The Design And Implementation Of The Load Balancing System In The Non-uniform Memory Access Cluster

The time of processor accessing to memory depends on the memorylocation relative to the processor in non-uniform memory access cluster, andthe processor access remote memory will costs much more time than accesslocal memory. At the same time, if the application does not specify whichprocessor the process will run on explicitly, process migration on the processorwill reduce the cache hit rate and reduce application performance.This thesis is about load balancing in the non-uniform memory accesscluster. The project is to allocate hardware resources in non-uniform memoryaccess architecture, based on user’s job requests. The user’s request, includingCPU resources, memory resources and job distribution policy. The request ofCPU resource contains type, size, scope and resources is exclusive or not; therequest of memory resources contains size, the memory should be local or canbe remote; job distribution strategy means the tasks are distributed centralizedor balance. The system will meet the user’s request, allocate most suitable hostresources to jobs and hard binding resources.The purpose is to improve theuser application performance, and save resource costs.In this thesis, it will design and implement six modules, user input module,topology information collection module, the mark available resources module,the job distribution policy module, the scheduling policy propagation moduleand the log module. In user input module, user submits job resource requestsand query jobs and host related information; In topology informationcollection module, the hardware resources will be collected in cluster,including CPU and memory resources; In marks available resources module,indexing structure will be built to improve the resource discovery efficiency;In jos distribution policy module, select the most suitable host resource to job,to meet user requirements about job distribution centralized, reduced cachemiss latency, or to meet the user requirements job distribution balance, avoidresource contention; In scheduling policy propagation module, the hosts willbe divided, improving the speed of propagation scheduling strategy greatly; Inlogging module, record job and system information, track system status andtrack system status and diagnose system exceptions.The system has been tested and the results prove that the allocation ofresources users can meet the user’s request about CPU and memory resources,and distribute jobs centralized or balanced, according to job distribution policy.At the same time, the system also meets the requirements of usability, robustness and scalability, it has high practical value.