Least Squares Reverse Time Migration on GPUs - Balancing IO and Computation

Inverse imaging in 3D creates huge computational challenges, particularly in terms of data handling (IO) and the number of floating point operations required. Adapting inverse imaging, such as Least Squares Reverse Time Migration, to a GPU system can accelerate certain aspects of the computation however it inevitably leads to an IO dominated scheme with serious data movement bottlenecks. There are several methods, such as restricting disk accesses, asynchronously moving wavefields, saturating CPU memories and using random boundaries that can reduce the amount of compute time spent on IO. We present, through the context of LSRTM in 3D, how it is possible to map the algorithm and adapt certain aspects to create a processing scheme that is balanced between wavefield computation, IO and imaging.