Design And Realization Of The Global Scheduling Mechanism In The Isomerous Multiple-cluster System

Grid is one of the popular High Performance Computing Systemsin the world. It can collect resources information from ResourceLayer,allocate resource to job or reservation according to certainpolicies,and submit job to the relevant compute node. The resourcefor execute job and reservation,is transparent to user completely.But it has some defects,such as grid can not be enough in queueworks and scheduling jobs between two suppositional organizations.Cluster even has more computing efficiency than grid. But thesum total of nodes in a cluster is very limite and the nodes of asimple cluster are all computers,so the sum total of computers ina cluster is very limite. And because the different clusters usedifferent protocols,a job worked in one cluster is hard totransplant to another cluster,so the development of cluster isrestricted.This paper presents the defects of grid and cluster,pose theimagine that the Load Shared Facility,LSF and the Sun Grid Engine,SGE can be composed together to build a isomerous multiple-clustersystem.Then a special scheduling algorithm named DistributionDepending on Probability Algorithm,which does well in the globalscheduling, is designed. And then,one of the independent jobscheduling algorithms , Maximum Intercal Heuristic SchedulingAlgorithm,Max-Int Algorithm,ischosen, and use it to schedule jobsfor the main cluster.In the end,a plug-in, which can compose LSFand SGE by the help of the plug-in mechanism of LSF scheduler,is designed, and realize it at last.And the measure and tests ofthe mechantism of the global scheduling and the independent jobscheduling in main cluster are in progress too. The imagine of theisomerous multiple-cluster system is proved buy the effective data.In the 1^st paragraph of this paper,it introduces the presentconditions briefly. And In the 2^ed paragraph of this paper, itintroduces the function and principle of LSF and SGE at first, thenintroduce how we build the isomerous multiple-cluster system.The 3^rd paragraph is the focal point of this paper. In thisparagraph,it detailed expounds the Distribution Depending onProbability Algorithm , which does well in the isomerousmultiple-cluster system,and the global scheduling module which isdesigned under the Distribution Depending on Probability Algorithm.At first, the job queue works and job scheduling works of the globalscheduling module isdefined in the aim of design. Then, the globalscheduling module isdivided into three modules, queue module,allocating job share of each clusters module,allocating job module.In the queue module,it use the First In First Out Mechanism andthe Priority Mechanism to stand the jobs in line. In the allocatingjob share of each clusters module and allocating job module,itdepend on the statistics and probability,find the Probability ofTurn Around on Time. Then use it as the standard of the isomerousmultiple-cluster system,and realize the Distribution Depending onProbability Algorithm at last.The 4th paragraph is the addition of the 3 rd paragraph. Aftercomparing many kinds of different independent job schedulingalgorithms, the Max-Int Algorithm,iswe chosen and improved. Thenthe Improved Max-Int Algorithm is used to schedule jobs for the maincluster. Thus far,we defind and realize the scheduling in the maincluster,and the global scheduling module isfinished.In the 5th paragraph of this paper,it presents how to designthe scheduler plug-in,and introduce the plug-in mechanism of LSFscheduler,the API Framework of LSF and how to program with the APIFramework. Then the API Framework of LSF and the Shell of SGE ismade together,and definded the bridge between LSF and SGE.At the Last of this paper,the measure and tests of themechantism of the global scheduling and the independent jobscheduling in main cluster are in progress. After the compare ofthe Fast-Greedy Algorithm,the Optimal Selection Theory Algorithminproved by the mechanism of maximum-communication and theDistribution Depending on Probability Algorithm,the truth is foundthat the efficiency of the Distribution Depending on ProbabilityAlgorithm improves 5% than the Fast-Greedy Algorithm.So the designof the Distribution Depending on Probability Algorithm is relativesuccess.Then after the compare of the Fast-Greedy Algorithm,theMin-Min Algorithm,the Max-Min Algorithm and the Improved Max-IntAlgorithm,and find the Improved Max-Int Algorithm is the bestone.So we can say the Improvement of the Max-Int Algorithm isrelative success too.By the reseach of isomerous multiple-cluster system,This paperopens the defects of grid and cluster out,and pose the imagine thatLSF and SGE can be composed together to build a isomerousmultiple-cluster system. Then a special scheduling algorithm namedDistribution Depending on Probability Algorithm,which does wellin the global scheduling,and a plug-in for it,is designed andrealized. At last,measure and tests ,the Distribution Dependingon Probability Algorithm is define relative Valuable. So theresearch of the global scheduling mechanism in the isomerousmultiple-cluster system is relative Valuable too.