Intelligent GRID Scheduling System Ralf Gruber 1 , Vincent Keller 1 , Pierre Kuonen 5 , Marie-Christine Sawley 4 , Basile Schaeli 2 , Ali Tolou 1 , Marc Torruella 6 , Trach-Minh Tran 3 1 LIN-STI, Ecole Polytechnique F´ ed´ erale, CH-1015 Lausanne, Switzerland 2 LSP-I&C, Ecole Polytechnique F´ ed´ erale, CH-1015 Lausanne, Switzerland 3 CRPP, Ecole Polytechnique F´ ed´ erale, CH-1015 Lausanne, Switzerland 4 CSCS, CH-6928 Manno, Switzerland 5 Ecole d’ing´ enieurs et d’architectes, CH-1705 Fribourg, Switzerland 6 Universitat Politecnica Catalunya, E-08034 Barcelona, Espana Abstract The main objective of the Intelligent GRID Scheduling System (ISS) project is to provide a middleware infrastructure allowing op- timal positioning and scheduling of real life applications in a compu- tational GRID. According to data collected on the machines in the GRID, on the behaviour of the applications, and on the performance requirements demanded by the user, the best-suited computational re- source is detected and allocated to execute the application. The mon- itoring information collected during execution is put into a database and reused for the next resource allocation decision. In addition to providing scheduling information, the collected data allows to detect overloaded resources and to pin-point inefficient applications that could be further optimised. 1 Introduction The development of GRID technology for computational purposes is promis- ing. By harnessing a great number of different systems in a transparent manner, a user can have access to the computer architectures that are best suited to the constraints of its applications. The different communication needs of applications demand a GRID that can offer different parallel computer architectures including SMP and/or NUMA machines for shared memory parallel applications, a NoW (Net- work of Workstations) interconnected by a bus for embarassingly parallel applications, scalable but cost-effective networked clusters for applications dominated by point-to-point communications, and more expensive machines with faster networks for communication intensive applications. There is currently little feedback about applications that are not adapted to the hardware infrastructure, and little incentive to do so: if for instance a user notices that the network is too slow and hampers the performance of its application, he may try to find another machine to run it. On the 1