摘要
To solve the load balancing problem in a triplet-based hierarchical interconnection network（THIN） system, a dynamic load balancing （DLB）algorithm--THINDLBA, which adopts multicast tree （MT）technology to improve the efficiency of interchanging load information, is presented. To support the algorithm, a complete set of DLB messages and a schema of maintaining DLB information in each processing node are designed. The load migration request messages from the heavily loaded node （HLN）are spread along an MT whose root is the HLN. And the lightly loaded nodes（LLNs） covered by the MT are the candidate destinations of load migration; the load information interchanged between the LLNs and the HLN can be transmitted along the MT. So the HLN can migrate excess loads out as many as possible during a one time execution of the THINDLBA, and its load state can be improved as quickly as possible. To avoid wrongly transmitted or redundant DLB messages due to MT overlapping, the MT construction is restricted in the design of the THINDLBA. Through experiments, the effectiveness of four DLB algorithms are compared, and the results show that the THINDLBA can effectively decrease the time costs of THIN systems in dealing with large scale computeintensive tasks more than others.
为了解决基三分层互连网络(THIN)系统中的负载平衡问题,提出一种采用多播树技术提高节点间交换负载信息效率的动态负载平衡(DLB)算法——THINDLBA.设计了一套完整的DLB消息和各节点处的信息维护机制以辅助算法实现.重载节点的负载迁移请求消息沿着一棵以该节点为根的多播树传播,被该树覆盖的轻载节点均成为负载迁移的候选目标节点,可以沿着该树和重载节点交互负载信息,从而使重载节点能够在算法的一次执行中外迁最多的过载进程,尽快改善自身负载状态.算法设计中约束了多播树的构造过程,以避免因树间覆盖造成的消息误传或冗余.通过实验对比了4种DLB算法的性能,结果证明THINDLBA能更有效地缩减THIN系统处理计算密集型任务的时间.
基金
The National Natural Science Foundation of China(No.69973007).
