高性能计算过程中基于网络带宽代价的节点决策算法研究

A Better Seeker of Node in High Performance Computing Environment

针对高性能计算中由于网络延迟造成的地理位置分散的节点资源使用率不均匀现象,提出了一种自适应的节点决策算法。该算法通过估计待选节点的作业执行代价,选择一个代价最小的节点作为作业的执行节点,提高了作业的响应速度,高效地利用了计算资源。文中首先给出了远程作业执行的模型,其次给出了网络通信特征以及自适应代价模型,最后提出了基于网络带宽代价的节点决策算法并对该算法进行了系统的评价,取得良好的效果。

Aim. Existing methods for seeking node are,in our opinion,not quite efficient.We propose what we believe to be a better method.In the full paper,we explain our node-seeking method in detail.In this abstract,we just add some pertinent remarks to listing the four topics of explanation.The first topic is： the remote computing model,as shown in Fig.1 in the full paper.In this topic,we discuss the four procedures of the model： the startup for transferring start data from original node to target node,execution,checkpoint,and stage-out.The second topic is： the equations for estimating the delay in data transfer.Eq.3 in the full paper calculates the delay in stage-in and stage-out,while Eq.4 calculates the delay in producing checkpoint files during execution and that in communicating the original node.The third topic is： the available network bandwidth cost estimation model.In this topic,we present Eq.6 to calculate the cost,namely given a relative cost due to data transfer,the computer can produce the smallest bandwidth,which is,in turn,the decision-making criterion.The criterion decides the time when a job is transferred to the mode that satisfies the network bandwidth for the job to be efficiently executed.The fourth topic is： an algorithm for seeking the best adaptive node.In this topic,we use the network bandwidth cost estimation criterion to seek again and again the optimal node which executes the job.Finally we do experiments to appraise our method.Table 1 in the full paper shows preliminarily that the response time of our better method based on node-seeking algorithm is quite efficient in comparison with the methods that use data node priority,task node priority and the combination of the two.The application of the algorithm to a small job management system achieves an average of 70% utilization of idle resources.