参数扫描应用的网格自适应调度

提出一种自适应调度方法，可使参数扫描应用在运行时保持近似的网格负载平衡，为适应资源性能的动态性，一个参数应用被分成若干作业顺序调度，每个作业运行后反馈的网格负载信息用于调整下一个作业的调度。使之适应资源性能波动，从而使后一个作业在运行时实现网格负载的近似平衡。每个作业被分成若干子作业，分别指派到不同的网格资源。子作业的运行时间构成一个网格负载向量，从中可计算出网格负载失衡因子，失衡因子表示作业运行时网格负载失衡的程度。负载向量用于调整下一个作业的划分方法，失衡因子用于调整下一个作业的规模。较小的失衡因子可使作业的规模有更快的增长，这样可使调度成本的增长速度慢于应用规模的增大速度。实验表明，自适应调度可保持近似的网格负载平衡，与其他资源性能的静态调度相比，可有效缩短参数应用的总运行时间。

Adaptive Scheduling for Parameter Sweep Applications in Grid

A kind of adaptive scheduling was introduced to keep the approximate balance of grid load during parameter sweep applications' run. To adapt to the dynamic of resource capabilities, a parameter application was divided into some jobs scheduled sequentially. After each job finished its run, the feedback information about grid load was used to adjust the scheduling of next job to fit into the fluctuations of resource capacities, thus approximate grid load balancing could be implemented while the next job was running. Each job was divided into some sub-jobs which were assigned to different grid resources. According to sub-job execution time, a grid load vector was constructed, from which a gird load unbalance index indicating the unbalance extent of gird load during job's run was obtained. The load vector was used to adjust the partition of next job, and the unbalance index was used to change the scale of next job. The smaller indexes made job scale increase more rapidly, thus the increase of scheduling cost was slower than that of application scale. The experiments show that the adaptive scheduling can keep approximate balance of grid load, and can obviously shorten the total execution time of parameter applications in contrast with the static scheduling based on resource performance.