Implementation of non uniform memory address (NUMA) parallel computation in order to speed up the common reflection surface (CRS) stack optimization process
Fernando Lawrens, Aditya Jiwandono, Rachmat Sule · 2015
The Common Reflection Surface (CRS) Stack is an alternative method of seismic data processing, with an assumption that a reflector consist of a group of reflector segments. As implication, one must determine three wave attributes, i.e. emergence angle, radius curvature of normal-incident-point-wave and radius curvature of Normal wave, before stacking process is performed. The cascaded method is commonly used, which consist of four main steps: automatic CMP-Stack, the plane wave search at zero offset, the hyperbola search at zero offset and is finalized with the optimization of CRS-Stack attributes. Although this method will give better seismic section, especially if the subsurface condition is undulated, this method consumes longer computation time. To solve this problem, NUMA parallel computation scheme is applied in the optimization CRS-Stack step. In order to obtain good NUMA parallel computation performance, some tunning method are implemented, such as load balancing, disk I/O access optimization, and cache optimization. By using all these tunning methods, the maximum speed-up of NUMA parallel computation for CRS-stack is 11.3 times, if 12-core system computer is used.