Speaking at the French high performance ORAP event earlier this year and later at the SEG, Total’s HPC guru Henri Calandra revealed how GPU-based accelerators and SGI hardware are contributing to its seismic imaging effort. Calandra presented a roadmap of algorithmic complexity and compute horsepower. Today, state of the art for Total is a sub petaflop machine running anisotropic reverse time migration. But before 2020 we can expect hundred petaflop machines running full waveform visco-elastic inversion at higher frequencies.
Total has seen a huge increase in its compute power this year with the arrival of an SGI ICE+ MPP machine which is being extended with Nvidia GPUs. This is currently offering 450 teraflop bandwidth. The SGI ICE+ houses a low latency hypercube interconnect, necessary because, although seismic code is ‘embarrassingly parallel,’ Amdahl’s ‘law’ works the other way in that ‘only 99.99% parallel code is scalable.’
The ICE+ has 16,284 Intel Harpertown cores with 2GB memory/core and 256 NVidia GPUs. A ‘Fat Tree’ hypercube interconnect links all groups and can scale to 32,000 nodes. The interconnect is the key to efficient implementation of the 3D wave equation finite difference solver. According to Calandra, ‘SGI’s hypercube topology offers a scalable single system that is easy to manage.’ The GPU is considered a very promising route to RTM—with a 35x advantage over a single Intel core. As has been noted elsewhere (see out reports from the SEG HPC session and SC09 in this edition), all this comes at the expense of increased programming complexity. Here Total leverages Caps Enterprise HMPP (Oil ITJ April 2008) to overcome the difficulty of adapting its code base to multi core/GPU architectures. More from sgi.com and caps-enterprise.com.
© Oil IT Journal - all rights reserved.