并行数值模式超线性加速问题的个例研究

Case Study on the Superlinear Speedup of a Parallel Numerical Model

利用区域分解方法,对模拟大气重力波在中高层大气中非线性传播过程的数值模式进行了并行化处理,从而建立了并行数值模式.在两个并行机群上对并行数值模式的测试结果表明:在千兆以太网和内存配置较低的机群系统上,并行效率随着进程个数的增加而减低,当进程个数增加到16时,并行效率仅为0.65.在内存配置较高的机群系统上,并行效率随着进程个数的增加并不减小,甚至出现了超线性加速现象.比如当进程个数为160个时,并行效率达到1.28.通过比较研究可以说明,机群系统的缓存、物理内存和网络通讯延时都是影响加速比的重要因素.尽可能大的缓存和物理内存能够有效提高加速比,网络交换机的选择需要在延时和带宽间进行权衡.

Domain decomposition method is employed to parallelize the numerical model, which is used to simulate the nonlinear propagation of gravity waves in the middle and upper atmosphere. The parallel numerical model has been tested on two parallel computer clusters. For the parallel computer cluster with gigahit ethernet and lower memory, the parallel efficiency is only 0.65 when the processors increases to 16. On the contrary, for the parallel computer cluster with Myrinet 2000 and large memory, the parallel efficiency increases with the number of processors. The superlinear speedup also occurs on this clus- ter since the computation efficiency can be as large as 1.28 when 160 processors are used. This study suggests that speedup is highly related to the cache, memory and network bandwidth. Thus, it is better to equip large cache and memory to increase the speedup of cluster. As for the network switch, it should be chosen on the tradeoff between time-delay and bandwidth.