A job is submitted for execution to Condor using the condor_ submit command. condor_ submit takes as an argument the name of a file called a submit description file. This file contains commands and keywords to direct the queuing of jobs. In the submit description file, Condor finds everything it needs to know about the job. Items such as the name of the executable to run, the initial working directory, and command-line arguments to the program all go into the submit description file. condor_ submit creates a job ClassAd based upon the information, and Condor works toward running the job.
The contents of a submit file can save time for Condor users. It is easy to submit multiple runs of a program to Condor. To run the same program 500 times on 500 different input data sets, arrange your data files accordingly so that each run reads its own input, and each run writes its own output. Each individual run may have its own initial working directory, stdin, stdout, stderr, command-line arguments, and shell environment. A program that directly opens its own files will read the file names to use either from stdin or from the command line. A program that opens a static filename every time will need to use a separate subdirectory for the output of each run.
The condor_ submit manual page
is on page and
contains a complete and full description of how to use condor_ submit.
In addition to the examples of submit description files given in the condor_ submit manual page, here are a few more.
Example 1 is the simplest submit description file possible. It queues up one copy of the program foo(which had been created by condor_ compile) for execution by Condor. Since no platform is specified, Condor will use its default, which is to run the job on a machine which has the same architecture and operating system as the machine from which it was submitted. No input, output, and error commands are given in the submit description file, so the files stdin, stdout, and stderr will all refer to /dev/null. The program may produce output by explicitly opening a file and writing to it. A log file, foo.log, will also be produced that contains events the job had during its lifetime inside of Condor. When the job finishes, its exit conditions will be noted in the log file. It is recommended that you always have a log file so you know what happened to your jobs.
#################### # # Example 1 # Simple condor job description file # #################### Executable = foo Log = foo.log Queue
Example 2 queues two copies of the program mathematica. The first copy will run in directory run_1, and the second will run in directory run_2. For both queued copies, stdin will be test.data, stdout will be loop.out, and stderr will be loop.error. There will be two sets of files written, as the files are each written to their own directories. This is a convenient way to organize data if you have a large group of Condor jobs to run. The example file shows program submission of mathematica as a vanilla universe job. This may be necessary if the source and/or object code to program mathematica is not available.
#################### # # 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
The submit description file for Example 3 queues 150 runs of program foo which has been compiled and linked for Sun workstations running Solaris 8. This job requires Condor to run the program on machines which have greater than 32 megabytes of physical memory, and expresses a preference to run the program on machines with more than 64 megabytes, if such machines are available. It also advises Condor that it will use up to 28 megabytes of memory when running. Each of the 150 runs of the program is given its own process number, starting with process number 0. So, files stdin, stdout, and stderr will refer to in.0, out.0, and err.0 for the first run of the program, in.1, out.1, and err.1 for the second run of the program, and so forth. A log file containing entries about when and where Condor runs, checkpoints, and migrates processes for the 150 queued programs will be written into file foo.log.
#################### # # Example 3: Show off some fancy features including # use of pre-defined macros and logging. # #################### Executable = foo Requirements = Memory >= 32 && OpSys == "SOLARIS28" && Arch =="SUN4u" Rank = Memory >= 64 Image_Size = 28 Meg Error = err.$(Process) Input = in.$(Process) Output = out.$(Process) Log = foo.log Queue 150
The requirements and rank commands in the submit description file are powerful and flexible. Using them effectively requires care, and this section presents those details.
Both requirements and rank need to be specified
as valid Condor ClassAd expressions, however, default values are set by the
condor_ submit program if these are not defined in the submit description file.
From the condor_ submit manual page and the above examples, you see
that writing ClassAd expressions is intuitive, especially if you
are familiar with the programming language C. There are some
pretty nifty expressions you can write with ClassAds.
A complete description of ClassAds and their expressions
can be found in section 4.1 on
page .
All of the commands in the submit description file are case insensitive, except for the ClassAd attribute string values. ClassAds attribute names are case insensitive, but ClassAd string values are case preserving.
Note that the comparison operators
(<
, >
, <=
, >=
, and ==
)
compare strings
case insensitively. The special comparison operators
=?=
and =!=
compare strings case sensitively.
The allowed ClassAd attributes are those that appear in a machine or a job ClassAd. To see all of the machine ClassAd attributes for all machines in the Condor pool, run condor_ status -l. The -l argument to condor_ status means to display all the complete machine ClassAds. The job ClassAds, if there jobs in the queue, can be seen with the condor_ q -l command. This will show you all the available attributes you can play with.
To help you out with what these attributes all signify, descriptions follow for the attributes which will be common to every machine ClassAd. Remember that because ClassAds are flexible, the machine ads in your pool may include additional attributes specific to your site's installation and policies.
X
for the VM with
name="vmX@full.hostname"For non-SMP machines with one virtual machine, the value will be 1.
In addition, there are a few attributes that are automatically inserted into the machine ClassAd whenever a resource is in the Claimed state:
There are a few attributes that are only inserted into the machine ClassAd if a job is currently executing. If the resource is claimed but no job are running, none of these attributes will be defined.
152.3
), as seen from condor_ q
on the submitting machine.
Finally, the single attribute, CurrentTime, is defined by the ClassAd environment.
CurrentTime - EnteredCurrentStatuswill equal the number of seconds that the job has been on hold.
Integer Code | Reason for Hold |
1 | The user put the job on hold with condor_ hold. |
2 | Globus middleware reported an error. HoldReasonSubCode is the GRAM error number. |
3 | The PERIODIC_HOLD expression evaluated to True. |
4 | The credentials for the job are invalid. |
5 | A job policy expression evaluated to Undefined. |
6 | The condor_ starter failed to start the executable. HoldReasonSubCode is the Unix error number. |
7 | The standard output file for the job could not be opened. HoldReasonSubCode is the Unix error number. |
8 | The standard input file for the job could not be opened. HoldReasonSubCode is the Unix error number. |
9 | The standard output stream for the job could not be opened. HoldReasonSubCode is the Unix error number. |
10 | The standard input stream for the job could not be opened. HoldReasonSubCode is the Unix error number. |
11 | An internal Condor protocol error was encountered when transferring files. |
12 | The condor_ starter failed to download input files. HoldReasonSubCode is the Unix error number. |
13 | The condor_ starter failed to upload output files. HoldReasonSubCode is the Unix error number. |
14 | The initial working directory of the job cannot be accessed. HoldReasonSubCode is the Unix error number. |
condor_q -hold
.
It can be used to determine if a job should be released or not.
Value | Status |
0 | Unexpanded (the job has never run) |
1 | Idle |
2 | Running |
3 | Removed |
4 | Completed |
5 | Held |
Value | Universe |
1 | standard |
4 | PVM |
5 | vanilla |
7 | scheduler |
8 | MPI |
9 | grid |
10 | java |
RemoteWallClockTime - CumulativeSuspensionTimeNote that this number does not get reset to zero when a job is forced to migrate from one machine to another.
When considering the match between a job and a machine, rank is used to choose a match from among all machines that satisfy the job's requirements and are available to the user, after accounting for the user's priority and the machine's rank of the job. The rank expressions, simple or complex, define a numerical value that expresses preferences.
The job's rank expression evaluates to one of three values. It can be UNDEFINED, ERROR, or a floating point value. If rank evaluates to a floating point value, the best match will be the one with the largest, positive value. If no rank is given in the submit description file, then Condor substitutes a default value of 0.0 when considering machines to match. If the job's rank of a given machine evaluates to UNDEFINED or ERROR, this same value of 0.0 is used. Therefore, the machine is still considered for a match, but has no rank above any other.
A boolean expression evaluates to the numerical value of 1.0 if true, and 0.0 if false.
The following rank expressions provide examples to follow.
For a job that desires the machine with the most available memory:
Rank = memory
For a job that prefers to run on a friend's machine on Saturdays and Sundays:
Rank = ( (clockday == 0) || (clockday == 6) ) && (machine == "friend.cs.wisc.edu")
For a job that prefers to run on one of three specific machines:
Rank = (machine == "friend1.cs.wisc.edu") || (machine == "friend2.cs.wisc.edu") || (machine == "friend3.cs.wisc.edu")
For a job that wants the machine with the best floating point performance (on Linpack benchmarks):
Rank = kflopsThis particular example highlights a difficulty with rank expression evaluation as currently defined. While all machines have floating point processing ability, not all machines will have the kflops attribute defined. For machines where this attribute is not defined, Rank will evaluate to the value UNDEFINED, and Condor will use a default rank of the machine of 0.0. The rank attribute will only rank machines where the attribute is defined. Therefore, the machine with the highest floating point performance may not be the one given the highest rank.
So, it is wise when writing a rank expression to check if the expression's evaluation will lead to the expected resulting ranking of machines. This can be accomplished using the condor_ status command with the -constraint argument. This allows the user to see a list of machines that fit a constraint. To see which machines in the pool have kflops defined, use
condor_status -constraint kflopsAlternatively, to see a list of machines where kflops is not defined, use
condor_status -constraint "kflops=?=undefined"
For a job that prefers specific machines in a specific order:
Rank = ((machine == "friend1.cs.wisc.edu")*3) + ((machine == "friend2.cs.wisc.edu")*2) + (machine == "friend3.cs.wisc.edu")If the machine being ranked is "friend1.cs.wisc.edu", then the expression
(machine == "friend1.cs.wisc.edu")is true, and gives the value 1.0. The expressions
(machine == "friend2.cs.wisc.edu")and
(machine == "friend3.cs.wisc.edu")are false, and give the value 0.0. Therefore, rank evaluates to the value 3.0. In this way, machine "friend1.cs.wisc.edu" is ranked higher than machine "friend2.cs.wisc.edu", machine "friend2.cs.wisc.edu" is ranked higher than machine "friend3.cs.wisc.edu", and all three of these machines are ranked higher than others.
If vanilla, java, parallel (or MPI) universe jobs are submitted without using the File Transfer mechanism, Condor must use a shared file system to access input and output files. In this case, the job must be able to access the data files from any machine on which it could potentially run.
As an example, suppose a job is submitted from blackbird.cs.wisc.edu, and the job requires a particular data file called /u/p/s/psilord/data.txt. If the job were to run on cardinal.cs.wisc.edu, the file /u/p/s/psilord/data.txt must be available through either NFS or AFS for the job to run correctly.
Condor allows users to ensure their jobs have access to the right shared files by using the FileSystemDomain and UidDomain machine ClassAd attributes. These attributes specify which machines have access to the same shared file systems. All machines that mount the same shared directories in the same locations are considered to belong to the same file system domain. Similarly, all machines that share the same user information (in particular, the same UID, which is important for file systems like NFS) are considered part of the same UID domain.
The default configuration for Condor places each machine in its own UID domain and file system domain, using the full hostname of the machine as the name of the domains. So, if a pool does have access to a shared file system, the pool administrator must correctly configure Condor such that all the machines mounting the same files have the same FileSystemDomain configuration. Similarly, all machines that share common user information must be configured to have the same UidDomain configuration.
When a job relies on a shared file system, Condor uses the requirements expression to ensure that the job runs on a machine in the correct UidDomain and FileSystemDomain. In this case, the default requirements expression specifies that the job must run on a machine with the same UidDomain and FileSystemDomain as the machine from which the job is submitted. This default is almost always correct. However, in a pool spanning multiple UidDomains and/or FileSystemDomains, the user may need to specify a different requirements expression to have the job run on the correct machines.
For example, imagine a pool made up of both desktop workstations and a dedicated compute cluster. Most of the pool, including the compute cluster, has access to a shared file system, but some of the desktop machines do not. In this case, the administrators would probably define the FileSystemDomain to be cs.wisc.edu for all the machines that mounted the shared files, and to the full hostname for each machine that did not. An example is jimi.cs.wisc.edu.
In this example, a user wants to submit vanilla universe jobs from her own desktop machine (jimi.cs.wisc.edu) which does not mount the shared file system (and is therefore in its own file system domain, in its own world). But, she wants the jobs to be able to run on more than just her own machine (in particular, the compute cluster), so she puts the program and input files onto the shared file system. When she submits the jobs, she needs to tell Condor to send them to machines that have access to that shared data, so she specifies a different requirements expression than the default:
Requirements = UidDomain == "cs.wisc.edu" && \ FileSystemDomain == "cs.wisc.edu"
WARNING: If there is no shared file system, or the Condor pool administrator does not configure the FileSystemDomain setting correctly (the default is that each machine in a pool is in its own file system and UID domain), a user submits a job that cannot use remote system calls (for example, a vanilla universe job), and the user does not enable Condor's File Transfer mechanism, the job will only run on the machine from which it was submitted.
Condor works well without a shared file system. The Condor file transfer mechanism is utilized by the user when the user submits jobs. Condor will transfer any files needed by a job from the machine where the job was submitted into a temporary working directory on the machine where the job is to be executed. Condor executes the job and transfers output back to the submitting machine. The user specifies which files to transfer, and at what point the output files should be copied back to the submitting machine. This specification is done within the job's submit description file.
The default behavior of the file transfer mechanism varies across the different Condor universes, and it differs between UNIX and Windows machines.
For jobs submitted under the standard universe, the existence of a shared file system is not relevant. Access to files (input and output) is handled through Condor's remote system call mechanism. The executable and checkpoint files are transfered automatically, when needed. Therefore, the user does not need to change the submit description file if there is no shared file system.
For the vanilla, java, MPI, and parallel universes, access to files (including the executable) through a shared file system is presumed as a default on UNIX machines. If there is no shared file system, then Condor's file transfer mechanism must be explicitly enabled. When submitting a job from a Windows machine, Condor presumes the opposite: no access to a shared file system. It instead enables the file transfer mechanism by default. Submission of a job might need to specify which files to transfer, and/or when to transfer the output files back.
For the grid universe, jobs are to be executed on remote machines, so there would never be a shared file system between machines. See section 5.3.2 for more details.
For the PVM universe, file transfer other than the master's executable and files given in input, output, and error commands is not supported. This is not usually an impediment (shared file system or not), since PVM jobs are set up to have the master direct the workers, and I/O from the workers is usually passed back to the master via PVM messages, not files.
For the scheduler universe, Condor is only using the machine from which the job is submitted. Therefore, the existence of a shared file system is not relevant.
To enable the file transfer mechanism, two commands are placed in the job's submit description file: should_transfer_files and when_to_transfer_output. An example is:
should_transfer_files = YES when_to_transfer_output = ON_EXIT
The should_transfer_files command specifies whether Condor should transfer input files from the submit machine to the remote machine where the job executes. It also specifies whether the output files are transferred back to the submit machine. The command takes on one of three possible values:
YES
: Condor always transfers both input and output files.
IF_NEEDED
: Condor transfers files if the job is
matched with (and to be executed on) a machine in a
different FileSystemDomain than the
one the submit machine belongs to.
If the job is matched with a machine in the local FileSystemDomain,
Condor will not transfer files and relies
on a shared file system.
NO
: Condor's file transfer mechanism is disabled.
The when_to_transfer_output command tells Condor when output files are to be transferred back to the submit machine after the job has executed on a remote machine. The command takes on one of two possible values:
ON_EXIT
: Condor transfers output files back to the submit
machine only when the job exits on its own.
ON_EXIT_OR_EVICT
:
Condor will always do the transfer,
whether the job completes on its own, is preempted by another job,
vacates the machine, or is killed.
As the job completes on its own, files are transferred back
to the directory where the job was submitted, as expected.
For the other cases, files are transferred back at eviction time.
These files are placed in
the directory defined by the configuration
variable SPOOL, not the directory from which the
job was submitted.
The transferred files are named using the
ClusterId and ProcId job ClassAd attributes.
The file name takes the form:
cluster<X>.proc<Y>.subproc0where
<X>
is the value of ClusterId, and
<Y>
is the value of ProcId.
As an example, job 735.0 may produce the file
$(SPOOL)/cluster735.proc0.subproc0
This is only useful if partial runs of the job are valuable. An example of valuable partial runs is when the application produces its own checkpoints.
There is no default value for when_to_transfer_output.
If using the file transfer mechanism,
this command must be defined.
If when_to_transfer_output is specified in the submit
description file,
but should_transfer_files is not, Condor assumes a
value of YES
for should_transfer_files.
NOTE: The combination of:
should_transfer_files = IF_NEEDED when_to_transfer_output = ON_EXIT_OR_EVICTwould produce undefined file access semantics. Therefore, this combination is prohibited by condor_ submit.
When submitting from a Unix platform,
the file transfer mechanism is unused by default.
If neither when_to_transfer_output or should_transfer_files
are defined, Condor assumes
should_transfer_files = NO
.
When submitting from a Windows platform, Condor does not provide any way to use a shared file system for jobs. Therefore, if neither when_to_transfer_output or should_transfer_files are defined, the file transfer mechanism is enabled by default with the following values:
should_transfer_files = YES when_to_transfer_output = ON_EXIT
NOTE: Prior to Condor version 6.5.2, different attributes were used to
control when and if files should be transferred.
Previously, a single attribute was used to control both things, and
the IF_NEEDED
value was not supported.
This older attribute is still allowed in newer versions of Condor but
it is now deprecated.
when_to_transfer_output and
should_transfer_files should be used instead.
However, beware that these settings will not work with Condor versions
older than 6.5.2.
If the file transfer mechanism is enabled, Condor will transfer the following files before the job is run on a remote machine.
transfer_input_files = file1,file2
As a default, for jobs other than those submitted to the grid universe, any files that are modified or created by the job in the temporary directory at the remote site are transferred back to the machine from which the job was submitted. Most of the time, this is the best option. To restrict the files that are transferred, specify the exact list of files with transfer_output_files. Delimite these file names with a comma. When this list is defined, and any of the files do not exist as the job exits, Condor considers this an error, and re-runs the job.
WARNING: Do not specify transfer_output_files (for other than grid universe jobs) unless there is a really good reason - it is best to let Condor figure things out by itself based upon what output the job produces.
For grid universe jobs, files to be transferred (other than standard output and standard error) must be specified using transfer_output_files in the submit description file.
The file transfer mechanism specifies file names and/or paths on both the file system of the submit machine and on the file system of the execute machine. Care must be taken to know which machine (submit or execute) is utilizing the file name and/or path.
Files in the transfer_input_files command are specified as they are accessed on the submit machine. The program (as it executes) accesses files as they are found on the execute machine.
There are three ways to specify files and paths for transfer_input_files:
Before executing the program, Condor copies the executable, an input file as specified by the submit command input, along with any input files specified by transfer_input_files. All these files are placed into a temporary directory (on the execute machine) in which the program runs. Therefore, the executing program must access input files without paths. Because all transferred files are placed into a single, flat directory, input files must be uniquely named to avoid collision when transferred. A collision causes the last file in the list to overwrite the earlier one.
If the program creates output files during execution, it must create them within the temporary working directory. Condor transfers back all files within the temporary working directory that have been modified or created. To transfer back only a subset of these files, the submit command transfer_output_files is defined. Transfer of files that exist, but are not within the temporary working directory is not supported. Condor's behavior in this instance is undefined.
It is okay to create files outside the temporary working directory on the file system of the execute machine, (in a directory such as /tmp) if this directory is guaranteed to exist and be accessible on all possible execute machines. However, transferring such a file back after execution completes may not be done.
Here are several examples to illustrate the use of file transfer. The program executable is called my_program, and it uses three command-line arguments as it executes: two input file names and an output file name. The program executable and the submit description file for this job are located in directory /scratch/test.
The directory tree for all these examples:
/scratch/test (directory) my_program.condor (the submit description file) my_program (the executable) files (directory) logs2 (directory) in1 (file) in2 (file) logs (directory)
This simple example explicitly transfers input files. These input files to be transferred are specified relative to the directory where the job is submitted. The single output file, out1, created when the job is executed will be transferred back into the directory /scratch/test, not the files directory.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs/err.$(cluster) Output = logs/out.$(cluster) Log = logs/log.$(cluster) should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = files/in1, files/in2 Arguments = in1 in2 out1 Queue
This second example is identical to Example 1, except that absolute paths to the input files are specified, instead of relative paths to the input files.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs/err.$(cluster) Output = logs/out.$(cluster) Log = logs/log.$(cluster) should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = /scratch/test/files/in1, /scratch/test/files/in2 Arguments = in1 in2 out1 Queue
This third example illustrates the use of the
submit command initialdir, and its effect
on the paths used for the various files.
The expected location of the
executable is not affected by the
initialdir command.
All other files
(specified by input,
output,
transfer_input_files,
as well as files modified or created by the job
and automatically transferred back)
are located relative to the specified initialdir.
Therefore, the output file, out1,
will be placed in the files
directory.
Note that the logs2 directory
exists to make this example work correctly.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs2/err.$(cluster) Output = logs2/out.$(cluster) Log = logs2/log.$(cluster) initialdir = files should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = in1, in2 Arguments = in1 in2 out1 Queue
This example illustrates a job that will fail. The files specified using the transfer_input_files command work correctly (see Example 1). However, relative paths to files in the arguments command cause the executing program to fail. The file system on the submission side may utilize relative paths to files, however those files are placed into a single, flat, temporary directory on the execute machine.
Note that this specification and submission will cause the job to fail and reexecute.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs/err.$(cluster) Output = logs/out.$(cluster) Log = logs/log.$(cluster) should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = files/in1, files/in2 Arguments = files/in1 files/in2 files/out1 Queue
This example fails with the following error:
err: files/out1: No such file or directory.
As with Example 4, this example illustrates a job that will fail. The executing program's use of absolute paths cannot work.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs/err.$(cluster) Output = logs/out.$(cluster) Log = logs/log.$(cluster) should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = /scratch/test/files/in1, /scratch/test/files/in2 Arguments = /scratch/test/files/in1 /scratch/test/files/in2 /scratch/test/files/out1 Queue
The job fails with the following error:
err: /scratch/test/files/out1: No such file or directory.
This example illustrates a failure case where the executing program creates an output file in a directory other than within the single, flat, temporary directory that the program executes within. The file creation may or may not cause an error, depending on the existence and permissions of the directories on the remote file system.
Further incorrect usage is seen during the attempt to transfer the output file back using the transfer_output_files command. The behavior of Condor for this case is undefined.
# file name: my_program.condor # Condor submit description file for my_program Executable = my_program Universe = vanilla Error = logs/err.$(cluster) Output = logs/out.$(cluster) Log = logs/log.$(cluster) should_transfer_files = YES when_to_transfer_output = ON_EXIT transfer_input_files = files/in1, files/in2 transfer_output_files = /tmp/out1 Arguments = in1 in2 /tmp/out1 Queue
The requirements expression for a job must depend
on the should_transfer_files
command.
The job must specify the correct logic to ensure that the job is matched
with a resource that meets the file transfer needs.
If no requirements expression is in the submit description file,
or if the expression specified does not refer to the
attributes listed below, condor_ submit adds an
appropriate clause to the requirements expression for the job.
condor_ submit appends these clauses with a logical AND, &&
,
to ensure that the proper conditions are met.
Here are the default clauses corresponding to the different values of
should_transfer_files
:
should_transfer_files = YES
results in the addition of
the clause (HasFileTransfer)
.
If the job is always going to transfer files, it is required to
match with a machine that has the capability to transfer files.
This is a backward compatibility issue, since all versions
of Condor since version 6.3.3 support file transfer and have
HasFileTransfer defined to TRUE
.
should_transfer_files = NO
results in the addition of
(TARGET.FileSystemDomain == MY.FileSystemDomain)
.
In addition, Condor automatically adds the
FileSystemDomain attribute to the job ad, with whatever
string is defined for the condor_ schedd to which the job is
submitted.
If the job is not using the file transfer mechanism, Condor assumes
it will need a shared file system, and therefore, a machine in the
same FileSystemDomain as the submit machine.
should_transfer_files = IF_NEEDED
results in the addition of
(HasFileTransfer || (TARGET.FileSystemDomain == MY.FileSystemDomain))If Condor will optionally transfer files, it must require that the machine is either capable of transferring files or in the same file system domain.
To ensure that the job is matched to a machine with enough local disk space to hold all the transfered files, Condor automatically adds the DiskUsage job attribute. This attribute includes the total size of the job's executable and all input files to be transferred. Condor then adds an additional clause to the Requirements expression that states that the remote machine must have at least enough available disk space to hold all these files:
&& (Disk >= DiskUsage)
If should_transfer_files = IF_NEEDED
and the job prefers
to run on a machine in the local file system domain
over transferring files,
(but are still willing to allow the job to run remotely and transfer
files), the rank expression works well. Use:
rank = (TARGET.FileSystemDomain == MY.FileSystemDomain)
The rank expression is a floating point number, so if other items are considered in ranking the possible machines this job may run on, add the items:
rank = kflops + (TARGET.FileSystemDomain == MY.FileSystemDomain)
The value of kflops can vary widely among machines, so this rank expression will likely not do as it intends. To place emphasis on the job running in the same file system domain, but still consider kflops among the machines in the file system domain, weight the part of the rank expression that is matching the file system domains. For example:
rank = kflops + (10000 * (TARGET.FileSystemDomain == MY.FileSystemDomain))
The should_transfer_files
and when_to_transfer_output
commands in the submit description file result in two corresponding string
attributes in the job ClassAd: ShouldTransferFiles and
WhenToTransferOutput.
These attributes are only defined when the job is matched with an
execute machine running Condor version 6.5.3 or a more recent version.
So, for backward compatibility, condor_ submit also includes the old
attribute used to control this feature: TransferFiles.
If you examine a job with the -long option to condor_ q, and
you see TransferFiles, that attribute is only there for
backward compatibility, and it is ignored if matched with a machine
running version 6.5.3 or greater.
There were problems with this old attribute, since it
was not flexible enough to handle the new IF_NEEDED
functionality, and it was confusing for users.
Therefore, TransferFiles is deprecated, and we will no longer
document its use.
If your submit file refers to transfer_files
,
consider switching it to use the settings described here.
The environment under which a job executes often contains information that is potentially useful to the job. Condor allows a user to both set and reference environment variables for a job or job cluster.
Within a submit description file, the user may define environment variables for the job's environment by using the environment command. See the condor_ submit manual page at section 9 for more details about this command.
The submittor's entire environment can be copied into the job ClassAd for the job at job submission. The getenv command within the submit description file does this. See the condor_ submit manual page at section 9 for more details about this command.
Commands within the submit description file may reference the
environment variables of the submitter as a job is submitted.
Submit description file commands use $ENV(EnvironmentVariableName)
to reference the value of an environment variable.
Again,
see the condor_ submit manual page at
section 9 for more details about this usage.
Condor sets several additional environment variables for each executing job that may be useful for the job to reference.
If executables are available for the different platforms of machines in the Condor pool, Condor can be allowed the choice of a larger number of machines when allocating a machine for a job. Modifications to the submit description file allow this choice of platforms.
A simplified example is a cross submission. An executable is available for one platform, but the submission is done from a different platform. Given the correct executable, the requirements command in the submit description file specifies the target architecture. For example, an executable compiled for a Sun 4, submitted from an Intel architecture running Linux would add the requirement
requirements = Arch == "SUN4x" && OpSys == "SOLARIS251"Without this requirement, condor_ submit will assume that the program is to be executed on a machine with the same platform as the machine where the job is submitted.
Cross submission works for both standard and vanilla universes. The burden is on the user to both obtain and specify the correct executable for the target architecture. To list the architecture and operating systems of the machines in a pool, run condor_ status.
A more complex example of a heterogeneous submission occurs when a job may be executed on many different architectures to gain full use of a diverse architecture and operating system pool. If the executables are available for the different architectures, then a modification to the submit description file will allow Condor to choose an executable after an available machine is chosen.
A special-purpose Machine Ad substitution macro can be used in the executable, environment, and arguments attributes in the submit description file. The macro has the form
$$(MachineAdAttribute)Note that this macro is ignored in all other submit description attributes. The $$() informs Condor to substitute the requested MachineAdAttribute from the machine where the job will be executed.
An example of the heterogeneous job submission has executables available for three platforms: LINUX Intel, Solaris26 Intel, and Solaris 8 Sun. This example uses povray to render images using a popular free rendering engine.
The substitution macro chooses a specific executable after a platform for running the job is chosen. These executables must therefore be named based on the machine attributes that describe a platform. The executables named
povray.LINUX.INTEL povray.SOLARIS26.INTEL povray.SOLARIS28.SUN4uwill work correctly for the macro
povray.$$(OpSys).$$(Arch)
The executables or links to executables with this name are placed into the initial working directory so that they may be found by Condor. A submit description file that queues three jobs for this example:
#################### # # Example of heterogeneous submission # #################### universe = vanilla Executable = povray.$$(OpSys).$$(Arch) Log = povray.log Output = povray.out.$(Process) Error = povray.err.$(Process) Requirements = (Arch == "INTEL" && OpSys == "LINUX") || \ (Arch == "INTEL" && OpSys =="SOLARIS26") || \ (Arch == "SUN4u" && OpSys == "SOLARIS28") Arguments = +W1024 +H768 +Iimage1.pov Queue Arguments = +W1024 +H768 +Iimage2.pov Queue Arguments = +W1024 +H768 +Iimage3.pov Queue
These jobs are submitted to the vanilla universe to assure that once a job is started on a specific platform, it will finish running on that platform. Switching platforms in the middle of job execution cannot work correctly.
There are two common errors made with the substitution macro. The first is the use of a non-existent MachineAdAttribute. If the specified MachineAdAttribute does not exist in the machine's ClassAd, then Condor will place the job in the machine state of hold until the problem is resolved.
The second common error occurs due to an incomplete job set up. For example, the submit description file given above specifies three available executables. If one is missing, Condor report back that an executable is missing when it happens to match the job with a resource that requires the missing binary.
Jobs submitted to the standard universe may produce checkpoints. A checkpoint can then be used to start up and continue execution of a partially completed job. For a partially completed job, the checkpoint and the job are specific to a platform. If migrated to a different machine, correct execution requires that the platform must remain the same.
In previous versions of Condor, the author of the heterogeneous submission file would need to write extra policy expressions in the requirements expression to force Condor to choose the same type of platform when continuing a checkpointed job. However, since it is needed in the common case, this additional policy is now automatically added to the requirements expression. The additional expression is added provided the user does not use CkptArch in the requirements expression. Condor will remain backward compatible for those users who have explicitly specified CkptRequirements-implying use of CkptArch, in their requirements expression.
The expression added when the attribute CkptArch is not specified will default to
# Added by Condor CkptRequirements = ((CkptArch == Arch) || (CkptArch =?= UNDEFINED)) && \ ((CkptOpSys == OpSys) || (CkptOpSys =?= UNDEFINED)) Requirements = (<user specified policy>) && $(CkptRequirements)
The behavior of the CkptRequirements expressions and its addition to
requirements is as follows.
The CkptRequirements expression guarantees correct operation
in the two possible cases for a job.
In the first case, the job has not produced a checkpoint.
The ClassAd attributes CkptArch and CkptOpSys
will be undefined, and therefore the meta operator (=?=
)
evaluates to true.
In the second case, the job has produced a checkpoint.
The Machine ClassAd is restricted to require further execution
only on a machine of the same platform.
The attributes CkptArch and CkptOpSys
will be defined, ensuring that the platform chosen for further
execution will be the same as the one used just before the
checkpoint.
Note that this restriction of platforms also applies to platforms where the executables are binary compatible.
The complete submit description file for this example:
#################### # # Example of heterogeneous submission # #################### universe = standard Executable = povray.$$(OpSys).$$(Arch) Log = povray.log Output = povray.out.$(Process) Error = povray.err.$(Process) # Condor automatically adds the correct expressions to insure that the # checkpointed jobs will restart on the correct platform types. Requirements = ( (Arch == "INTEL" && OpSys == "LINUX") || \ (Arch == "INTEL" && OpSys =="SOLARIS26") || \ (Arch == "SUN4u" && OpSys == "SOLARIS28") ) Arguments = +W1024 +H768 +Iimage1.pov Queue Arguments = +W1024 +H768 +Iimage2.pov Queue Arguments = +W1024 +H768 +Iimage3.pov Queue