Practical VTI RTM

In this abstract we describe a production implementation of 3D anisotropic VTI RTM using a finite-difference scheme on production cluster. Our end-to-end code is about four times faster than a blade that consists of dual Intel quadcore Harpertown running a similar VTI RTM code, including vectorization. It is interesting to note that in the final implementation disk IO is the bottleneck. This means that faster compute nodes (increased FLOPS per core) will not improve performance. What is needed is more disk bandwidth (>200MB/s per processor) to make it run faster, and/or more computationally intensive versions of RTM to generate better images.