三维波动方程正演的三级并行加速

发布时间：2018-06-28 13:12

本文选题：三维波动方程正演模拟 + 并行计算　；参考：《石油地球物理勘探》2016年05期

【摘要】：在常规并行方案的基础上,继续挖掘CPU计算潜力,从CPU指令级优化入手,借用CPU的矢量运算单元(VALU)和SSE指令集,实现了在一个指令周期内并行完成四个浮点数据运算,得到以下认识:1对有限差分类方法的波动方程正演而言,使用SSE指令集可以取得较好的加速效果,能够实现CPU的二次加速;2引入SSE后的加速比会随着正演模型数据量的增大而缓慢增大,但因其一次最多完成4个浮点型数据运算,理论加速比最大不会超过4;3SSE加速不需要额外增加硬件配置就可实现计算效率提升,加速成本低,有很广泛的适用性;4单机执行三级并行能够获得最佳的执行效率,多机间执行三级并行效率与机间数据传输网络速度有关。通过数值模拟实验发现,新的并行方案较常规并行方案在运算速度上有大幅提升,获得了明显的加速效果。
[Abstract]:On the basis of the conventional parallel scheme, we continue to excavate the CPU computing potential, starting with the CPU instruction level optimization, using the vector operation unit (VALU) and SSE instruction set of CPU to achieve four floating point data operations in one instruction cycle, and get the following understanding: 1 the use of the SSE instruction for the wave equation forward of the finite difference classification method. The set can achieve better acceleration and achieve two acceleration of CPU; 2 the acceleration ratio after the introduction of SSE will increase slowly with the increase of the data of the forward model, but the theoretical acceleration will not exceed 4 because of the maximum completion of 4 floating-point data operations at one time, and the 3SSE addition speed does not need additional hardware configuration to achieve the calculation. The efficiency is improved, the accelerated cost is low, and it has a wide range of applicability; 4 single machine execution of three level parallel can obtain the best execution efficiency. The efficiency of the multi machine execution level three parallel is related to the speed of the inter machine data transmission network. An obvious acceleration.
【作者单位】：西南石油大学地球科学与技术学院;
【分类号】：P631

【相似文献】