The HTCondor daemons do not run authenticated to AFS; they do not possess AFS tokens. Therefore, no child process of HTCondor will be AFS authenticated. The implication of this is that you must set file permissions so that your job can access any necessary files residing on an AFS volume without relying on having your AFS permissions.
If a job you submit to HTCondor needs to access files residing in AFS, you have the following choices:
The Center for High Throughput Computing hopes to improve upon how HTCondor deals with AFS authentication in a subsequent release.
Please see section 3.14.1 for further discussion of this problem.
If the current working directory when a job is submitted is accessed via an NFS automounter, HTCondor may have problems if the automounter later decides to unmount the volume before the job has completed. This is because condor_submit likely has stored the dynamic mount point as the job’s initial current working directory, and this mount point could become automatically unmounted by the automounter.
There is a simple work around. When submitting the job, use the submit command initialdir to point to the stable access point. For example, suppose the NFS automounter is configured to mount a volume at mount point /a/myserver.company.com/vol1/johndoe whenever the directory /home/johndoe is accessed. Adding the following line to the submit description file solves the problem.
HTCondor attempts to flush the NFS cache on a submit machine in order to refresh a job’s initial working directory. This allows files written by the job into an NFS mounted initial working directory to be immediately visible on the submit machine. Since the flush operation can require multiple round trips to the NFS server, it is expensive. Therefore, a job may disable the flushing by setting
in the job’s submit description file. See page 2411 for a definition of the job ClassAd attribute.
HTCondor is normally installed such that the HTCondor daemons have root permission. This allows HTCondor to run the condor_shadow daemon and the job with the submitting user’s UID and file access rights. When HTCondor is started as root, HTCondor jobs can access whatever files the user that submits the jobs can.
However, it is possible that the HTCondor installation does not have root access, or has decided not to run the daemons as root. That is unfortunate, since HTCondor is designed to be run as root. To see if HTCondor is running as root on a specific machine, use the command
where <machine-name> is the name of the specified machine. This command displays the full condor_master ClassAd; if the attribute RealUid equals zero, then the HTCondor daemons are indeed running with root access. If the RealUid attribute is not zero, then the HTCondor daemons do not have root access.
NOTE: The Unix program ps is not an effective method of determining if HTCondor is running with root access. When using ps, it may often appear that the daemons are running as the condor user instead of root. However, note that the ps command shows the current effective owner of the process, not the real owner. (See the getuid(2) and geteuid(2) Unix man pages for details.) In Unix, a process running under the real UID of root may switch its effective UID. (See the seteuid(2) man page.) For security reasons, the daemons only set the effective UID to root when absolutely necessary, as it will be to perform a privileged operation.
If daemons are not running with root access, make any and all files and/or directories that the job will touch readable and/or writable by the UID (user id) specified by the RealUid attribute. Often this may mean using the Unix command chmod 777 on the directory from which the HTCondor job is submitted.
A job lease specifies how long a given job will attempt to run on a remote resource, even if that resource loses contact with the submitting machine. Similarly, it is the length of time the submitting machine will spend trying to reconnect to the (now disconnected) execution host, before the submitting machine gives up and tries to claim another resource to run the job. The goal aims at run only once semantics, so that the condor_schedd daemon does not allow the same job to run on multiple sites simultaneously.
If the submitting machine is alive, it periodically renews the job lease, and all is well. If the submitting machine is dead, or the network goes down, the job lease will no longer be renewed. Eventually the lease expires. While the lease has not expired, the execute host continues to try to run the job, in the hope that the submit machine will come back to life and reconnect. If the job completes and the lease has not expired, yet the submitting machine is still dead, the condor_starter daemon will wait for a condor_shadow daemon to reconnect, before sending final information on the job, and its output files. Should the lease expire, the condor_startd daemon kills off the condor_starter daemon and user job.
A default value equal to 40 minutes exists for a job’s ClassAd attribute JobLeaseDuration, or this attribute may be set in the submit description file, using job_lease_duration, to keep a job running in the case that the submit side no longer renews the lease. There is a trade off in setting the value of job_lease_duration. Too small a value, and the job might get killed before the submitting machine has a chance to recover. Forward progress on the job will be lost. Too large a value, and an execute resource will be tied up waiting for the job lease to expire. The value should be chosen based on how long the user is willing to tie up the execute machines, how quickly submit machines come back up, and how much work would be lost if the lease expires, the job is killed, and the job must start over from its beginning.
As a special case, a submit description file setting of
as well as utilizing submission other than condor_submit that do not set JobLeaseDuration (such as using the web services interface) results in the corresponding job ClassAd attribute to be explicitly undefined. This has the further effect of changing the duration of a claim lease, the amount of time that the execution machine waits before dropping a claim due to missing keep alive messages.