Introduction

A goal of grid computing is to allow an authorized batch scheduler to send jobs to run on some remote pool, even when that remote pool is running a non-HTCondor system.

There are several mechanisms in HTCondor to do this.

Flocking allows HTCondor jobs submitted from one pool to execute on another, separate HTCondor pool. Flocking is enabled by configuration on both of the pools. An advantage to flocking is that jobs migrate from one pool to another based on the availability of machines to execute jobs. When the local HTCondor pool is not able to run the job (due to a lack of currently available machines), the job flocks to another pool. A second advantage to using flocking is that the submitting user does not need to be concerned with any aspects of the job. The user’s submit description file (and the job’s universe) are independent of the flocking mechanism. Flocking only works when the remote pool is also an HTCondor pool.

Glidein is the technique where condor_startds are submitted as jobs to some remote batch systems, and configured with report to, and expand the local HTCondor batch system. We call these jobs that run startds “pilot jobs”, to distinguish them from the “payload jobs” which run the real user’s domain work. HTCondor itself does not provide an implementation of glidein, there is a very complete implementation the HEP community has built, named GlideinWMS, and several HTCondor users have written their own glidein systems.

Other forms of grid computing are enabled by using the grid universe and further specified with the grid_type. For any HTCondor job, the job is submitted on a machine in the local HTCondor pool. The location where it is executed is identified as the remote machine or remote resource. These various grid computing mechanisms offered by HTCondor are distinguished by the software running on the remote resource. Often implementations of Glidein use grid universe to send the pilot jobs to a remote system.

When HTCondor is running on the remote resource, and the desired grid computing mechanism is to move the job from the local pool’s job queue to the remote pool’s job queue, it is called HTCondor-C. The job is submitted using the grid universe, and the grid_type is condor. HTCondor-C jobs have the advantage that once the job has moved to the remote pool’s job queue, a network partition does not affect the execution of the job. A further advantage of HTCondor-C jobs is that the universe of the job at the remote resource is not restricted.

One disadvantage of grid universe is the destination must be declared in the submit file when condor_submit is run, locking the job to that remote site. The condor job router is a condor daemon which can periodically scan the scheduler’s job queue, and change a vanilla universe job intended to run on the local cluster into a grid job, destined for a remote cluster. It can also be configured so that if this grid job is idle for too long, it can undo the transformation, so that the job isn’t stuck forever in a remote queue.

Further specification of a grid universe job is done within the grid_resource command in a submit description file.