Milwaukee’s elegant Pfister Hotel hosted approximately 100 attendees for the 66th HPC User Forum (September 5-7, 2017). In the original home city of Pabst Blue Ribbon and Harley Davidson motorcycles the agenda addressed artificial intelligence, self-driving vehicles and drug repositioning. The printed agenda neglected to suggest that we would actually be served PBR and be accompanied by two HD cruisers complements of the House of Harley local dealership. The Hyperion folks surprised and delighted us further with live Germany music, rouladen, bratwurst and sauerkraut.
Below are a few of my observations from the forum.
Exascale Efforts in the USA and EU
While the European Union’s single digital market strategy is moving forward with a legal and procurement framework, the USA is thinking through metrics. Specifically, it appears the exascale community has abandoned metrics for theoretical peak performance and the percentage utilization of a CPU as key metrics. This may portend that less attention is being given to solver-heavy physics to this new generation of supercomputers. We shall see, but perhaps this is influenced by the fraction of non-recurring engineering costs involved in developing exascale systems in a non-incremental way. Seemingly arbitrary power limitations and an observed pullback on metrics in the US model may correlate with some community observations of under-investment. But perhaps an approach that doesn’t require as much double precision will broaden the market.
Key takeaway: Exascale will emerge in unexpected ways following a retrenchment in HPC metrics used for decades
Artificial Intelligence/Machine Learning/Deep Learning
There were several presentations on AI, machine learning and deep learning ranging from Michael Garris who is co-chair of the NIST ML/AI subcommittee to Maarten Sierhuis (Nissan Research Center in Silicon Valley), Tim Barr (Cray), Arno Kolster (Providentia Worldwide) and Graham Anthony (BioVista). While each of us knows intuitively that we have cognitive assistance in our pockets I was especially interested in the comments that accuracy and speed is often a tradeoff (logical), reduction in error rates occur when 10x more data is used (nice quantification) and pattern detection is very specific to the use case (less intuitive).
Maarten Sierhuis predicted that multiple lane highway scans for automobiles will be available in 2018 and for urban intersections in 2020. Full autonomy is extremely difficult, especially when attempting to identify non-car objects and mimic human decision making in complex situations. High-definition maps aren’t the only missing piece – AI must be present in the cloud.
Arno Kolster was especially targeted in his message that interoperability and workflow management lag pattern detection and algorithm development, concluding that general solutions are a long way off. Algorithm and data formats are very closely linked now – a lockstep that is predictable but inflexible. Ideally, algorithm performance would detect and adjust to system capabilities, along with fluid workflow, integrated message flow, visualization tuned to the customer and well exposed KPIs.
A breath of fresh air came from Graham Anthony who spoke about the pursuit of sustainable healthcare through personalized medicine. The BioVista website calls it ‘drug repositioning’ when HPC drives the ability to more effectively and quickly combine patient and general biomedical data to transform medicine. The key challenge is to get the cost of these services down to fit into standard cost reimbursement codes and the time-frame for doctor use to fit into a 15-minute visit.
Key takeaway: High qualitative impact in a variety of sectors may dwarf the use of today’s research HPC
Innovation Award Winners Paul Morin (The Polar Geospatial Center University of Minnesota) and Leigh Orf (University of Wisconsin at Madison)
Dr. Morin caught my attention when he claimed he could use all possible cycles in the world to analyze geospatial mapping of the poles. Perhaps he said he could use all cycles ever provided but I got rather lost in just the current realm. His plan is to process 80 trillion pictures of the entire arctic at a resolution of two meters. Then repeat – effectively providing time-dependent photography that can track changes in elevation. He uses Blue Waters as a capacity machine today but its scheduler had to be rewritten to handle thousands of job launches. My first thought was that other use cases could benefit from a high-capacity scheduler, such as bioinformatics. Then my second thought was a bit cynical, thinking that most capacity computing proposals wither and die among policy makers who believe our nation’s largest machines should be reserved for capability computing. He is willing to try other technologies. Perhaps the cloud’s existing exascale capacity could help – its current business model notwithstanding.
Dr. Orf’s tornadoes were among the best I’ve seen. He uses 15-meter resolutions knowing that doubling the resolution needs 10x more compute power and 8x more memory. His biggest bottleneck is I/O because of the frequency of time-step saving. His biggest achievement may be that he effectively created a new file system by allocating one core per node to build an HDF5 file. His key desire is to issue probabilistic forecast warnings by looking at radar as storms are forming and differentiating between predictions of EF1 and EF5.
Key takeaway: These researchers are heroes interested in impact that transcends both basic and applied research. So why is ready access to huge, highly-tuned capacity computing so impossible?
NCSA/Hyperion Industry Study
I have some unique perspectives on the report released August 22, 2017, by NCSA since I was the initial PI for the NSF award. This work complements a 2012 NCSA survey that I completed on the impact on scientific discovery using simulation-based engineering and science and a 2015 book on industrial applications of HPC that captures 40 contributions from eleven countries from HPC centers that engage closely with industry. I’ll share my observations on this study for a separate article.
As for my takeaways from beyond the printed agenda I would simply observe that the dinner speaker from the Pabst Museum was informative and inspirational. Captain Pabst married into a brewing family and became an unlikely company president given his first love as a steamer captain on Lake Michigan. Pabst Brewing Company ultimately grew to become the world’s largest brewery, selling 15.6 million barrels of beer in 1978. I highly recommend a tour of the 22,000 square-foot 1890s-era Pabst mansion on Milwaukee’s original Grand Avenue. It offers deep learning of a different kind.
About the Author
Merle Giles is currently CEO of Moonshot Research LLC. He directed NCSA’s Private Sector Program for ten years.