Organization of distributed computer systems' functioning at processing of sets of moldable jobs

Currently, high-performance information processing is based on the using of distributed computingsystems (DCS). Amount of elementary machines in such systems can reach up to 106. Themonoprogram mode is considered to be quite difficult.The paper considers the multiprogram mode of processing jobs of parallel programs. Thismode is used when the amount of programs and their characteristics are known in advance. Thecharacteristics of the programs include the programs rank (the number of parallel branches), theexecution time, the penalty for delaying of program execution, etc.Programs can be moldable, i.e. ones may have several options for the parameters which characterizenecessary for its execution configuration of the computing resources. Particular attention ispaid to the possibility of admissible rankings for program configurations of computing resources andtime spent by tasks in the DCS (due to the penalty function for the delay in the solution).The efficiency of a distributed computing system depends on methods of organizing functiontheir resources. The report proposed an approach for moldable jobs of schedule for execution forresource of distributed computing system with genetic optimization method. Schedules are basedon distribution jobs for packages, which will be subsequently load on the resources of DCS. Startsolutions time of each program is defined as the start solutions time corresponding to the package.Branches of parallel programs consistently distributed over EM.Each package is presented in the form of the gene. Genome is the collection of packages thatcontain all of the tasks set, each of which selected one of the possible values for the parameters.The initial set of schedules (a population) is constructed using algorithms for packing into containers.The algorithm of gene shuffling is used as the crossover. Mutation is a random selectionof one of the parameters of the job.The paper considers the implementation of serial and parallel genetic algorithms.

Keywords

Authors

References