One of the most promising use cases for “Big Data” is to help advance climate research. At SC11, Reinhard Budich (Max Planck Institute for Meteorology), John Feo (Pacific Northwest National Laboratory) Tobias Weigel (DKRZ) and Per Nyberg (Cray) will co-host the second Climate Knowledge Discovery (CKD) workshop to explore new data-intensive methods. HPCwire talked about this with Budich and Feo.
HPCwire: Climate modeling has always been one of the most compute-hungry applications, and ensemble models have exacerbated this hunger. Talk about the “data tsunami” this has been creating?
Reinhard Budich: Well, to put it simply, we can’t afford the data tsunami! At least not if we try to deal with it in the traditional way, which would be to simply buy as much storage as possible and then try to cope with the shortage in storage space this implies – storage costs go down much more slowly than production costs. We need to work on our methods as intensively as possible, try to generate knowledge from the data we produce as fast as we can, and avoid as much as possible any mid- or long term storage.
Another important aspect which applies to climate science maybe more than to other science areas, is that we need to provide knowledge “downstream,” that is, decision-makers having nothing to do with climate science in the narrow sense of the word need our knowledge for their decisions.
HPCwire: How different are the data analysis challenges of scientific computing and business computing?
John Feo: Just as many scientific domains share fundamental mathematic and scientific processes, many fundamental data analysis technologies can be shared by different domains. Consequently, large market forces will drive and provide sustainable markets for advanced and one-off technologies.
The specifics of science and business analysis are different, but many of the underlying concepts, requirements, algorithms, and data organizations will be similar. For example, the businessman may ask “Who is the thought leader in this social network?” whereas as the scientist may ask “Which process in these metabolic pathways is dominant?” Very different questions, but both might be answered by finding the most central node in the network.
That said, there are significant differences between analyzing scientific and business data. Two big differences are, first, scientific measurements are relatively accurate compared to data on human activities, and second, unlike humans, science does not change behavior overnight. So, a scientific analysis process that is useful today will be useful tomorrow, but the same isn’t necessarily true when humans are involved.
HPCwire: Data-intensive science has been called the fourth paradigm of scientific discovery. What is the current state of data-intensive science in the climate community?
Budich: We have been a data-intensive science since the beginning of numerical climate sciences, I would claim. You can see that, for example, by looking at the balance between investments in compute on the one hand and in storage on the other hand in climate-specific compute centers like NCAR or the DKRZ, Investments are much higher on the storage side than in other centers.
But the methods to create knowledge from these huge amounts of data have also been traditional, since scientists need to evolve their methods rather slowly to make sure they are on safe grounds in terms of reasoning and justifying their results. It also means that these methods are slow, and may be not satisfyingly efficient in all cases.
To increase efficiency, the data processing in climate science needs a more systematic approach, relying much more on metadata not only for provenance, but also for the content of the data. Unfortunately, the development of the according methods is very person-power-intensive and therefore not easy to get funding for. This is one more reason to try to make people aware of the facts.
HPCwire: How do data warehouses or knowledge repositories differ from high-performance analytics or knowledge discovery?
Budich: I’m not sure what this question is aiming for, but to me, knowledge repositories and knowledge discovery differ very much. A data warehouse is not comparable with high-performance analytics. Probably all these terms have a well-defined meaning for people at home in the data-intensive world, but climate scientists are not used to these terms.
They tend to think in terms of flat files and their latest self developed tools to tackle these files, to compare them with the flat files their colleagues have, which are organized and annotated slightly differently and need a lot of work to be made comparable. The IPCC/CMIP process [Intergovernmental Panel on Climate Change/Coupled Model Intercomparison Project] helps a lot in this area, providing the necessary platform to develop metadata generation methods and common standards.
Feo: The two are very different in words and capability. Data warehouses and knowledge repositories refer to the storage, organization, and tagging of data. High-performance analytics or knowledge discovery refers to the discovery of new facts and relationships on the data stored.
For example, a data warehouse may organize a set of Earth-Moon sightings as a data structure of the mass of the Earth and Moon, the distance between them, and the right ascension and declination of the Moon at different times of the year. Knowledge discovery is the realization that the gravitational force implied by the Moon’s orbit is given by Newton’s equation.
HPCwire: Why are new data integration and analysis methods needed for climate science? What’s wrong with the existing methods?
Budich: We do not have enough metadata, and where we have some, they are too cumbersome to produce and so not – or not sufficiently well – integrated into the everyday workflow of the scientist. So everyone is depending to a high degree upon knowledge that is not directly attached to the data, but available only in the form of articles, gray literature or, very often, knowledge stored in the brains of colleagues.
HPCwire: Based on the first Climate Knowledge Discovery Workshop held in Hamburg earlier this year, what new approaches and tools are being considered in climate science?
Budich: According to my understanding we saw that climate science needs to develop more and better methods to produce metadata, and also for use on the content for the data. It needs to develop ontologies and semantics to annotate the data. And it needs to reconsider systematically the way it post-processes the data once they have been produced. To give you a more striking example for this need, at MPI-M we conducted in recent years a Millennium experiment, which produced hundreds of terabytes of data.
We tried to design it in a way that at least provenance data and some quality control metadata were produced during the experiments, as well as subsets giving enough information to make it possible to have a good first analysis of the data without the necessity to keep the raw data on disk. These are normally stored on tape immediately after production. These attempts turned the post-processing into such an extensive process that the production took less time than the post-processing. And this is a trend you can observe in many sites. Again, we need to make the knowledge generation process much more efficient.
HPCwire: How might approaches using network science, machine learning, semantic analysis and graphs improve our understanding of the earth’s climate? What new insights might they be able to provide?
Budich: What I would hope, apart from keeping the cost of the post-processing reasonable, is that we can find much easier than today relationships between different physical, and chemical, biological, and sociological, phenomena. Processes like tele-connections, pattern propagation, etcetera, might be easier to detect than nowadays. I believe that applying methods from other fields does good to any science. We can learn a lot from the efforts made by other communities, such as bioinformatics, where semantics have been applied successfully.
HPCwire: Are any of these approaches being used today?
Budich: This is why we carry out these workshops, which hopefully will turn into a series: We want to find out who uses what, why, and with what success. We try to bring those people together who try to employ semantics and graphs. A group at the Potsdam Institute for Climate Impact Research, for instance, has used graphs successfully for climate knowledge discovery.
The group around Karsten Steinhauser at the University of Minnesota has set up a large project now doing exactly those things we discussed at the workshop. Climate informatics activities point in exactly the same direction. Nevertheless it is a young discipline, and a bit too early to announce any ground-breaking successes influencing climate science methods at large. But we think it has a huge potential, given the right initial investments.
Feo: Many different approaches are being applied to different problems by individual practitioners. The literature is full of papers describing how some technique was applied to some problem to discover something. Analysis tools have been built by individual scientific groups for their specific problem; but a small set of universal techniques has not yet emerged.Also, there is no programming environment for data analysis as there is for scientific computing. What we need is a MATLAB for data analysis.
HPCwire: People in other domains, such as life sciences, tell us there aren’t good tools yet for data integration and analysis on really big data sets of 100 terabytes and beyond. They say Hadoop is useful, but it’s too cumbersome to use on really big data sets and it doesn’t allow you to pose intelligent questions. Can you comment from the perspective of climate science?
Budich: I fully support these statements, and I do not think there are easy ways to design such tools. But it needs to be done for the reasons given earlier. For the time being, you need pre-existing knowledge to reduce the size of the data sets before you begin your investigations, so that you do not need to employ the complete data set of that size for it. For example, our 100 terabytes would probably contain data from atmosphere, ocean and land. If you want to find out about stratospheric interconnections, you might not need land and ocean data for a first attempt, so you could reduce the data size from the beginning.
Feo: Relational databases and Hadoop systems solve some problems very well. It is not the amount of data that makes them fail, but rather the data’s sparsity and the question’s complexity as a function of outer joins. In such cases, graph analytics are more efficient. A single technology won’t solve all problems well. Efficiency and scalability will require the decomposition of analysis workflows into separate components, each executed on a potentially different platform. The integration of smart storage systems, Hadoop clusters, and computer systems able to run large graph algorithms at scale is a daunting challenge facing all communities interested in analyzing large data sets.
HPCwire: The climate community has several data-related efforts, including the Earth System Grid, the Earth System Curator project and the European METAFOR project. How is the Climate Knowledge Discovery different?
Budich: The projects you mention are extremely important for our science field, since they are really putting the importance of metadata and provenance data into the view of many people in our science for the first time. To generate knowledge from data, you need to know that these data exist, where they are and how they were created. These projects care about this knowledge.
But I think our community needs to go one step further by asking, scientifically, what exactly do these data contain, and how do these scientific contents relate to each other? This is the question we try to tackle with the CKD workshops. So far, we have mostly been looking at the data from the outside. The CKD effort is trying to break into this black box and extract knowledge from the contents of the data.
HPCwire: What were some of the main outcomes of the April 2011 Climate Knowledge Discovery Workshop in Hamburg?
Budich: We have seen that data from reanalysis projects can be turned into complex networks, which can help to discover changes in the climate system or show complex dependencies in atmospheric or oceanic circulation patterns. Regional climate predictions can be enhanced by using network-based clusters, which help interpret teleconnections with phenomena like El Niño/La Niña. So there are projects applying these modern data analysis methods to climate science today. But they are not very prevalent, and they still need to prove their efficiency potential.
We found that it would also be beneficial to apply ontologies, representing knowledge of a set of concepts and the relationships among these concepts, to climate data. Annotating complex networks with ontologies improves interpretability using downstream tools. Construction of ontologies about climate data is a huge task that only large community projects can handle. Getting top experts from multiple disciplines together is the crucial success factor for such projects and is one of the most difficult tasks in ontology engineering.
HPCwire: What are the primary objectives for the second workshop at SC11?
Budich: To carry on the process we started in Hamburg, against the background of a potentially much larger audience, hoping that more common projects are initiated or fostered between informatics and climate scientists.
Feo: Eventually, we want to write a high-level requirements document for the technologies that the climate community needs for efficient and useful analysis of large data sets. The document will include many end-user scenarios highlighting clearly the limitations of current technologies and the aspirations of the community.
HPCwire: Who should participate in this workshop?
Budich: What I would like to achieve is community involvement in these activities, in the informatics and the climate communities. So SC11 should be a great venue for this, since people from both worlds will be attending. So if someone needs to find out about new methods to apply to climate data, he or she should attend.
Feo: Participants can register for the workshop at the SC11 website or at the conference hall prior to the start of the workshop. Anyone registered for the SC11 technical program is already registered for the workshop.
Budich: Our next meetings will probably be held alongside or at the European Geophysical Union assembly in April 2012, in Vienna, Austria. We are looking for European colleagues who would like to get involved. If any of your readers are interested, they can get in touch with me ([email protected]) or John ([email protected]).