Analysis of empty ATLAS pilot jobs

In this analysis we quantify the wallclock time used by short empty pilot jobs on a number of WLCG compute resources. Pilot factory logs and site batch logs are used to provide independent accounts of the usage. Results show a wide variation of wallclock time used by short jobs depending on the site and queue, and changing with time. For a reference dataset of all jobs in August 2016, the fraction of wallclock time used by empty jobs per studied site ranged from 0.1% to 0.8%. The variation in wallclock usage may be explained by different workloads for each resource with a greater fraction when the workload is low. Aside from the wall time used by empty pilots, we also looked at how many pilots were empty as a fraction of all pilots sent. Binning the August dataset into days, empty fractions between 2% and 90% were observed. The higher fractions correlate well with periods of few actual payloads being sent to the site.