６４位集群计算平台波动方程叠前深度偏移的性能优化

详细介绍了iCluster波动方程叠前深度偏移软件系统所采用的集群并行计算策略，以及在６４位集群计算平台上开展的计算性能优化工作和取得的成果。给出了大量软件优化和性能测试结果，包括波动方程叠前深度偏移处理程序运行分析、MPI并行计算运行分析、不同编译器和编译参数测试对比、不同FFT算法效率测试对比、IA32和IA64平台软件运行效率对比测试及并行加速比测试等结果。以原来基于2.6GHZ Xeon系统上的软件运行效率作为基准，在基于1.4GHz Itanium2的系统上，经过优化的波动方程叠前深度偏移程序其计算性能提高了９倍左右。其性能的提高主要归功于３个因素，即６４位处理器硬件性能的提高、优化编译器的贡献和Intel MKL数学库中DFT程序的贡献。试验和分析表明，对于波动方程叠前深度偏移计算，基于千兆以太网的集群计算机系统可以在较大规模配置下，仍然能保持很高的并行加速比。面对数据规模和计算规模的挑战，向６４位计算平台的迁移是未来几年内地球物理计算的必然发展方向。IA64平台上软件计算性能的优化，既有巨大潜力，又有可观的经济价值。

erformance optimization of wave equation prestack depth migration package on 64-bit cluster computing platform

This paper described in detail the cluster parallel computing strategy of iCluster,a wave equation prestack depth migration package, and performance optimization carried out on 64-bit cluster computing platform. Results of package optimization and performance testing were presented, including operation analysis of wave equation prestack depth migration package and MPI parallel computing, performance comparison between different compilers or compiling parameters, efficiency comparison between different FFT realizations, software operation efficiency and speedup ratio comparison between IA32 and IA64 architectures. Taking the package running on 2. 6GHz Xeon system as benchmark, the performance of the optimized wave equation prestack depth migration package increases about 9 times on 1. 4GHz Itanium2 system. This is owing to 3 major factors: performance increment of 64-bit CPU, optimized compiler, and contribution from DFT, a FFT software in Intel MKL mathlab. Tests and analysis show that for wave equation depth migration computing, kilomega-Ethernet based cluster computer can provide very high speed-up ratio, even if under large scale configuration. To cope with massive data volume and computing, migration to 64-bit computing platform is an inevitable development trend for geophysical computing. Software performance optimization on IA64 platform has shown great potentials.