ATLAS IaaS Draft

In collaboration with the ATLAS Distributed Computing Cloud R&D group we have developed a method for running regular ATLAS jobs from Panda workload management system using dynamically provisioned Virtual Machines (VMs). We use Condor with a suporting package called Cloud Scheduler(https://github.com/hep-gc/cloud-scheduler) to respond dynamically to the load on a Panda queue. The system has been in operation since the first week of April and has completed a total of 42000 jobs in it's first two months of operation (Figure 1), with a total cumulative CPU time of 19 years (Figure 2). All data is stored at UVic and the VMs are excuted onthe FutureGrid cloud resources physically located at the University of Chicago (Hotel Cloud), and the University of Florida (Foxtrot Cloud).

Figure 1: Completed Jobs.

Figure 2: CPU Consumption

We are effectively using ATLAS applications to backfil available resources on these sites. Figure 3 shows the number of running cores which are dispersed across the the available clouds. As can been the number of available cores changes depending on the number of free IaaS slots.

Figure 3: Average number of cores used per day