|Name:||Computational Challenges in the Energy Industry – What Seismic Processing, Reservoir Simulation & Quantum Chemistry Have in Common|
|Time:||Wednesday, June 19, 2013
12:30 PM - 1:00 PM
CCL - Congress Center Leipzig
|Breaks:||1:00 PM - 2:15 PM Lunch|
|Speakers:||Detlef Hohl, Shell International Exploration & Production|
|Abstract:||The Oil & Gas industry is one of the largest consumers of HPC capacity worldwide. Applications in the O&G sector fall into two large categories: Data processing and analysis applications (e.g. seismic imaging and inversion for exploration and surveillance, hydrocarbon reservoir modeling and inversion) and engineering design applications (e.g. chemical and other process designs, facilities design). Legacy algorithms and software are often either sequential or rely on low levels of concurrency and trivial, fully independent, parallelism. They must now be replaced with fundamentally new ones that can exploit thousands and, in only a few years, millions of closely coupled possibly heterogeneous cores. This is necessary both for exploiting the continued increase in compute power provided by Moore's exponential growth law and for coping with the “Big Data” challenge posed by the even faster exponential rate at which data volumes and rates grow. In this contribution, we will present three example applications from Shell’s portfolio that leverage massively parallel computing with common approaches and a common programming model but at different levels and with different means: i) Seismic imaging, ii) seismic inversion for reservoir modeling and iii) atomistic simulations with quantum physical Density Functional Theory (DFT).
Compute intensive modern seismic imaging methods employing complex wave propagation methods provide higher resolution and more detailed subsurface images in difficult geological settings. Reverse-Time Migration (RTM) with high-order finite difference stencils is a particularly challenging one to parallelize efficiently. We present how a complete practical scaling proprietary RTM application with all necessary industry features was implemented on a massively parallel GPGPU-based platform in the pursuit of maximum performance.
Seismic inversion is a subsequent higher-level processing step using seismic data for reservoir modeling. Shell’s proprietary probabilistic seismic inversion application PROMISE is based on an underlying adaptive Markov-chain Monte Carlo algorithm to sample Bayes posterior density function for common reservoir properties. The Markov chains are interacting and not trivially independent due to geostatistical constraints, and large-scale parallelization therefore encounters special difficulties.
Shell uses Density Functional Theory in the nanomaterials design and chemistry area for catalyst modeling, the calculation of high-pressure, high-temperature materials properties and in several petrochemical applications. The computational demand for a specific problem is highly system specific, and the most widely used pseudopotential plane wave implementation relies heavily on Fast Fourier Transforms and linear algebra kernels. We present the performance and scaling characteristics for two typical applications of coupled DFT and MD (“Car Parrinello method”).