The 2017 Rice University Oil & Gas HPC Conference: 3 Key Takeaways from Day 1

UNIVA UD's picture

Ian Lumb, Solutions Architect, Univa

Day 1 of the 2017 Rice University Oil and Gas HPC Conference wrapped up yesterday. Here are 3 takeaways I extracted (in no specific order):

  • Seismics still a Grand Challenge.  Reverse Time Migration (RTM) is performance-challenged. From brute force, to innovative algorithms involving GPUs (or possibly Apache Spark), these challenges can be addressed. And although RTM remains an essential step in processing seismic data, there’s an even-more performance-challenging step that many organizations are attempting to grapple with. Full Waveform Inversion (FWI) is inherently demanding as it requires solution of The Wave Equation. For realistic wave propagation physics and complex geologies, iterative solution of this equation on petroleum-exploration-scale datasets remains … well, extremely performance challenging. In a keynote presentation by John Eastwood of ExxonMobil, concern was expressed that the requirements of FWI and trends in next-gen HPC systems architecture, were misaligned – so much so, that he viewed the two as evolving in opposite directions. In other words, it’ll likely take innovative algorithms (that exploit GPUs?), for real progress to be made.
  • Shifting times for HPC? Self-proclaimed “Data Girl” (aka. Intel’s Debra Goldfarb) shared numerous data points – almost too many for us to internalize in real time … and that’s a good thing! In one example, she identified a potential shift in HPC from an explicit to implicit mode. The explicit mode refers to the way in which HPC is deployed today, with many organizations in the oil and gas industry making use of on-premise supercomputers for their applications. In the emerging implicit mode, HPC takes place in the cloud and is more service oriented. This is a trend that we’ll continue to track, as our Navops Command for Kubernetes-based container clusters could be a key enabler of such a cloud-native, micro-services based shift.
  • Deep Learning hitting the wall.  Although Goldfarb, Univa and others alluded to Machine Learning in their presentations, Fraunhofer ITWM’s Janis Keuper addressed the topic of scaling algorithms directly in his plenary presentation. In highlighting results from a technical paper he presented at SC16, Keuper made it clear that there are limits to scaling Deep Learning. Briefly, widely used algorithms (e.g., Stochastic Gradient Descent, SGD) require certain steps be performed sequentially, and are therefore exceedingly difficult to parallelize. And anyone familiar with Amdahl’s Law knows that the need for sequential steps is rate limiting for parallel computing. Whereas brute-force computing on GPUs has allowed us to progress to impressive performance results, we’ve definitely reached the point where the GPUs can’t be busy enough(!). In other words, it’s a software issue, not a hardware issue. Keuper assures us that many researchers are working on the development of improved algorithms that seek to reduce, hide or (even better) eliminate the need for sequential steps; even more assuring, he holds a high degree of optimism that this algorithmic breakthrough and subsequent implementation in software will occur over the next 12 months.

Rice University president David Leebron joined us for the day-end reception marking the 10th anniversary of the Oil & Gas HPC Conference. In part, he shared that events like this were perfectly aligned with one of his missions – namely, to re-invent the city of Houston as a tech hub. The collaboration and community that characterizes the annual Conference was held up as exemplary in doing just that – helping to place a different profile on one of the US’s largest urban centers.

Day 2 of the 2017 Conference is now underway. Please follow us on Twitter for real-time takeaways from the event. Tomorrow, we’ll share our quick takes for Day 2.

Copy this html code to your website/blog to embed this press release.


Post new comment

2 + 15 =

To prevent automated spam submissions leave this field empty.