DEPLOYING PARALLEL NUMERICAL LIBRARY ROUTINES TO CLUSTER COMPUTING IN A SELF ADAPTING FASHION
Kenneth J. Roche, Jack J. Dongarra · 2002
This paper discusses middleware under development which couples cluster system information with the specifics of a user problem to launch cluster based applications on the best set of resources available. The user is responsible for stating his numerical problem and is assumed to be working in a serial environment. He calls the middleware to execute his application. The middleware assesses the possibility of solving the problem faster on some subset of available resources based on information describing the state of the system. If so, the user's data is redistributed over the subset of processors, the problem is executed in parallel, and the solution is returned to the user. If it is no faster, or slower, the user's problem is solved with the best appropriate serial routine. The feasibility of this approach is empirically investigated on a typical target system and results reported which validate the method