Third EAGE Workshop on High Performance Computing for Upstream

In this talk, I will discuss our approach to the formulation and the performance optimization of finite difference methods for PDEs arising in FWI. Our framework consists of a stack of domain specific languages and optimizing compilers. The mathematical specification of a finite difference method is translated by a compiler, Devito, into C code, applying a sophisticated sequence of transformations. These include standard loop transformations, such as blocking and vectorization, as well as symbolic manipulations to reduce the unusually high arithmetic intensity of the stencils arising in forward and adjoint operators. These include common subexpressions elimination, factorization, code motion and approximation of transient functions. I will show the impact of these transformations on standard Intel Xeon architectures as well as on Intel Knights Landing. Compelling evidence points in the direction that our stencil kernels are significantly bound by the L1 cache. I will conclude discussing future challenges and goals of our work.

The extension of full waveform inversion to multi-parameter is the straightforward step towards achieving its full capability, i.e. extracting high resolution information about the sub-surface properties of the Earth based on seismic data. Elastic and visco-elastic FWI begin to be computationally affordable at large scale using the latest HPC technologies and thus to introduce the option to invert for more parameters such as shear wave velocity and attenuation. FWI code design is a key point to allow geophysicists to test new inversion parameterization with minimal code developments. In this study we present a framework based on the abstraction of the parameters inverted. We describe how our framework allows us to use different wave equation propagators with different optimization schemes in a transparent and flexible way. We also show how using a symbolic expression parser to automatically compute required chain rules can greatly reduce the effort needed to implement and test new parameterizations while preserving computational efficiency of the code. We finally present how this symbolic parser can be used to easily implement master-slave relationships between inverted and passive parameters. We then present some FWI tests using different types of parameterizations on the SEAM synthetic model.

The new container technology Singularity focuses on the compute portability, this work covers the creation of portable containers using multiple resources, and showing how simple is to deploy a complex MPI+CUDA+Infiniband seismic application across multiple supercomputers with only a regular user account.

In our work, we consider Discontinuous Galerkin methods (DGm) to solve 3D elastic equations in frequency domain. In 3D, the large size of the linear system represents a challenge even with the use of High Performance Computing (HPC).

Our solution is to develop a new class of DGm, a hybridizable Discontinuous Galerkin method (HDGm). It consists of expressing the unknowns of the initial problem in terms of the trace of the numerical solution on each face of the mesh cells.

We, first, compared the computational performances, in terms of CPU time and memory consumption, of the HDGm with the ones of a classical DGm and of a classical finite elements method (FEm).

Then, since the global matrix of HDGm is very sparse, we also compared performances of the two solvers when solving 3D elastic wave propagation over HDGms: a parallel sparse direct solver MUMPS ( MUltifrontal Massively Parallel sparse direct Solver) and a hybrid solver MaPHyS (Massively Parallel Hybrid Solver) which combines direct and iterative methods.

Wavefield modeling is necessary in modern seismic imaging applications such as reverse time migration and full-waveform inversion. When the medium has complex structures such as salt bodies or carbonate reservoirs finite-difference methods (FDM) for wavefield simulation (extrapolation) are typically used to handle those cases. FDM allows us to simulate a multitude of realistic wave phenomena, but in some cases it makes our applications computationally intensive. When large numbers of sources and receivers are considered, a large number of wavefield extrapolations in the process of inversion is executed. To accelerate the 3-D wavefield simulation in elastic orthorhombic anisotropic media we rely on GPU technology. With the OpenAcc PGI compiler we create a pool of automatically managed memory that is shared between the CPU and GPU, thus achieving data management with minimal code modifications. We collapse the tightly nested loops used for velocity and stress updates which allows us to improve the execution time of the whole code by about ten percent. We report a performance speedup as we compare to a 16 core dual socket Haswell server of 1.15X on a K80 GPU and 2.32X when using the Pascal Tesla P100 GPU.

We examine the performance of a 3D finite difference seismic reverse time migration (RTM) kernel on systems with Intel® Xeon® processors and Intel® Xeon Phi™ x200 processors. Using an identical tuned RTM TTI source code, an system with one Intel Xeon Phi 7250 processor outperforms a system with two-socket Intel Xeon 2697v4 processors by a factor of 1.5.

Large memory overhead of LU decomposition during the factorization stage of the truncated SPIKE algorithm represents a common bottleneck. For large banded diagonally-dominant linear systems, their structure can be leveraged to avoid said LU decompositions. Specifically, Neumann series can be used to approximate these inverse, with smaller memory consumption complexity thus avoiding the excessive growth of the reduced linear system. Convergence of the Neumann series is ensured via a simple scaling of the input matrix. We present results showing the achieved accuracy providing bounds for memory and FLOP count.

a new and novel dynamic approach for writers selection by mapping writing tasks dynamically and adaptively to MPI processes at runtime. The objective of this method is not to deduce the optimal number of writers that should be used. Instead, the method determines and recommends better mapping that enhances the balance of I/O workload across hosts and maximize the overall I/O performance during the lifespan of the simulation run.

Oil & Gas companies are facing an increasing challenge on extracting performances from parallel IO filesystems for check-pointing and writing results. Those efforts will increasingly become an issue due to the complexity introduced by the integration of more sophisticated seismic imaging equations in order to sustain the need of more detailed images to face the next exploration and production challenges. As current and future generations High Performance Computing (HPC) systems are evolving toward an increase computing power, IO bandwidth is remaining relatively constant. Henceforth, the need to manage IO efficiently at scale will become a strict requirement.

We will present an implementation and optimization study of ADIOS in the context of seismic imaging. We will exhibit a performance study made on a proxy application and on a RTM for several computing kernels for different HPC systems in the context of check-pointing. We will show that using ADIOS can provide good performances, manage more efficiently IO and reduce meta-data contention. We will highlight that advanced IO libraries such as ADIOS provide the opportunity to overcome the challenges of maintaining performances with high level IO interfaces at scale.

HPC and Big Data are progressing towards new challenges in an era of explosion of machine data. Computing power is a given, data management is the challenge. Memory-driven computing is going to change dramatically the systems architectures, and HPE is working on the building blocks that will enable this new paradigma of computing. The talk will present the innovative technologies and the roadmap that will bring to that new ecosystem.

An offshore petroleum is the main focus of the current research with Big Data and high performance computing motivations in the study area. We undertake a joint exploration study focusing on Romanian offshore Black Sea (Western) basin, using volumes and varieties of datasets in a Big Data scale. The exploration datasets include more than 175 2D seismic lines, 30 km2 of 3D seismic data, information on more than 8 exploratory drilled wells including check-shot and VSP data including existing petrophysical and production data.

Big Data opportunities are explored in the current upstream business research by proposing data modelling, visualization and data interpretation schemes. In spite of the data quality issues in the study area, several isochrones, isochores and other geological information are integrated and made based on which depositional models are drawn for risk minimizing the exploration and field development plans. The conclusions are based on structural, strati-structural interpretation, organic geochemistry and identification of new opportunity areas. Several data models, visualization and interpretation artefacts can handle the volumes and varieties in Big Data scale minimizing the risk involved in the upstream business in the investigating area. Several new opportunities are identified in the shelf, slope and deep marine areas.

The Sobel filter is a discrete differentiation operator widely used in seismic image processing algorithms for automatic fault detection and extraction. The filter approximates the local gradient by combining derivatives of the amplitude between neighboring traces along the x, y, and z directions. However, 3D-Sobel Seismic Fault Detection algorithm runs very slowly and is computationally intensive, even with dual Sandy Bridge Xeon 8-core 2.6 GHz CPU, it requires more than one minute to process a 560×390×320 seismic volume data.

Herein we present a multiple parallel computing designs leveraging shared memory (OpnMp), distributed memory (MPI) and many-core Graphical processing Unit (GPU) to reduce total execution time of 3D-Soble Seismic Fault Detection; experiments demonstrate a significant speed up over serial CPU version. For a 250 MB poststack seismic volume, the speedup using the parallel algorithm with MPI was six times faster than the serial version of the algorithm.

We first introduce the new 3D-Sobel seismic fault detection algorithm and the parallel implementations, and then show the experimental results.