This report describes the design, implementation and test of a software able to monitor the load of a cluster IBM Power 6, acting on its nodes switching them on and off on the basis of their load (jobs), scheduled by LSF, taking into account a linear, integer variables optimization model which produces a feasible schedule, possibly optimal, of pending jobs, given running jobs. Optimizations happens in two phases: optimization of the energy cost function, associated to each cluster node; optimization of makespan in order to avoid holes in the schedule produced by the 1st optimization. Machines which are idle in the schedule can be powered off. The aim is to save energy, when possible, consumed by IBM Power 6 cluster. Moreover, turning off idle nodes will also imply less heat produced by the cluster, and additional savings on the cooling system.
CMCC InstitutesCMCC Divisions