|
Research communities
are encouraged to choose parallel computers according to the needs
of the different user applications. Optimal GRID scheduling strategies
can help to choose a set of parallel machines with an optimal price/performance
ratio. For this purpose, the applications and the different computing
resources are parameterized (three of them could be used to characterize
the TOP500 machines). These parameters as well as information on the
machine
loads and functionalities are used to evaluate a cost function model that predicts
the best-suited machine. Besides economical quantities, this cost function
can also include ecological aspects. We propose to adapt processor
cycle periods to the application needs in order to drastically reduce
power consumption. These strategies are demonstrated with three specific
applications executed on different capacity and capability machines. |