As we are heading towards extreme-scale HPC coupled with data intensive analytics like machine learning, the necessary integration of big data and HPC is a current hot topic of research that is, as Rashid Mehmood notes, “still in its infancy”.[i] Mehmood is the Research Professor of Big Data Systems and the Director for Research, Training and Consultancy at the High Performance Computing Centre, King Abdulaziz University (KAU) in Saudi Arabia.
A driving force to incorporate big data into HPC, Mehmood observed in his presentation at the first Middle East meeting of the Intel Extreme Performance Users Group at KAUST (King Abdullah University of Science and Technology) that, “Increasingly more data is being produced by scientific experiments from areas such as bioscience, physics, and climate, and therefore, HPC needs to adopt data-driven paradigms.”
Mehmood is not alone in his observation. Over the past four years the Big Data and Exascale Computing (BDEC) project organized a series of five international workshops that explored ways in which new forms of data-centric discovery might be integrated with the established, simulation-centric paradigm of the high performance computing (HPC) community. [ii]
Looking toward the future of cyberinfrastructure for science and engineering, BDEC produced a whitepaper that highlights the critical problems involved in the diverse patterns of when, where, and how data is to be produced, transformed, shared, and analyzed. We view the main points of the BDEC whitepaper in light of current efforts in the HPC community, such as the Wrangler data analytics supercomputer at the Texas Advanced Computing Center (TACC), the Argonne lab-wide data service, and data management efforts at NERSC.
Understanding the bifurcation between the two software ecosystems
Comparing HPC to High-end Data Analysis (HDA) people use a different vernacular and focus on different key concepts.
Those who work in HDA speak of the 4Vs of big data which are: volume (scale of the data), velocity (speed of intake particularly with streaming data), variety (different forms of data), and veracity (the uncertainty of the data). Meanwhile HPC scientists tend to speak in terms of performance, scaling, and the power efficiency of a computation.
This difference in focus is reflected in the representative big data and HPC software stacks as summarized by Reed and Dongarra. [iii]
The BDEC committee attributes this bifurcation in software stacks to the natural evolution of the two separate communities (e.g. scientists vs. academics and commercial software developers) working to address their separate problem domains.
Working over the past four decades, the HPC scientific community focused in increasing the ability of scientists to model and simulate using numerical models. Meanwhile, the data analytics ecosystem has been rapidly developed over the past fifteen to process the torrents of business, industrial process, and social network data now being generated by consumer devices and the burgeoning Internet of Things. For the most part, the data analytics software ecosystem was not developed by the scientific computing community as they work to adapt to the massive increases in data that is being produced by new instruments and sensor systems.
Both paradigms are collapsing from the data deluge
The BDEC whitepaper observes that both HPC and HDA workflows are eroding, if not collapsing under the onslaught of an apparently ever-growing data deluge[iv]. The future, they advocate, is to stop thinking in terms of a “big machine” but rather focus on the many unsolved problems surrounding wide-area, multi-stage workflows.
Such workflows represent a remarkable reversal in thinking about data, where the issue is not connecting the edge via “the last mile”. Instead, these workflows present a multidimensional “first mile problem” that is not currently addressed by either cloud-based HDA or on-premises based HPC solutions. The BDEC whitepaper authors state, “Arguably, the main cyberinfrastructure challenge of the Big Data era is to adapt or replace the legacy paradigm with a new type of distributed services platform (DSP), one that combines computing, communication, and buffer/storage resources in a data processing network that is far more integrated than anything hitherto available”.
Current efforts to address the HPC data challenge
Both vendors and the HPC community are working to address the big data challenge in a variety of ways – especially with the general acceptance of AI and its dependence on large data sets. One example is how Intel is working with the ecosystems to develop a reference platform to guide the development of future infrastructure to leverage the growing data and the power of HPC supercomputers.
Academic projects such as the ones listed below have shown remarkable success and have provided valuable “lessons learned” to the HPC community.
The Argonne lab-wide data service
At Argonne National Laboratory, researchers are preparing for the exascale era by exploring ways to improve collaboration, eliminate barriers to using next-generation systems like Aurora, and facilitate seamless workflows.
In one example, a team at Argonne’s Data Science and Learning Division is developing a lab-wide service that will make it easier to access, share, analyze, and reuse large-scale datasets.
“Our motivation,” Ian Foster (Argonne Data Science and Learning Division Director and Distinguished Fellow) explains, “is to create increasingly rich data services so people don’t just come to the ALCF for simulation but for simulation and data-centric activities.” Foster also observes that, “It’s becoming increasingly impractical for supercomputing facility users to move their data to their home institution’s system for analysis”.
Aimed at enabling more effective data capture and discovery, as well as association of machine learning models with data collections for improved reproducibility and simpler deployment at scale, the service leverages well-known tools including Globus for research data management and the Argonne’s Petrel storage system.
The Texas Advanced Computing Center (TACC) Wrangler supercomputer is the first of its kind and the most powerful data analysis system allocated in the Extreme Science and Engineering Discovery Environment (XSEDE). [v]
The system is designed to support HDA in an HPC environment. It provides around a half a petabyte (0.5 PB) high speed flash storage system that can be used to handle data analysis and processing workflows not practical on other systems. TACC notes, “Wrangler’s unique architecture handles the many aspects of the volume, velocity, and variety that can make digital data research difficult to handle on standard high performance systems”. [vi]
Very importantly, the system is dynamically provisioned by the users to handle different data workflows, including databases (both relational database systems and the newer noSQL style databases), Hadoop/HDFS based workflows (including MapReduce and Spark), and more custom workflows leveraging the flash-based parallel file system.
The success of Wrangler can be seen in the several hundred projects in the TACC Wrangler Data Portal that range from Advanced 3D Microscopy to a Zebrafish map that identifies recessive mutations in Zebrafish.
Recent research shows TACC at the forefront of deep-learning with a new algorithm that speeds training on the Stampede 2 supercomputer so it only take 11 minutes to train ImageNet.
Addressing the challenge of the two paradigm splits
The end goal, according to the BDEC whitepaper is to, “define a new, common and open Distributed Services Platform (DSP), one that offers programmable access to shared processing, storage and communication resources, and that can serve as a universal foundation for the component interoperability that novel services and applications will require”.[vii]
The following schematic reflects this vision.
As the future recipient of the nation’s first exascale supercomputer, Argonne National Laboratory is particularly vested in taking a leadership role in testing the wide-area, multi-stage workflows recommended by the BDEC whitepaper. The Argonne Petrel project appears to be a good start. In particular, the ability to ingest data from instruments and simulation as well as collaborate and publish data regardless of the size of the data set are particularly valuable. An experimental effort using Kubernetes containers may help to democratize the software stack as well as data by providing HDA and HPC convergence through applications containers. The ability to dynamically provision the machine is a “lesson learned” from TACC.
It makes sense to cross-fertilize as much as possible between the HDA and HPC software stacks for big data while looking ahead to an even bigger data future. There is much to be gained as we know that big data is here to stay and exascale supercomputers will certainly play an essential role in helping scientists use this data to make ground-breaking scientific discoveries.
Rob Farber is a global technology consultant and author with an extensive background in HPC and in developing machine learning technology that he applies at national labs and commercial organizations. Rob can be reached at [email protected]
[i] Usman S., Mehmood R., Katib I. (2018) Big Data and HPC Convergence: The Cutting Edge and Outlook. In: Mehmood R., Bhaduri B., Katib I., Chlamtac I. (eds) Smart Societies, Infrastructure, Technologies and Applications. SCITA 2017. Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, vol 224, pp. 11–26. Springer, Cham. DOI: https://doi.org/10.1007/978-3-319-94180-6_4
[ii] See http://www.exascale.org/bdec/ and specifically the report which can be downloaded here: http://www.exascale.org/bdec/sites/www.exascale.org.bdec/files/whitepapers/bdec_pathways.pdf.
[iii] The freely available BDEC whitepaper credits Reed and Dongarra citing Daniel A. Reed and Jack Dongarra. Exascale computing and big data. Commun. ACM, 58(7):56–68, June 2015. ISSN 0001-0782. doi: 10.1145/2699414. URL http://doi.acm.org/10.1145/2699414.
[viii] Usman S., Mehmood R., Katib I. (2018) Big Data and HPC Convergence: The Cutting Edge and Outlook. In: Mehmood R., Bhaduri B., Katib I., Chlamtac I. (eds) Smart Societies, Infrastructure, Technologies and Applications. SCITA 2017. Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, vol 224, pp. 11–26. Springer, Cham. DOI: https://doi.org/10.1007/978-3-319-94180-6_4
[ix] Sardar Usman, Rashid Mehmood and Iyad Katib HPC & Big Data Convergence: The Cutting Edge & Outlook Poster presented at the first Middle East meeting of the Intel Extreme Performance Users Group, Intel IXPUG, KAUST, April 2018 https://epostersonline.com/ixpug-me2018/node/19