Condor for Debian
=================

The documentation for Condor can be found on the World Wide Web at:

http://www.cs.wisc.edu/condor/manual

At that web site, you will find both an on-line version of manual and
directions for downloading your own copy in a variety of formats.


Package upgrades
----------------

When the Condor package is upgraded it will restart the Condor daemons. If this
happens on the central manager it will result in jobs being killed on execute
nodes. To avoid this, it is best to gracefully shutdown the Condor pool before
upgrading. The Debian package does not perform a graceful shutdown on upgrades,
because there is a variety of possible procedures, where subjective
appropriateness depends on the configuration and utilization of a particular
Condor pool. Please see this page for more information on graceful shutdowns:

  https://condor-wiki.cs.wisc.edu/index.cgi/wiki?p=HowToShutDownCondor


Snapshotting of vanilla universe jobs
-------------------------------------

The Debian package comes with built-in support for snapshotting of vanilla
universe jobs with DMTCP. This is useful for migrating running jobs between
machines of a condor pool without having to relink the respective binaries with
Condor's standard universe LIBC wrapper.  The main advantage of DMTCP over
Condor's native mechanism is that it also works for scripts (interpreter
sessions like Python), and closed-source systems (e.g. Matlab). The necessary
shim script (a wrapper around the actual job executable) is available at
/usr/lib/condor/shim_dmtcp. The condor-doc package comes with example submit
files for self-compiled executables and system-wide installed software.  A
submit configuration for a job with on-demand snapshotting on a Debian cluster
could look like this:

  # DMTCP checkpointing related boilerplate (should be applicable in most cases)
  universe = vanilla
  executable = /usr/lib/condor/shim_dmtcp
  should_transfer_files = YES
  when_to_transfer_output = ON_EXIT_OR_EVICT
  kill_sig = 2
  output = $(CLUSTER).$(PROCESS).shimout
  error = $(CLUSTER).$(PROCESS).shimerr
  log = $(CLUSTER).$(PROCESS).log
  dmtcp_args = --log $(CLUSTER).$(PROCESS).shimlog --stdout $(CLUSTER).$(PROCESS).out --stderr $(CLUSTER).$(PROCESS).err
  dmtcp_env = DMTCP_TMPDIR=./;JALIB_STDERR_PATH=/dev/null;DMTCP_PREFIX_ID=$(CLUSTER)_$(PROCESS)

  # job-specific setup, uses arguments and environment defined above
  arguments = $(dmtcp_args) some_command some_arg1 some_arg2
  environment = $(dmtcp_env)

  queue 1

 -- Michael Hanke <michael.hanke@gmail.com>  Fri, 09 Mar 2012 8:19:25 +0200