论文部分内容阅读
针对MASNUM全球海浪数值模式,设计了一种负载近绝对均衡的非规则类矩形剖分及计算格点序列化排序与交换方案,建立了高效并行海浪数值模式.基于国家超算济南中心的神威蓝光平台,对建立的全球并行海浪模式进行了1°×1°,(1/2)°×(1/2)°,(1/4)°×(1/4)°和(1/8)°×(1/8)°四种不同水平分辨率的无IO并行测试,测试规模达到131072进程.测试结果表明:随着进程数目的增多,四种分辨率并行模式的计算速度均在提高.当进程数目为基准进程数目的4倍时,四种分辨率的加速效率均在80%以上;当进程数目达到8倍基准进程数目时,1°×1°,(1/2)°×(1/2)°和(1/4)°×(1/4)°三种分辨率加速效率都还在80%以上,而此时(1/8)°×(1/8)°分辨率使用131072进程的加速效率仍可达62%;当进程数目为24倍基准进程数目时,1°×1°分辨率的加速效率可达72%,(1/2)°×(1/2)°分辨率的加速效率为62%,而此时(1/4)°×(1/4)°分辨率使用98304进程的加速效率仍可达到38%.从加速比和加速效率来看,采用非规则类矩形剖分的并行方案可以使得全球海浪数值模式具有高并行效率.
According to MASNUM global ocean wave numerical model, a non-regular rectangular subdivision with approximate absolute equilibrium load and serialization and exchange of grids are designed, and an efficient numerical model of parallel ocean waves is established.Based on the state- (1 ° × 1 °, (1/2) ° × (1/2) °, (1/4) ° × (1/4) °, and (1/8)) of the established global parallel wave model. ° × (1/8) ° without IO test, the test scale reaches 131072. The test results show that as the number of processes increases, the computational speed of the four resolution parallel modes increases. When the number of processes is four times of the number of processes, the acceleration efficiency of the four resolutions is above 80%. When the number of processes reaches eight times the number of processes, 1 ° × 1 °, (1/2) ° × 1/2) ° and (1/4) ° × (1/4) ° Acceleration efficiencies are still above 80% at this time (1/8) ° × (1/8) ° Resolution The acceleration efficiency of the 131072 process can still reach 62%. When the number of processes is 24 times the number of reference processes, the acceleration efficiency of 1 ° × 1 ° resolution can reach 72%, (1/2) ° × (1/2) The resolution of the acceleration efficiency of 62%, while (1/4) ° × (1/4) ° resolution The rate of 98304 acceleration can still reach 38% with the acceleration rate and acceleration efficiency, the parallel scheme with irregular rectangular split can make the global wave numerical model have high parallel efficiency.