S. Dal Pra (INFN-CNAF)

The default behaviour of a batch system is to dispatch jobs to nodes having the lower value of some load index. Whilst this causes jobs to be equally distributed among all the nodes in the farm, there are cases when different types of behaviour may be desirable, such as having a completely full node before dispatching jobs to another one, or having similar jobs dispatched to nodes already running jobs of the same kind. This work defines the packing concept, different packing policies and useful metrics to evaluate how good the policy is. A simple farm simulator has been written to evaluate the expected impact on a farm of different packing policy. The simulator is run against a sequence of real jobs, whose parameters have been taken from the accounting database of INFN-Tier1. The effectiveness of two packing policies of interest, namely relaxed and exclusive, are compared. The exclusive policy proves to be better, at the cost of unused cores in the farm, whose number is estimated. The possibility of implementing the exclusive policy on a specific batch system, LSF 7.06, is exploited. Relevant configurations are shown and an overall description of the mechanism is presented.