Barcelona has been absolutely lovely; the weather, the food, the people. I am, sadly, finishing my last day at PRACEdays 2017 with two sessions: an in-depth look at the oil and gas industry’s use of HPC and a panel discussion on bridging the gap between scientific code development and exascale technology.
Henri Calandra of Total SA spoke on the challenges of increased HPC complexity and value delivery for the oil and gas industry.
The main challenge Total and other oil and gas companies are finding is that discoveries of oil deposits are becoming more rare. To stay competitive, they need to first and foremost open new frontiers for oil discovery, but do this while reducing risk and costs.
In the 1980s, seismic data was reviewed in the 2 dimensional space. The 1990’s started development of 3D seismic depth imaging. Continuing into the 2000’s, 3D depth imaging was improved as wave equations were added to the traditional imaging. The 2010’s brought more physics, more accurate images, and more complex processes to visually view the seismic data.
The industry continues to see drastic improvements. A seismic simulation that in 2010 took four weeks to run, in 2016 takes one day. Images have significantly higher resolution and the amount of detail seen in the images enables Total to be more precise in identifying seismic fields and potential hazards in drilling.
If you look closely at the pictures (shown on the slide), you can make out improvements the image quality. Although it may seem slight to our eye, the geoscientist can see the small nuances in the images that help them be more precise, identify hazards, and achieve a better positive acquisition rate.
How did this change over the last 30+ years happen? Improved technology, integrating more advanced technologies, improved processes, more physics, more complex algorithms – basically more HPC.
Using HPC, Total has been able to reduce their risks, become more precise and selective on their explorations, identify potential oil fields faster, and optimize their seismic depth imaging.
What’s next: Opening new frontiers enabled by the better appraisal of potential new opportunities. HPC has enabled seismic depth imaging methods that can do more iterations, more physics, and more complex approximations. Models are larger, there are multiple resolutions, and 4D data. There is interactive processing happening during the drilling and these multi real-time simulations allow adjustments to the drilling, thus improving the success rate of finding oil.
Developing new algorithms is a long-term process and typically last across several generations of supercomputers. Of course, the oil and gas industry is looking forward to exascale. But the future is complex — in the compute in the form of manycore, with accelerators, and heterogeneous systems. Complexity in the storage with the abundance of data and movement between tiers of storage via multiple storage technologies. Complexity in the tools such as OpenCL, CUDA, OpenMP, OpenACC, and compilers. There is a need for standardized tools to hide the hardware complexity and help the users of the HPC systems.
None of this can be addressed without HPC specialists. Application development cannot be done without a strong collaboration between the physicist, scientist, and HPC team. This constant progress will continue to improve the predictions Total relies on for finding productive oil fields.
The second session of the day was a panel moderated by Inma Martinez: titled “Bridging the gap between scientific code development and exascale technology.” Much of the focus was on the software challenges for extreme scale computing faced by the community.
Henri Calandra: Total
Lee Margetts: NAFEMS
Erik Lindahl: PRACE Scientific Steering Committee
Frauke Gräter: Heidelberg Institute for Theoretical Studies
Thomas Skordas, European Commission
This highly anticipated session looked at the gap between hardware, software, and application advances and the role of industry, academia and the European Commission in the development of software for HPC systems.
Thomas Skordas pointed out that driving leadership in exascale is important and it’s about much more than hardware. It’s the next generation code, training, and understanding the opportunities exascale can accomplish.
Frauke Gräter sees data as a significant challenge; the accumulation of more and more data and the analysis of that data. In the end, scientists are looking for insights and research organizations will invest in science.
Parallelizing the algorithms is the key action according to Erik Lindahl. There is too much focus on the exascale machine but algorithms need to be good to make the best use of the hardware. Exascale, expected to happen around 2020, is not expected to be a staple in commercial datacenter until 2035. There is not a supercomputer in the world that does not run open source software, and exascale machines will follow this practice.
Lee Margetts talked of “monster machines” — the large compute clusters in every datacenter. As large vendors adopt artificial intelligence and machine learning, will we see the end of the road for the large “monster” machines? We have very sophisticated algorithms and are using very sophisticated computing. What if this technology that is used in something like oil and gas were used to predict volcanoes or earthquakes — the point being, can technologies be used for more than one science?
Henri Calandra noted that data analytics and storage will become a huge issue. If we move to exascale, we’ll have to deal with thousands of compute nodes and update code for all these machines.
The biggest challenge is the software challenge.
When asked about the new science we will see, the panel had answers that fit their sphere of knowledge. Thomas spoke of brain modeling and self-driving cars. Frauke added genome assembly and new scientific disciplines such as personalized medicine. She says, “To attract young people, we need to marry machine learning and deep learning into HPC.” Erik notes that we have a revolution of data because of accelerators. Data and accelerators enabling genome resource will drive research in this area. Lee spoke of integrating machine learning into manufacturing processes.
As Lee said, “Diversity in funding through the European commission is really important – we need to fund the mavericks as well as the crazy ones.”
My takeaway is that the accomplishment of an exascale machine is not the goal that will drive the technology forward. It’s the analysis of the data. The algorithms. Parallelizing code. There will be some who will buy the exascale machine, but it will be years after it’s available before it’s broadly accepted. As Lee said, “the focus is not the machine, the algorithms or the software, but delivering on the science. Most people in HPC are domain scientists who are trying to solve a problem.”