Section: New Results
Optimization of the GSP for Depth Imaging
Participants : Hélène Barucq, Nicolas Le Goff, Roland Martin, Frank Prat.
We have considered the question of optimizing the GSP code by adopting different approaches, as follows. First of all we have speed up the computation of the wave field by changing its representation as a series in which we have used a suitable arrangment of the terms in such a way that at least the 2j -th and the (2j + 1) -th terms are computed in the same time. Hence the time computing is reduced considerably since it is at least divided by two. Secondly, the structure of the code has been parallelized. This work was really necessary because the code, which was developed by Magique-3D has been next installed at Total . For a run at 3D, the initial version required 7 Go of high memory which is prohibitive to go further into the solution of the inverse problem. The parallelization has been optimized by enforcing a better share of the works between each process while the first version consisted in sending the auxiliary computations to the main process which had to work a lot and penalized the computing performances. The new algorithm improves significantly the computing performances of the GSP code. For a velocity model consisting of 400×400×200 points, we have used 8 processes and we got the numerical results after 2 hours, by using 200 Mo of high memory. Using the first version of the code, the main process required 2 Go of high memory alone for a smaller velocity model of 128×128×50 points and it was not possible to consider a model of 400×400×200 points. At present time, some investigations are done to estimate the improvements we realize when comparing the performance of the GSP code with the SPECFEM code.
Very encouraging results are obtained between the two methods. For a 2D three horizontal layered and acoustic medium, the seismograms obtained with both methods show the same travel times and the reflections due to the several interfaces are quite the same. The parallel versions of both codes provide similar computational times for few processes while the GSP code is faster (from a few seconds to a few minutes) than the Specfem code (from few seconds to 1 hour) by one order of magnitude for 20, 50 and 100 processors. At 26Hz, and for a heterogeneous medium model provided by TOTAL, the same kind of mesh (18km by 11km = 15000 by 4000 points) is considered for both methods. The GSP code seems to last few minutes when Specfem2D lasts an hour. The main interfaces can be captured in both cases but Specfem2D is much more accurate than GSP. This seems to be due to the fact that the periodic boundary conditions are not tappered efficiently at the lateral boundaries (except the upper free surface) and that the lateral variations of the material properties are not adequately taken into account by GSP. However, It is extremely important to emphasiwe that GSP seems to be a good candidate for seismic reflection, and a much faster procedure than Specfem2D to discrimintae the main features of the subsurface geological structures. Both programs have been optimized in synchrone and asynchrone mode for High Performance Computing applications. Improved padding techniques will be added in the near future in order to reduce the spurious modes appearing in the seismograms produced by GSP and coming from the lateral periodic boundary conditions.
