NAME¶

pegasus-kickstart - remote job wrapper

SYNOPSIS¶

pegasus-kickstart [-n tr] [-N dv] [-H] [-R site] [-W | -w dir]
                  [ -L lbl -T iso] [-s p | @fn] [-S p | @fn] [-i fn]
                  [ -o fn] [-e fn] [-X] [-l fn sz] [-F] (-I fn | app [appflags])
pegasus-kickstart -V

DESCRIPTION¶

pegasus-kickstart is a wrapper program which manages and monitors the execution of jobs on remote resources. Sitting in between the remote scheduler and the application process, it is possible for pegasus-kickstart to gather additional information about the process' run-time behavior and resource usage, including the exit status of jobs. This information is important for Pegasus invocation tracking as well as detecting Globus GRAM job failures. pegasus-kickstart allows the optional execution of jobs before and after the main application job that run in chained execution with the main application job. See section SUBJOBS for details about this feature. It also allows stdin, stdout, and stderr to be redirected from/to specific files. All jobs with relative path specifications to the application are part of search relative to the current working directory (yes, this is unsafe), and by prepending each component from the PATH environment variable. The first match is used. Jobs that use absolute pathnames, starting in a slash, are exempt. Using an absolute path to your executable is the safe and recommended option. pegasus-kickstart rewrites the command line of any job (pre, post and main) with variable substitutions from Unix environment variables. See section VARIABLE REWRITING below for details on this feature.

OPTIONS¶

-n tr -N dv -H -R site -L lbl, -T iso -S l=p -s fn -i fn -o fn -e fn -l logfn -w dir -W dir -X -B sz -F -I fn -f -t -q app appflags

RETURN VALUE¶

pegasus-kickstart will return the return value of the main job. In addition, the error code 127 signals that the call to exec failed, and 126 that reconnecting the stdio failed. A job failing with the same exit codes is indistinguishable from pegasus-kickstart failures.

SEE ALSO¶

pegasus-plan(1), condor_submit_dag(1), condor_submit(1), getrusage(3c).

SUBJOBS¶

Subjobs are a new feature and may have a few wrinkles left. In order to allow specific setups and assertion checks for compute nodes, pegasus-kickstart allows the optional execution of a prejob. This prejob is anything that the remote compute node is capable of executing. For modern Unix systems, this includes #! scripts interpreter invocations, as long as the x bits on the executed file are set. The main job is run if and only if the prejob returned regularly with an exit code of zero. With similar restrictions, the optional execution of a postjob is chained to the success of the main job. The postjob will be run, if the main job terminated normally with an exit code of zero. In addition, a user may specify a setup and a cleanup job. The setup job sets up the remote execution environment. The cleanup job may tear down and clean-up after any job ran. Failure to run the setup job has no impact on subsequent jobs. The cleanup is a job that will even be attempted to run for all failed jobs. No job information is passed. If you need to invoke multiple setup or clean-up jobs, bundle them into a script, and invoke the clean-up script. Failure of the clean-up job is not meant to affect the progress of the remote workflow (DAGMan). This may change in the future. The setup-, pre-, and post- and cleanup-job run on the same compute node as the main job to execute. However, since they run in separate processes as children of pegasus-kickstart, they are unable to influence each others nor the main jobs environment settings. All jobs and their arguments are subject to variable substitutions as explained in the next section. To specify the prejob, insert the the application invocation and any optional commandline argument into the environment variable GRIDSTART_PREJOB. If you are invoking from a shell, you might want to use single quotes to protect against the shell. If you are invoking from Globus, you can append the RSL string feature. From Condor, you can use Condor’s notion of environment settings. In Pegasus use the profile command to set generic scripts that will work on multiple sites, or the transformation catalog to set environment variables in a pool-specific fashion. Please remember that the execution of the main job is chained to the success of the prejob. To set up the postjob, use the environment variable GRIDSTART_POSTJOB to point to an application with potential arguments to execute. The same restrictions as for the prejob apply. Please note that the execution of the post job is chained to the main job. To provide the independent setup job, use the environment variable GRIDSTART_SETUP. The exit code of the setup job has no influence on the remaining chain of jobs. To provide an independent cleanup job, use the environment variable GRIDSTART_CLEANUP to point to an application with possible arguments to execute. The same restrictions as for prejob and postjob apply. The cleanup is run regardless of the exit status of any other jobs.

VARIABLE REWRITING¶

Variable substitution is a new feature and may have a few wrinkles left. The variable substitution employs simple rules from the Bourne shell syntax. Simple quoting rules for backslashed characters, double quotes and single quotes are obeyed. Thus, in order to pass a dollar sign to as argument to your job, it must be escaped with a backslash from the variable rewriting. For pre- and postjobs, double quotes allow the preservation of whitespace and the insertion of special characters like \a (alarm), \b (backspace), \n (newline), \r (carriage return), \t (horizontal tab), and \v (vertical tab). Octal modes are not allowed. Variables are still substituted in double quotes. Single quotes inside double quotes have no special meaning. Inside single quotes, no variables are expanded. The backslash only escapes a single quote or backslash. Backticks are not supported. Variables are only substituted once. You cannot have variables in variables. If you need this feature, please request it. Outside quotes, arguments from the pre- and postjob are split on linear whitespace. The backslash makes the next character verbatim. Variables that are rewritten must start with a dollar sign either outside quotes or inside double quotes. The dollar may be followed by a valid identifier. A valid identifier starts with a letter or the underscore. A valid identifier may contain further letters, digits or underscores. The identifier is case sensitive. The alternative use is to enclose the identifier inside curly braces. In this case, almost any character is allowed for the identifier, including whitespace. This is the only curly brace expansion. No other Bourne magic involving curly braces is supported. One of the advantages of variable substitution is, for example, the ability to specify the application as $HOME/bin/app1 in the transformation catalog, and thus to gridstart. As long as your home directory on any compute node has a bin directory that contains the application, the transformation catalog does not need to care about the true location of the application path on each pool. Even better, an administrator may decide to move your home directory to a different place. As long as the compute node is set up correctly, you don’t have to adjust any Pegasus data. Mind that variable substitution is an expert feature, as some degree of tricky quoting is required to protect substitutable variables and quotes from Globus, Condor and Pegasus in that order. Note that Condor uses the dollar sign for its own variables. The variable substitution assumptions for the main job differ slightly from the prejob and postjob for technical reasons. The pre- and postjob command lines are passed as one string. However, the main jobs command line is already split into pieces by the time it reaches pegasus-kickstart. Thus, any whitespace on the main job’s command line must be preserved, and further argument splitting avoided. It is highly recommended to experiment on the Unix command line with the echo and env applications to obtain a feeling for the different quoting mechanisms needed to achieve variable substitution.

EXAMPLE¶

You can run the pegasus-kickstart executable locally to verify that it is functioning well. In the initial phase, the format of the performance data may be slightly adjusted. Please take note a few things in the above example: The output from the postjob is appended to the output of the main job on stdout. The output could potentially be separated into different data sections through different temporary files. If you truly need the separation, request that feature. The log file is reported with a size of zero, because the log file did indeed barely exist at the time the data structure was (re-) initialized. With regular GASS output, it will report the status of the socket file descriptor, though. The file descriptors reported for the temporary files are from the perspective of pegasus-kickstart. Since the temporary files have the close-on-exec flag set, pegasus-kickstarts file descriptors are invisible to the job processes. Still, the 'stdio of the job processes are connected to the temporary files. Even this output already appears large. The output may already be too large to guarantee that the append operation on networked pipes (GASS, NFS) are atomically written. The current format of the performance data is as follows:

OUTPUT FORMAT¶

Refer to http://pegasus.isi.edu/wms/docs/schemas/iv-2.2/iv-2.2.html for an up-to-date description of elements and their attributes. Check with http://pegasus.isi.edu/documentation for invocation schemas with a higher version number.

RESTRICTIONS¶

There is no version for the Condor standard universe. It is simply not possible within the constraints of Condor. Due to its very nature, pegasus-kickstart will also prove difficult to port outside the Unix environment. Any of the pre-, main-, cleanup and postjob are unable to influence one another’s visible environment. Do not use a Pegasus transformation with just the name null and no namespace nor version. First Condor, and then Unix, place a limit on the length of the command line. The additional space required for the gridstart invocation may silently overflow the maximum space, and cause applications to fail. If you suspect to work with many argument, try an argument-file based approach. A job failing with exit code 126 or 127 is indistinguishable from pegasus-kickstart failing with the same exit codes. Sometimes, careful examination of the returned data can help. If the logfile is collected into a shared file, due to the size of the data, simultaneous appends on a shared filesystem from different machines may still mangle data. Currently, file locking is not even attempted, although all data is written atomically from the perspective of pegasus-kickstart. The upper limit of characters of command line characters is currently not checked by pegasus-kickstart. Thus, some variable substitutions could potentially result in a command line that is larger than permissible. If the output or error file is opened in append mode, but the application decides to truncate its output file, as in the above example by opening /dev/fd/1 inside keg, the resulting file will still be truncated. This is correct behavior, but sometimes not obvious.

FILES¶

/usr/share/pegasus/schema/iv-2.2.xsd

ENVIRONMENT VARIABLES¶

GRIDSTART_TMP TMP TEMP TMPDIR GRIDSTART_SETUP GRIDSTART_PREJOB GRIDSTART_POSTJOB GRIDSTART_CLEANUP

HISTORY¶

As you may have noticed, pegasus-kickstart had the name kickstart in previous incantations. We are slowly moving to the new name to avoid clashes in a larger OS installation setting. However, there is no pertinent need to change the internal name, too, as no name clashes are expected.

AUTHORS¶

Michael Milligan <mbmillig@uchicago.edu> Mike Wilde <wilde@mcs.anl.gov> Yong Zhao <yongzh@cs.uchicago.edu> Jens-S. Vöckler <voeckler@isi.edu> Gideon Juve <gideon@isi.edu> Pegasus Team http://pegasus.isi.edu/