The Design And Implementation Of Allocating Job Resources By Multi-granularity In Grid

Grid aims at providing the sharing of platform or resources, can integrateheterogeneous computing resources distributed in a certain geographical area. Tosolve large-scale parallel computational problems, the grid provides a relativelyinexpensive high-performance computing model.LSF (Load Sharing Facility), the well-known product of the grid community, isa kind of grid computing software, which can manage computing resources. It canprovide users with a single system image of different computing resources. In thecluster that is managed by LSF, it is common to run parallel jobs that require a lot ofcomputing resources provided by multiple hosts. When the job is running,interactive information communicating between these hosts will consume a largeamount of network bandwidth, so that the performance of the entire cluster is slowdown.The issue of the paper proposes system-level solutions to the problem ofparallel jobs affecting system performance in the cluster managed by LSF. Whenscheduling jobs, users can request resources of different granularity and differentamounts, according to the cluster’s network topology characteristics. At the sametime, users can also restrict the methods of allocating resources by adding someconstraints. For different constraints, LSF will follow different algorithms toallocate resources, and ultimately run the job.The module of allocating job resources by multi-granularity in grid implementsthis solution, and adds it to LSF as an functional extension. On the bases ofunchanging LSF’s existing framework, the module added three sub-modules, whichimplement different job scheduling strategies respectively. The three sub-modulesuse the LSF existing frameworks and API, can help LSF identify user’sconfiguration of resource granularity information, and allocate resources inaccordance with the users requirements for the job. How to save the resourceinformation, scheduling jobs correctly and updating the status of resources promptlyare the difficulties of the system implementation. Therefore the system adds datastructures to save the resource information, designs and realises several parallel jobscheduling algorithms.After the completion of system development, when LSF scheduling paralleljobs across multiple hosts, the scheduling time will increase, but the job executiontime will decrease, i.e, the system efficiency increases significantly. Therefore, the subject achieves the purpose of significant increase in the LSF overall systemperformance when running parallel jobs.