpicc myprogram.c -o myprogram	$ module load gcc $ gcc -fopenmp myprogram.c -o myprogram
C++	$ module load gcc $ g++ myprogram.cpp -o myprogram	$ module load mpich2-gcc $ mpiCC myprogram.cpp -o myprogram	$ module load gcc $ g++ -fopenmp myprogram.cpp -o myprogram

Other versions of the GNU compiler and/or libraries may also be available. To see which versions are currently installed, use the command "module avail".

More information on compiler options can be found in the official man pages, which can be accessed if the appropriate module is loaded using the "man" command, or online here:

Here is some more documentation from other sources on the GCC compilers:

http://gcc.gnu.org/onlinedocs/

PGI Compiler Set

To use the PGI compiler set (compilers and associated libraries) on Radon, load the "pgi" module, using the "module" command.

Here are some examples for the PGI compilers:

Language	Serial Program	MPI Program	OpenMP Program
Fortran77	$ module load pgi $ pgf77 myprogram.f -o myprogram	$ module load mpich2-pgi $ mpif77 myprogram.f -o myprogram	$ module load pgi $ pgf77 -mp myprogram.f -o myprogram
Fortran90	$ module load pgi $ pgf90 myprogram.f90 -o myprogram	$ module load mpich2-pgi $ mpif90 myprogram.f90 -o myprogram	$ module load pgi $ pgf90 -mp myprogram.f90 -o myprogram
Fortran95	$ module load pgi $ pgf95 myprogram.f95 -o myprogram	(not available)	$ module load pgi $ pgf95 -mp myprogram.f95 -o myprogram
C	$ module load pgi $ pgcc myprogram.c -o myprogram	$ module load mpich2-pgi $ mpicc myprogram.c -o myprogram	$ module load pgi $ pgcc -mp myprogram.c -o myprogram
C++	$ module load pgi $ pgCC myprogram.cpp -o myprogram	$ module load mpich2-pgi $ mpiCC myprogram.cpp -o myprogram	$ module load pgi $ pgCC -mp myprogram.cpp -o myprogram

Other versions of the PGI compiler and/or libraries may also be available. To see which versions are currently installed, use the command "module avail".

More information on compiler options can be found in the official man pages, which can be accessed with the "man" command after loading the appropriate compiler module.

Here is some more documentation from other sources on the PGI compilers:

http://www.pgroup.com/doc/pgiug.pdf [PDF]

Running Jobs on Radon

There are a number of different compilers and programs installed on the RCAC systems. To access them, use module load <program>. To see the available modules, type module avail. To read more about the "module" command, look here.

There are two methods for submitting jobs to the Radon community cluster. First, you may submit jobs directly to a queue on Radon. These jobs may be serial, message-passing, or shared-memory in nature. You use the Portable Batch System (PBS) to submit jobs to a queue. Secondly, the Radon cluster is a part of BoilerGrid. You may submit serial jobs to BoilerGrid and specifically request that the serial jobs be run on the resources on Radon.

Running Jobs via PBS

Radon uses PBS version 9.x. The newer versions have a few minor differences from the older versions (before 8.0).

Differences are mainly:

Use of "cpp" no longer supported
Hosts in $PBS_NODEFILE ordered differently

The Portable Batch System (PBS) is a richly featured workload management system providing job scheduling and job management interface on computing resources, including Linux clusters. With PBS, a user requests resources and submits a job to a queue. A description of Radon's queues follows further down.

Note that you should never run big, long, multi-threaded, or CPU-intensive jobs on the front-end host. The front-end hosts are community-owned and running anything but the smallest test-job will slow them down for everyone. Use PBS to submit the job as a job submission file (called a job script in the official manual) or run an interactive PBS job session.

Radon PBS Tips

PBS Queues. Always use qstat -Q to determine which queues are available. There will usually be queues which are available to everyone with an account on that system. On most systems these are called either "standby" or "workq".
Any program that is installed can be run interactively. You must still use the "module load <program>" command to access it.
Programs which open a display can also be run interactively. Just use the -v DISPLAY option to qsub.
You can see which nodes you are using on one of the cluster machines with the command: cat $PBS_NODEFILE
When running OpenMP programs, you need the processors to be on the same node to get the advantage of shared memory.
The order of the processors is random. There is no way to tell which processor will do what and in which order in a parallel program.
Remember that ncpus can not be larger than the number of processors on each node on the machine in question.

Radon PBS Queues

To see a list of queues, type qstat -Q.

There are a total of 2 queues, but 'workq' is the default one to which everyone has access. It has a default wall time limit of 30 minutes, and maximum wall time limit of 336 hours.

No priorities are currently enforced on the ITaP/RCAC Linux clusters. All users have an equal chance at resources (note that this is not to say that all jobs have an equal chance at resources).

The nodes in Radon are partitioned into subclusters. By default, user jobs will be given nodes from the same subcluster (meaning that all machines will share an Ethernet switch and will have the highest possible interconnect bandwidth). If it is necessary to run jobs that span multiple subclusters, or to run a job that uses more than the number of nodes in any one subcluster, please contact rcac-help@purdue.edu.

The command qstat -Q will give you a list of all the queues. If you use qstat -Q -f it gives you more information about the various queues, including number of jobs, users and walltime.

Queues and some information about them:

Name of queue	Node Count	Default walltime	Max walltime
workq	240/240	00:30:00	336:00:00
radon_hold	N/A	00:30:00	N/A

Node Count is the number of accessible nodes for the queue. It is given as max/default.

Note that the number of processors that can be used for a certain queue is the total number of processors that are available to that queue. It does not mean that every submitter to that queue can get "Node Count" processors. For example, if a queue has a total of 32 accessable processors and one user has requested 16 processors and another has asked for 12 processors, then another submitter to that queue can only get 4 processors as long as the others are in use. If this user asks for more than 4 processors, his/her job will wait in the queue until enough processors are available.

Radon PBS Submission Script

A job submission file (job script) can contain any of the commands that you would otherwise issue yourself from the command line. You can, for example both compile and run a program and also set any necessary environment values. The results from compiling or runnning your programs can usually be seen after the job has run. They will show up in your directory as the files <script_name>.e<job number> and <script_name>.o<job number>. The first file will contain any errors that were reported (hopefully none), and the second file will give any results that your program may have output to the screen. If the program is supposed to write the results to a file, this will of course still happen. The job number is a number which PBS gives to every job. This will be reported when the job is submitted.

It may take quite a while before the job finishes running. How long will, among other things, depend on the number of nodes you have requested, how large the program is, which queue you are running it in, and how many other people are using the system at the same time.

A job submission file may consist of PBS directives, comments and executable statements. A PBS directive provides a way of specifying job attributes in addition to the command line options. For example:

	#PBS -N Job_name
	#PBS -l select=4:mem=320kb,walltime=10:30
	#PBS -m be
	#
	step1 arg1 arg2
	step2 arg3 arg4

The -N Job-name replaces script_name of the error and output files. The qsub command scans the lines of the script file for directives. An initial line in the script that begins with the characters "#!" or the character ":" will be ignored and scanning will start with the next line. Scanning will continue until the first executable line, that is a line that is not blank, not a directive line, nor a line whose first non-white space character is "#". If directives occur on subsequent lines, they will be ignored.

The remainder of the directive line consists of the options to qsub in the same syntax as they appear on the command line. The option character is to be preceded with the "-" character.

If an option is present in both a directive and on the command line, that option and its argument, if any, will be ignored in the directive. The command line takes precedence.

If an option is present in a directive and not on the command line, that option and its argument, if any, will be processed as if it had occurred on the command line.

How you run a program depends on whether it is a serial program, an OpenMP program, or a MPI program. There is no difference in how to run the program for the various compilers.

Important: You must 'module load' the same compiler (and MPICH2 if needed) that you used for compiling. Note that it is not necessary to load the standard compiler if you have loaded the corresponding compiler with the MPICH2 libraries included.

PBS Job Submission

The command to submit the job submission file is the following:

	qsub -q standby -l select=4,walltime=1:00 run_hello

This example submits a job to queue 'standby' and requests 4 nodes. It has a walltime of 1 minute. The job submission file is called run_hello. The names of the queues will be different on the various RCAC systems. You can find a list of their names with the command qstat -q or look at the section 'Queues'.

Some useful options for the qsub command includes (in the list below, note that a chunk is defined as a set of resources that are to be allocated as a unit):

-q <name>: tells which queue you want the job to run in. A list of available queues can be seen using the command qstat -Q. If none is choosen, the batch server will be the default server.
-l select=[N:]chunk[+[N:]chunk ...], where N specifies how many of that chunk, and a chunk is of the form: resource_name=value[:resource_name=value ...]

Job-wide resource_name=value requests are of the form: -l resource_name=value[,resource_name=value ...]. The most important resource_name's are: node (required), ncpus (how many processors), mpiprocs (how many processes).
The place statement has this form: -l place=[ arrangement ][: sharing ][: grouping] where
- arrangement is one of free | pack | scatter
- sharing is one of excl | shared
- grouping can have only one instance of group=resource
and where
- free: Place job on any node(s). Only good if you have a job that does not need much memory, so you do not mind it sharing the node with other. Will most likely give you access quicker than the other options.
- pack: You will get processors on one node only - all jobs will be placed on one node. Good for OpenMP.
- scatter: The chunks with any MPI processes will be spread out across as many of the nodes as possible, attempting to put only one process on each. A chunk with no MPI processes may be taken from the same node as another chunk.
- excl: Only this job uses the nodes chosen.
- shared: This job can share the nodes chosen.
- group=resource: Chunks will be grouped according to a resource. All nodes in the group must have a common value for the resource, which can be either the built-in resource host or a site-defined node-level resource.
Note that nodes can have sharing attributes that override job placement requests.
-I: Job is to be run interactively.
-v variable_list: Expands the list of environment variables that are exported to the job. This can also be environment variables from the qsub command environment which are made available to the job when it executes. The variable_list is a comma separated list of strings of the form variable or variable=value. These variables and their values are passed to the job.
-V: Declares that all environment variables in the qsub command's environment are to be exported to the batch job.

Note that ncpus can not be larger than the number of processors on each node on the machine in question.

Some environment variables can be set. They are then available to PBS. They include:

PBS_O_HOST: the name of the host upon which the qsub command is running
PBS_O_QUEUE: the name of the original queue to which the job was submitted
PBS_O_SYSTEM: the operating system name given by uname -s on the host on which qsub is running
PBS_O_WORKDIR: the absolute path of the current working directory of the qsub command
PBS_ENVIRONMENT: set to PBS_BATCH to indicate the job is a batch job, or to PBS_INTERACTIVE to indicate the job is a PBS interactive job, see -I option
PBS_JOBID: the job identifier assigned to the job by the batch system
PBS_JOBNAME: the job name supplied by the user
PBS_NODEFILE: the name of the file containing the list of nodes assigned to the job
PBS_QUEUE: the name of the queue from which the job is executed

If you wish to interrupt qsub prior to job start (before you get a command-line prompt), this can be done by typing control-C. It will then query if the user wishes to exit. If the user responds "yes", qsub exits and that job is aborted.

Instead of using a job submission file, qsub also accepts commands from standard input - the keyboard. To use this option, avoid giving a script operand or give the single character "-". When the script is being read from Standard Input, qsub will copy the file to a temporary file. This temporary file is passed to the library interface routine pbs_submit. The temporary file is removed by qsub after pbs_submit returns or upon the receipt of a signal which would cause qsub to terminate.

Once the job has started execution, input to and output from the job pass through qsub. Keyboard-generated interrupts are passed to the job. Entries beginning with the tilde ('~') character and containing special sequences are escaped by qsub. The recognized escape sequences are:

	~.      Qsub terminates execution. The batch job is 
                also terminated.
	
	~susp   Suspend the qsub program if running under the C shell. 
	        "susp" is the suspend character, usually CNTL-Z.
	
	~asusp  Suspend the input half of qsub (terminal to job), 
	        but allow output to continue to be displayed. Only 
                works under the C shell.  "asusp" is the auxiliary 
                suspend character, usually CNTL-Y.

If no script is provided, the qsub command reads the script from standard input. When the script is being read from Standard Input, qsub will copy the file to a temporary file. This temporary file is passed to the library interface routine pbs_submit. The temporary file is removed by qsub after pbs_submit returns or upon the receipt of a signal which would cause qsub to terminate.

Note: The following warning applies for users of the c-shell, csh. If the job is executed under the csh and a .logout file exists in the home directory in which the job executes, the exit status of the job is that of the .logout script, not the job submission file. This may impact any interjob dependencies. To preserve the job exit status, either remove the .logout file or place the following line as the first line in the .logout file:

	set EXITVAL = $status

and the following line as the last executable line in .logout

	exit $EXITVAL

PBS Job Status

Using the command qstat -a will show you the jobs currently running and their ID's.

Example (run on Steele):

$ qstat -a 

steele-adm.rcac.purdue.edu:
                                                            Req'd  Req'd   Elap
Job ID          Username Queue    Jobname    SessID NDS TSK Memory Time  S Time
--------------- -------- -------- ---------- ------ --- --- ------ ----- - -----
77025.steele-ad user123  standby  hello         --    1   8    --  00:05 Q   --
115505.steele-a user456  ncn      job4         5601   1   1    --  600:0 R 575:0
...
189479.steele-a user456  standby  AR4b          --    5  40    --  04:00 H   --
189481.steele-a user789  standby  STDIN        1415   1   1    --  00:30 R 00:07
189483.steele-a user789  standby  STDIN        1758   1   1    --  00:30 R 00:07
189484.steele-a user456  standby  AR4b          --    5  40    --  04:00 H   --
189485.steele-a user456  standby  AR4b          --    5  40    --  04:00 Q   --
189486.steele-a user123  tg_workq STDIN         --    1   1    --  12:00 Q   --
189490.steele-a user456  standby  job7        26655   1   8    --  04:00 R 00:06
189491.steele-a user123  standby  job11         --    1   8    --  04:00 Q   --
$

Where 'Q' = Queued, 'R' = Running, and 'H' = Held.

The list can be very long, making it difficult to find your own runs. If that is the case, use the following command to ask for jobs submitted by a specific user:

$ qstat -a -u user123

steele-adm.rcac.purdue.edu: 
                                                            Req'd  Req'd   Elap
Job ID          Username Queue    Jobname    SessID NDS TSK Memory Time  S Time
--------------- -------- -------- ---------- ------ --- --- ------ ----- - -----
182792.steele-a user123  standby job1   28422   1   4    --  23:00 R 20:19
185841.steele-a user123  standby job2   24445   1   4    --  23:00 R 20:19
185844.steele-a user123  standby job3   12999   1   4    --  23:00 R 20:18
185847.steele-a user123  standby job4   13151   1   4    --  23:00 R 20:18
$

PBS Job Cancellation

Stopping the job before it finishes.

qdel <job id>

You get the job id from the qstat -a or qstat -a -u [username] command.

PBS Interactive Jobs

To use the PBS queue interactively, you have to use the -I option. The command to submit such a job would then be like this command

	qsub -I -q standby -l select=2:ncpus=2

Where the options used means the following:

-q <name>: the queue you want the job to run in. A list of available queues can be seen using the command qstat -Q. If none are choosen, the batch server will be the default server.
-l select=[N:]chunk[+[N:]chunk ...], where N specifies how many of that chunk, and a chunk is of the form: resource_name=value[:resource_name=value ...]

Job-wide resource_name=value requests are of the form: -l resource_name=value[,resource_name=value ...]. An example of a resource_name is ncpus, which is number of CPU's.

The place statement has this form: -l place=[ arrangement ][: sharing ][: grouping]

where

- arrangement is one of free | pack | scatter
- sharing is one of excl | shared
- grouping can have only one instance of group=resource

and where

free: Place job on any node(s). Only good if you have a job that does not need much memory, so you do not mind it sharing the node with others. Will most likely give you access quicker than the other options.
pack: You will get processors on one node only - all jobs will be placed on one node. Good for OpenMP.
scatter: The chunks with any MPI processes will be spread out across as many of the nodes as possible, attempting to put only one process on each. A chunk with no MPI processes may be taken from the same node as another chunk.
excl: Only this job uses the nodes chosen.
shared: This job can share the nodes chosen.
group=resource: Chunks will be grouped according to a resource. All nodes in the group must have a common value for the resource, which can be either the built-in resource host or a site-defined node-level resource.

Note that nodes can have sharing attributes that override job placement requests.
-I: Job is to be run interactively.

As mentioned, the -I option must be specified for the job to be interactive. After opening an interactive session, we may run programs in the normal way. For running serial programs, you should only ask for one "chunk". Parallel programs can be run with the preferred number of nodes, which should be specified with -l select=<# nodes> when qsub is started.

Note that ncpus can not be larger than the number of processors on each node on the machine in question.

To open a display when running interactively, start the interactive session with the following command:

	qsub -I -q <queue> -l select=< (number of nodes)>:ncpus=< (number of processors)> -v DISPLAY

To end an interactive job, just type exit. If you wish to interrupt qsub prior to job start, this can be done by typing control-C. It will then query if the user wishes to exit. If the user responds "yes", qsub exits and the job is aborted.

To see which nodes your job is using:

	cat $PBS_NODEFILE

It is strongly suggested that you only use an interactive session for developmental tasks (such as debugging). Use a PBS job submission file when running the finished program.

PBS Examples

A large part of submitting a job involves understanding how to request computing resources. This section contains examples of submitting PBS jobs, both using a batch script and interactively. There will be separate examples for MPI and OpenMP jobs. Note that the sections 'batch' and 'interactive' have some examples which might also be relevant for, say, MPI and OpenMP.

Serial PBS Example

There are two ways to run a serial program under PBS: batch and interactively. For long jobs, batch submission is to be preferred. There is no difference in how you run a Fortran program, a C program or a C++ program, when they have been compiled.

Batch submission

Suppose that we want to run the C program 'hello.c' - where the executable is called 'hello'. Make a script and call it something meaningful, like run_hello. The script should then contain the following:

	#!/bin/bash
	cd $PBS_O_WORKDIR
	./hello

Since PBS will always start in your home directory, you should either do a cd $PBS_O_WORKDIR (which returns you to the directory you submitted the script from) or give the full path to the program.

The command to submit the job is the following:

	qsub -q standby -l select=1,walltime=1:00 run_hello

Where I am using the queue 'standby' on Steele, 1 node and a walltime of one min. My job submission file is called run_hello. It should be noted that if you want to use the default queue, you do not need to explicitly ask for it.

Submitting this script gives the following result. It will take a while before the job completes:

	$ qsub -q standby -l select=1,walltime=1:00 run_hello
	91.steele-adm.rcac.purdue.edu
	$

Doing a 'ls' in your directory will now show two new files:

	$ ls
	hello                            run_hello
	hello.c                          run_hello.e91
	hello.out                        run_hello.o91
	$

If everything went well, then the file 'run_hello.e91' will be empty, since it contains any error-messages your program gave while running. The file 'run_hello.o91' contains the output from your program. In this case the output is:

	$ cat run_hello.o91
	Hello World!
	$

Interactively

To use the PBS queue interactively, you must first give a command like the one below. Remember to type 'cd $PBS_O_WORKDIR' (or the path to the working directory), since you will have been returned to your home directory upon start of the interactive job.

$ qsub -I -q standby -l select=1
qsub: waiting for job 189639.steele-adm.rcac.purdue.edu to start
qsub: job 189639.steele-adm.rcac.purdue.edu ready

$

Where we are running in the queue 'standby' on Steele, and asking for 1 node.

We can now run the job the same way a serial job is normally run. Remember, interactive sessions are mostly for testing purposes, and longer jobs should always be submitted using a job submission file.

	$ ./hello
	Hello World!
	$

MPI PBS Example

The path to MPICH2 or MPICH is probably not setup on your account. Try typing mpicc at the prompt. If you get 'mpicc: command not found', then you need to either use the 'module load' command or add the path to the compilers you use to your setup file (.cshrc, .tcsh, .bash, .login, or .profile). The compilers and MPICH2/MPICH can be found in /opt. They have to be loaded to be able to run a MPI program.

Using 'module load': The easiest (and preferred) way to access the compilers with MPICH2 included (and the ordinary compilers) is to use the module load mpich2-<compiler> (and module load <compiler> for the ordinary compilers), where <compiler> is one of: intel, gcc, or pgi. Use module avail to see all the possibilities.

Example (loading Intel compilers with MPICH2 included):

	$ module load mpich2-intel
	$

To learn how to compile a MPI program, see the section Compiling a MPI program.

Running your program

To run your programs, you will need to use the PBS queue. This can either be done interactively or by submitting a job using a job submission file.

There is no difference in how you run a Fortran program, a C program or a C++ program, when they have been compiled.

Submitting the job to PBS using a script

Let us say that we want to run the C program example 'hello.c'. Make a script and call it something meaningful, like run_hello. The script should then contain something like the following:

	#!/bin/bash
	source /etc/profile
	module load mpich2-intel
	mpirun -np 4 -machinefile $PBS_NODEFILE mpi/hello

It is necessary to include the 'source /etc/profile' under bash/ksh, to be able to use the 'module load mpich2-<compiler>'. The <compiler> can be either intel, gcc, or pgi. Last, the program is run with either mpiexec or mpirun. Since PBS always goes to your home directory, you should give the full path to the program - here mpi/hello, or add cd $PBS_O_WORKDIR before running the program (puts you in the directory you were standing in when issuing the qsub command.

To submit a job:

	qsub -q workq -l select=4,walltime=1:00 run_program

Where the options used mean the following:

-q <name>: tells which queue you want the job to run in (here I have chosen workq). A list of available queues can be seen using the command qstat -Q.
-l select: tells the job how many nodes you want to use (4 in the example) and
walltime=hh:mm:ss defines how much wall clock time it has (in the example it is set to 1 minute).

Submitting this script now gives the following result (it will take a while before the job is completed):

	user123@radon-fe00:~/mpi$ qsub -q workq -l select=4,walltime=1:00 run_hello 
	119452.radon-adm.rcac.purdue.edu
	user123@radon-fe00:~/mpi$

Doing a 'ls' in your directory will now show two new files:

	user123@radon-fe00:~/mpi$ ls
	hello                            run_hello
	hello.c                          run_hello.e119452
	hello.out                        run_hello.o119452
	user123@radon-fe00:~/mpi$

If everything went well, then the file 'run_hello.e119452' will be empty, since it contains any error-messages your program gave while running. The file 'run_hello.o119452' contains the output from your program. In this case the output is:

	user123@radon-fe00:~/mpi$ less run_hello.o119452 
	Processor 2 of 4: Hello World!
	Processor 3 of 4: Hello World!
	Processor 1 of 4: Hello World!
	Processor 0 of 4: Hello World!
	user123@radon-fe00:~/mpi$

Mpiexec is a replacement program for the script mpirun, which is part of the MPICH2 (and MPICH) package. It is used to initialize a parallel job from within a PBS batch or interactive environment. Mpiexec uses the task manager library of PBS to spawn copies of the executable on the nodes in a PBS allocation. There are reasons to use mpiexec rather than a script (mpirun) or an external daemon (mpd):

Running interactively

Example (run on the queue 'workq' on Radon):

	user123@radon-fe00:~/mpi$ qsub -I -q workq -l select=2:ncpus=2,walltime=8:00
	qsub: waiting for job 119450.radon-adm.rcac.purdue.edu to start
	qsub: job 119450.radon-adm.rcac.purdue.edu ready
	
	user123@radon-b002:~$

Where the options used means the following:

-I means that we want to run the job interactively.
-q <name>: tells which queue you want the job to run in (here I run on workq). A list of available queues can be seen using the command qstat -Q.
-l select: tells the job how many nodes you want to use (2 in the example) and
ncpus is how many processors you want to use on each node. Here we want 2 - remember to not ask for more than is available on the given machine, and
walltime=hh:mm:ss defines how much wall clock time it has (in the example it is set to 8 minutes).

You can also just start up an interactive job without time constraints:

	qsub -I -q workq -l select=4

(Where the options used mean we ask for 4 nodes and 1 processor on each node). To end the job, you then type: exit.

Note that running an interactive job without time constraints means that you will keep the nodes allocated for the default time limit for that queue. If this is shorter than the time you need, your job will not finish. If, on the other hand, it is longer than what you need, you are keeping those nodes from other people's usage. Therefore, use this with caution.

Running the above PBS command gives:

	user123@radon-fe00:~/mpi$ qsub -I -q workq -l select=4                    
	qsub: waiting for job 119451.radon-adm.rcac.purdue.edu to start
	qsub: job 119451.radon-adm.rcac.purdue.edu ready
	
	user123@radon-b002:~$

We then need to change to the directory where our program is located. To run a program we use mpirun or mpiexec. You can not just start the program with ./program, since it will then just use one task process.

mpirun: To run a program with mpirun, you issue the following command (remember to do 'module load mpich2-<compiler>', source your file which contains the path, or simply give the full path to mpirun, if you haven't added either to your setup). We need to add '-machinefile $PBS_NODEFILE' for running with mpirun on Radon - this is not the case if we use mpiexec:

	mpirun -np <number of tasks> -machinefile $PBS_NODEFILE program

Running this on our program 'hello', for 4 tasks results in:

	user123@radon-b002:~/mpi$ mpirun -np 4 -machinefile $PBS_NODEFILE hello
	Processor 2 of 4: Hello World!
	Processor 1 of 4: Hello World!
	Processor 3 of 4: Hello World!
	Processor 0 of 4: Hello World!
	user123@radon-b002:~/mpi$

Note that the order of the processors is random. This can not be controlled in a parallel program.

mpiexec: To run with mpiexec, use the following command (only give number of tasks if you do not want to use all you asked for when entering the queue):

	mpiexec -n <number of tasks> program

It is not necessary to give the -n <number of tasks>, unless you wish to use a different amount than what you asked for when the job was originally started.

	user123@radon-b002:~/mpi$ mpiexec hello
	Processor 2 of 4: Hello World!
	Processor 0 of 4: Hello World!
	Processor 3 of 4: Hello World!
	Processor 1 of 4: Hello World!
	user123@radon-b002:~/mpi$ mpiexec -n 2 hello
	Processor 1 of 2: Hello World!
	Processor 0 of 2: Hello World!
	user123@radon-b002:~/mpi$

To see which nodes you are using:

	cat $PBS_NODEFILE

Notes

Starting tasks with the TM interface is much faster than invoking a separate rsh or ssh once for each process.
Resources used by the spawned processes are accounted correctly with mpiexec, and reported in the PBS logs, because all the processes of a parallel job remain under the control of PBS, unlike when using startup scripts such as mpirun.
Tasks that exceed their assigned limits of CPU time, wallclock time, memory usage, or disk space are killed cleanly by PBS. It is quite hard for processes to escape control of the resource manager when using mpiexec.
You can use mpiexec to enforce a security policy. If all jobs are required to startup using mpiexec and the PBS execution environment, it is not necessary to enable rsh or ssh access to the compute nodes in the cluster.
PBS Queues. Always use qstat -Q to determine which queues are available. The queueus which usually are available to everyone are generally called standby, and workq.

When using MPI_File_Write, the output to the file will be written in binary. To view it, use: od -d 'output.file'.

MPI_File_Write will not automatically delete the old contents of an output file, so you may have to remove it before writing new data to it, unless you want to keep the old data.

You can see which nodes you are using on one of the cluster machines with the command: cat $PBS_NODEFILE
Running a program on the Linux clusters with ./program is a bad idea, since it will just use one task process. So unless that is what you want (which is rarely the case for a MPI-program), you should use mpirun or mpiexec.
The order of the processors is random. There is no way to tell which processor will do what and in which order in a parallel program.

Common mistakes

For C programs, remember to write MPI_Send and MPI_Recv and NOT MPI_SEND and MPI_RECV. The compiler will not complain, but during the execution of the program you will get a very confusing error:

Invalid datatype (0) in MPI_Send, task 0...
In C programs, when sending arrays or values with MPI_Send/MPI_Scatter, you need to remember the & in front of the send/recieve buffer name. Example:
```
int array[5000];
int subarray[5000/4];
.
.
.
MPI_Scatter(&array,sendcount,MPI_INT,&subarray,recvcount,MPI_INT,\
0,MPI_COMM_WORLD);
```

In C programs, when using dynamically allocated arrays, they are really pointers. Therefore, when using the MPI_Send/MPI_Scatter, you must NOT put a & in front of the send/receive buffer name. Example:

int *array;
int *subarray;
.
.
.
array = (int *)malloc(array_size*sizeof(int));
subarray = (int *)malloc(subarray_size*sizeof(int));
.
.
.
MPI_Scatter(array,sendcount,MPI_INT,subarray,recvcount,MPI_INT,0,\
MPI_COMM_WORLD);

Remember that values must either be globally declared or sent/broadcast/scattered to every processor. Otherwise the other processor will not be able to see their values.

Also, remember that memory allocation must be done either globally or on each processor.

C/MPI programming examples

Most of the programs below are my answers to the exercises in the online "Introduction to MPI" course at NCSA.

hello.c: Each processor prints "Hello World" to the screen.

ping.c: Sending a message between processors 0 and 1. The program reports the initial values of inmsg and outmsg before sending and then after. Try running this program with 2, 3, and 4 processors and see that their order is random and that only processors 0 and 1 exchanges a value.

ArrayFindSerial.c: This is a serial program to find a target value in an array and report back the indices where that target value was found. The target value and the array is read from a file (data.dat below) and printed to an output file 'out.dat'. This serial program should just be run as ./program. It is included for comparison with the following parallel versions of the same program.

ArrayFindParallelChapter4.c: This is the first parallel version of the above serial program. I use the blocking MPI_Send and MPI_Recv. The Master processor reads the array in from the file. Then it divides it up in three equal parts and sends one to each slave processor. These (3) slave processors search through their subarray and finds the indices where the targetvalue are found. These indices are then converted to global indices and sent to the master processor. The master processor then assembles these and write them to a file. The program is meant to run on 4 task processes. I use a static size array, which is initialized in the beginning of the program.

ArrayFindParallelChapter5.c: This version still uses the blocking MPI_Send and MPI_Recv, but since I am now sending back both the index where the target value is found and the average of the index and the target value, I need to use a derived data type. This is first created and then used. The program runs on 4 task processes.

ArrayFindParallelChapter6.c: As in the last version, I create a derived data type, but I now use MPI_Bcast, MPI_Scatter, MPI_Barrier and MPI_Irecv. The program runs on 4 task processes.

ArrayFindParallelChapter7.c: I still use the derived data type and MPI_Bcast and MPI_Scatter, but I know also find the "neighbour processors" and have them send the first value of their subarray to each neighbour processor. The program uses 4 task processes.

ArrayFindParallelChapter8.c: Still using MPI_Bcast and MPI_Scatter. To accomplish finding the neighbour processor I this time create a topological ring, which is a much easier way of finding them.

ArrayFindParallelChapter9.c: Back to using the blocking MPI_Send and MPI_Recv again. This makes it easier to accomplish the goal of this program, which is to allow each of the 3 slave processors to write to the output file. The master processor (0) still reads from the input file.

ArrayFindParallelChapter12.c: I again use MPI_Bcast and MPI_Scatter. The arrays are no longer static and I allocate them dynamically when I know the length of the arrays. This value is read from the input file (b.dat).

data.dat: A file with the target value and array used by the 'ArrayFind' programs. The first value is the target value. All values are integers and separated by space.

b.dat: The input file used by the program 'ArrayFindParallelChapter12.c'. The value on the first line is the the number of elements in the array, the next line holds the target value to be found, and the following lines are the elements in the array. The values are all integers and they are separated by space.

Diagnostic Error messages from MPI

Click here and go to chapter 5 (p. 121) to see what the diagnostic error messages from MPI means.

Extra examples of MPI programs

To see a few other examples of running MPI programs go here.

Running Jobs via Condor

Condor allows users to run jobs on systems which would otherwise be idle for however long as those systems are not needed by their primary users. Condor is one of several distributed computing systems RCAC makes available. Most RCAC resources, in addition to being available through normal means, are a part of BoilerGrid and can be used via Condor. If a primary user needs a machine, the Condor job is immediately either checkpointed and/or migrated and the resource made available. Thus, shorter jobs will have a better completion rate via Condor than longer jobs; however, even though jobs may have to be restarted elsewhere, BoilerGrid can offer a vast amount of computational resources to users. Not only are nearly all RCAC systems part of BoilerGrid, so also are large numbers of lab machines at the West Lafayette and other Purdue campuses. BoilerGrid is one of the largest Condor pools in the world. Some machines at other institutions are also a part of a larger Condor federation known as DiaGrid and can be used as well. For more information, refer to the BoilerGrid documentation.

Radon Condor Tips

Code Preparation To get a job to run under Condor, it must be able to run as a background batch job. Since Condor runs the program unattended and in the background, it will not be able to do interactive input and output. Condor can redirect console output (stdout and stderr) and keyboard input (stdin) to and from files for you. Create any needed files that contain the proper keystrokes needed for program input. Make certain the program will run correctly with the files.
The Condor Universe Condor has more than one runtime environment (called a universe) from which to choose. The most used ones are:
- the standard universe, which allows a job running under Condor to handle system calls by returning them to the machine where the job was submitted and it also provides the mechanisms necessary to take a checkpoint and migrate a partially completed job, should the machine on which the job is executing become unavailable. To use the standard universe, it is necessary to relink the program with the Condor library using the condor_compile command. To read more about compiling for Condor, look at man condor_compile or in the longer manual.
- the vanilla universe, which provides a way to run jobs that cannot be relinked. There is no way to take a checkpoint or migrate a job executed under the vanilla universe. For access to input and output files, jobs must either use a shared file system, or use Condor's File Transfer mechanism.
Choose a universe under which to run the Condor program, and re-link the program if necessary.
Submit description file To control the details of a job submission, you use a submit description file. The file contains information about the job such as what executable to run, the files to use for keyboard and screen data, the platform type required to run the program, which universe use wish to use, and where to send e-mail when the job completes. You can also tell Condor how many times to run a program; it is simple to run the same program multiple times with multiple data sets. The sumbit description file is where the requirements and rank commands are defined.
Write a submit description file to go with the job. Look at this example for guidance.
Submit the Job Submit the program to Condor with the condor_submit command.

condor_q

condor_status

condor_prio

condor_rm

It is difficult to be allowed to run many (like 200+) jobs in the standard Condor universe, because of how the Purdue pool is.
Don't queue up thousands and thousands of jobs in a queue. Use DAGman to divide your jobs into reasonably-sized chunks. (500 jobs or so)
When a submit node is heavily used, don't run condor_q constantly. The condor_schedd is single-threaded, and schedules work in the same thread that you're using to list the queue.
Long jobs should run in the standard universe, not in the vanilla universe, since they will otherwise never finish.
Standard Universe is the most desirable due to checkpoint availability, but no possibility of sub-processes. Scripts can be used as executables. It is also necessary to link with Condor run-time library (use of Intel compilers is not possible). Only static links works. Good for longer jobs because of the checkpoint availability.
Vanilla Universe is the only possibility for Windows machines. It only has preemption by suspension or eviction and is thus bad for long jobs, but OK for short jobs (eviction is when the owner of the cluster bumps your job. It will then restart.) Can use Intel compilers (may run 30%-40% faster). Thus it may even be faster for somewhat longer jobs, because the speed gain may be bigger than the advantage from the checkpoint availability.
Generally, if the execution of your job runs less than 1/2 hour, then there is almost no eviction. If it is shorter than 1 hour, there will still only be a few evictions.
Purdue have both a scavenging/preempting and a scheduling system. Remember that the Condor pool is very heterogeneous, both regarding processor versions and OS versions/types (both Linux of different varieties and some Windows.)
It is a good idea to use static links regardless of universe, since you will never know which version of SUN libraries etc. you will find on a given machine.
Why no middleware (like Mycluster at TACC)? Middleware can be easier for the user, since it does Condor (and other stuff) 'behind the scene'. Middlewares are schedulers also and will don't start the job until it is guaranteed to run to completion (no eviction). However, it has a lot of job restarts and thus much overhead on many jobs. Therefore, for a large number of jobs the 'pure' Condor is better.

Limitations on Jobs which can be Checkpointed

Multi-process jobs are not allowed. This includes system calls such as fork(), exec(), and system().
Interprocess communication is not allowed. This includes pipes, semaphores, and shared memory.
Network communication must be brief. A job may make network connections using system calls such as socket(), but a network connection left open for long periods will delay checkpointing and migration.
Sending or receiving the SIGUSR2 or SIGTSTP signals is not allowed. Condor reserves these signals for its own use. Sending or receiving all other signals is allowed.
Alarms, timers, and sleeping are not allowed. This includes system calls such as alarm(), getitimer(), and sleep().
Multiple kernel-level threads are not allowed. However, multiple user-level threads are allowed.
Memory mapped files are not allowed. This includes system calls such as mmap() and munmap().
File locks are allowed, but not retained between checkpoints.
All files must be opened read-only or write-only. A file opened for both reading and writing will cause trouble if a job must be rolled back to an old checkpoint image. For compatibility reasons, a file opened for both reading and writing will result in a warning but not an error.
A fair amount of disk space must be available on the submitting machine for storing a job's checkpoint images. A checkpoint image is approximately equal to the virtual memory consumed by a job while it runs. If disk space is short, a special checkpoint server can be designated for storing all the checkpoint images for a pool.
On Digital Unix (OSF/1), HP-UX, and Linux, your job must be statically linked. Dynamic linking is allowed on all other platforms.

Note

Choosing a Condor Universe

Standard: The standard universe provides migration and reliability, but has some restrictions on the programs that can be run.
Vanilla: The vanilla universe provides fewer services, but has very few restrictions.
PVM: The PVM universe is for programs written to the Parallel Virtual Machine interface. See section 2.9 of the Condor manual for more about PVM and Condor.
MPI: The MPI universe is for programs written to the MPICH interface. See section 2.10 of the Condor manual for more about MPI and Condor. The MPI Universe has be superseded by the Parallel universe.
Globus or Grid: The Globus or Grid universe allows users to submit jobs using Condor's interface. These jobs are submitted for execution on grid resources. For Globus jobs, see http://www.globus.org for more information.
Java: The Java universe allows users to run jobs written for the Java Virtual Machine (JVM).
Scheduler: The scheduler universe allows users to submit lightweight jobs to be spawned by the condor_schedd on the submit host itself.
Local: The local universe allows a Condor job to be submitted and executed with different assumptions for the execution conditions of the job.
Parallel: The Parallel universe is for programs that require multiple machines for one job. See section 2.10 for more about the Parallel universe.

chapter 2.4.1 of the Condor manual

Radon Condor Submission Script

Example 1

Output = hello.out

  ####################                                                    
  # 
  # Example 1                                                            
  # Simple condor job description file                                    
  #                                                                       
  ####################                                                    
                                                                          
  Executable     = hello
  Log            = hello.log
  Queue

Example 2

Condor manual

  ####################     
  #                       
  # Example 2: demonstrate use of multiple     
  # directories for data organization.      
  #                                        
  ####################                    
                                         
  Executable     = mathematica          
  Universe = vanilla                   
  input   = test.data                
  output  = loop.out                
  error   = loop.error             
  Log     = loop.log                                                    
                                  
  Initialdir     = run_1         
  Queue                         
                               
  Initialdir     = run_2      
  Queue

Example 3

Condor manual

  ####################                    
  #
  # Example 3: Show off some fancy features including
  # use of pre-defined macros and logging.
  #
  ####################                                                    

  Executable     = foo                                                    
  Requirements   = Memory >= 32 && OpSys == "IRIX65" && Arch =="SGI"     
  Rank		 = Memory >= 64
  Image_Size     = 28 Meg                                                 

  Error   = err.$(Process)                                                
  Input   = in.$(Process)                                                 
  Output  = out.$(Process)                                                
  Log = foo.log

  Queue 150

Condor Job Submission

condor_submit

	condor_submit file

Example:

	condor_submit run_hello (my submit description file is called run_hello).

condor_submit

Condor Job Status

condor_status

condor_all

tg_login

/opt/condor/bin/condor_all

condor_status

here

condor_status

-available shows only machines which are willing to run jobs now.
-run shows only machines which are currently running jobs.
-l lists the machine ClassAds for all machines in the pool.

man condor_status

Condor Manual

condor_status

here

condor_status -l

here

Affecting the jobs execution

Placing a job on hold

condor_hold

condor_release

manual page

Changing the priority of jobs

condor_prio

	user123@radon:~$ condor_q user123
	
	
	-- Submitter: radon.rcac.purdue.edu : <128.210.9.35:35407> : radon.rcac.purdue.edu
	 ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD               
	260187.0   user123         8/30 13:59   0+00:00:00 I  0   19.5 hello             
	
	1 jobs; 1 idle, 0 running, 0 held
	user123@radon:~$ condor_prio -p -15 260187.0
	user123@radon:~$ condor_q user123
	
	-- Submitter: radon.rcac.purdue.edu : <128.210.9.35:35407> : radon.rcac.purdue.edu
	 ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD               
	260187.0   user123         8/30 13:59   0+00:00:03 R  -15 19.5 hello             
	
	1 jobs; 0 idle, 1 running, 0 held
	user123@radon:~$

Note

6. Managing a Job

man <command>

chapter 9

condor_status -submitters

	-bash-3.00$ condor_status -submitters
	
	Name                 Machine      Running IdleJobs HeldJobs
	
	user4@rcac.purdue. hamlet.rca         0        0        5
	user5@r            hamlet.rca       792        0        0
	user6@rcac.purdue. hamlet.rca         0        0        0
	user7@rcac.purdue. hamlet.rca         0        0        1
	user8@rcac.purdue.edu hamlet.rca         0        0      228
	user5@r            lear.rcac.         0        0        1
	user9@rcac.purdue.e radon.rcac         0        6        2
	user10@rcac.purdue radon.rcac         0        0        1
	user7@rcac.purdue radon.rcac         0        1        2
	user6@rcac.purdue.e radon.rcac         0        0        2
	user5@r             radon.rcac       882        0        0
	user11@rcac.purdue  radon.rcac         0        0        1
	user12@rcac.purdue.e radon.rcac         0        0        5
	user13@rcac.purdue.  radon.rcac         0      186        2
	user14@rcac.purdue   radon.rcac      1000        0        0
	user15@rcac.p        steele-fe0         0      220        1
	user16@rcac.purdue.ed steele-fe0         0        0        1
	user17@rcac.purdue steele-fe0         0    37472        1
	tg_user1@rcac.     tg-gatekee         0        1        0
	tg_user2@rcac.purdue.e tg-login64         0        1        0
	
	                           RunningJobs           IdleJobs           HeldJobs
	
	user9@rcac.purdue.e                 0                  6                  2
	user7@rcac.purdue                   0                  0                  1
	user15@rcac.p                        0                220                  1
	tg_user1@rcac.                       0                  1                  0
	user12@rcac.purdue.                  0                  0                  5
	user7@rcac.purdue                    0                  1                  2
	user6@rcac.purdue.e                  0                  0                  2
	user11@rcac.purdue.ed                0                  0                  1
	user5@r                           1674                  0                  1
	user17@rcac.purdue                   0              37472                  1
	tg_user1@rcac.purdue.e               0                  1                  0
	user6@rcac.purdue.                   0                  0                  0
	user16@rcac.purdue                   0                  0                  1
	user12@rcac.purdue.e                 0                  0                  5
	user13@rcac.purdue.                  0                186                  2
	user14@rcac.purdue                1000                  0                  0
	user10@rcac.purdue.                  0                  0                  1
	user8@rcac.purdue.edu                0                  0                228
	
	               Total              2674              37887                253
	-bash-3.00$

Checking on the progress of jobs

condor_q

Log = <name>.log

condor_shadow

condor_status

condor_status -constraint 'RemoteUser == "user123@rcac.purdue.edu"'

	user123@radon:~$ condor_status -constraint 'RemoteUser == "user123@rcac.purdue.edu"'
	
	Name          OpSys       Arch   State      Activity   LoadAv Mem   ActvtyTime
	
	ba-005.rcac.p LINUX       INTEL  Claimed    Busy       1.000   502  0+00:24:44
	ba-006.rcac.p LINUX       INTEL  Claimed    Busy       0.990   502  0+00:20:22
	ba-007.rcac.p LINUX       INTEL  Claimed    Busy       1.000   502  0+00:23:16
	ba-008.rcac.p LINUX       INTEL  Claimed    Busy       1.000   502  0+00:30:20
	...

condor_status -run

Condor Job Cancellation

Removing a job from the queue

condor_rm

	user123@radon:~$ 
	
	Submitter: radon.rcac.purdue.edu : <128.210.9.35:35407> : radon.rcac.purdue.edu
	 ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD               
	...
	260076.7   nice-user.user1 8/18 00:05   0+00:21:47 I  0   29.3 startfah.sh -oneun
	260076.9   nice-user.user1 8/18 00:05   0+01:40:44 I  0   136.7 startfah.sh -oneun
	260185.0   user123              8/30 13:01   0+00:00:00 R  0   19.5 hello             
	...

	user123@radon:~$ condor_rm 260185.0
	Job 260185.0 marked for removal
	user123@radon:~$ condor_q
	
	Submitter: radon.rcac.purdue.edu : <128.210.9.35:35407> : radon.rcac.purdue.edu
	 ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD               
	...
	260076.7   nice-user.user1 8/18 00:05   0+00:21:47 I  0   29.3 startfah.sh -oneun
	260076.9   nice-user.user1 8/18 00:05   0+01:40:44 I  0   136.7 startfah.sh -oneun
	...

Condor Examples

Condor manual

Simple Condor Example

Compiling

	condor_compile <compiler> <program>.<extension> -o <program name>

Example:

	condor_compile gcc hello.c -o hello

Submitting

condor_submit <job-name>

job-name

Example:

	Executable   = hello
	Log          = hello.log
	Output       = hello.out
	Queue

hello

condor_compile gcc hello.c -o hello

	user123@radon:~$ condor_submit run_hello 
	Submitting job(s).
	Logging submit event(s).
	1 job(s) submitted to cluster 260182.
	user123@radon:~$

condor_status

hello.log

hello.out

always

hello.log

	000 (260182.000.000) 08/29 16:21:31 Job submitted from host: <128.210.9.35:35407>
	...
	001 (260182.000.000) 08/29 16:22:42 Job executing on host: <128.211.131.51:32780>
	...
	005 (260182.000.000) 08/29 16:22:42 Job terminated.
	        (1) Normal termination (return value 13)
	                Usr 0 00:00:00, Sys 0 00:00:00  -  Run Remote Usage
	                Usr 0 00:00:00, Sys 0 00:00:00  -  Run Local Usage
	                Usr 0 00:00:00, Sys 0 00:00:00  -  Total Remote Usage
	                Usr 0 00:00:00, Sys 0 00:00:00  -  Total Local Usage
	        830  -  Run Bytes Sent By Job
	        13490672  -  Run Bytes Received By Job
	        830  -  Total Bytes Sent By Job
	        13490672  -  Total Bytes Received By Job
	...

hello.out

	Hello World!

Radon Info