基于前验负载差异的负载平衡性能模型  

Load-balancing evaluation model based on pre-defined variability

作　　者：张理论[1] 吴建平[1] 宋君强[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073

出　　处：《计算机应用》2009年第10期2849-2851,共3页journal of Computer Applications

基　　金：国家自然科学基金资助项目(40505023)

摘　　要：基于有限差分离散的并行应用非常普遍,针对此类问题的负载平衡性能评估,引入了一个刻画应用问题负载平衡能力的关键参数:最大负载变化率,推导了一个以并行效率为目标函数的负载平衡性能模型,涉及问题规模、并行通信计算比、离散格式复杂度和并行规模等。以POP全球海洋模式并行程序为测试实例,验证了该模型的性能。结果显示最大负载变化率作为衡量负载平衡程度的指标是有效的,基于模型的预测性能与实测性能在总体趋势上基本吻合。该性能模型对基于有限元、有限体积等其他局部离散格式的大型并行计算应用的负载平衡能力评估也具有参考价值。An evaluation model of load-balancing was presented for finite difference parallel computing. Maximal Load Variability （MLV） was introduced as a key index for load-balancing, and the quantitative model was constructed with parallel efficiency as the object function, involving ratio of communication to computing, degree of parallel, problem size and complexity of local numerical schemes. Both the parameter MLV and the model were verified by analyzing the POP global ocean circulation model benchmark. The results show that the performance derived from the evaluation model is consistent with that of wall-time measurement on the whole. And the model given in this paper is also useful for those parallel applications with local schemes such as finite element and finite volume.

关 键 词：数值并行计算 负载平衡 性能模型 负载变化率 通信计算比 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]