Following a brief welcome by Conference Chair Samuel Mabakane (CHPC), CHPC Director Happy Sithole called the 13^th annual conference to order on December 1, 2019, at the Birchwood Conference Centre in Kempton Park, Johannesburg.

One of the final in-person events the global HPC community attended before COVID-19 arrived, this conference drew 667 attendees who learned new cybersecurity defense strategies and how to prepare the workforce for the Fourth Industrial Revolution.

SADC HPC Forum

The Southern African Development Community Cyberinfrastructure (SADC-CI) Forum was co-located with the CHPC event for the seventh year. Delegates from SADC member states presented status reports from their countries; what they have accomplished, or challenges they face with cyberinfrastructure, cybersecurity, and related workforce development. In addition to the forum, many of the SADC member-states represented are also involved in the South African CHPC-led African HPC Ecosystems project that features HPC clusters and training programs at a dozen data centers throughout the region. Several Ecosystems workshops and presentations were held during the conference.

Six distinguished keynotes and 26 invited speakers provided a wealth of information, and some are highlighted here.

Thomas Sterling (left), professor of intelligent systems engineering at Indiana University School of Informatics, Computing and Engineering and “Father of Beowulf” delivered a talk titled, “The future of computing will be non-von Neumann.” The essence of his presentation was captured in his July 22, 2020 HPCwire article titled, “A Declaration of Interdependence through Non von Neumann Architecture.”

SC19 General Chair Michela Taufer (right) joined the CHPC National Conference just two weeks after SC19 in Denver, Colorado-U.S. In addition to SC anecdotes, Taufer shared highlights about her research at the University of Tennessee, including a precision agriculture project that exemplifies the emerging trends in data analytics at the intersection of HPC and edge computing.

Taufer explained that with data-intensive workflows, there is a tendency for many to extrapolate to the edge which is costly in terms of time, compute, and storage resources. “With meta data mapping and annotations curated at the moment of collection, it is not necessary to send all data to the HPC system right away,” she said.

Soil moisture data inform fluid dynamics simulations (FDS) that are used to predict flood surges and crop viability. FDS data help land managers make better decisions about a variety of conditions affected by climate change, including when and where to conduct controlled burning to prevent wildfires, when to plant and harvest, etc. “Models created with satellite data simply do not meet expectations,” she said. “Satellite data are course-grained at 27×27 km; ideally, data gathered from sensors at 1x1km granularity is preferred,” she added. This ‘ground truth’ helps managers get the most out of each parcel of land; conditions can vary over a matter of feet (depending on the slope, proximity to water, ground cover, etc.). As the climate warms, our ability to understand global trends is essential if we hope to produce enough food, feed, fiber, and fuel to satisfy a growing population. Taufer’s team has determined the need for hybrid approaches of collecting and analyzing data; using convolutional neural networks and predictive modeling to bridge gaps between satellite and sensor data. Taufer’s research was highlighted in a 2019 eScience paper she co-authored with D. Rorabaugh, et al. titled, “SOMOSPIE: A Modular SOil MOisture SPatial Inference Engine based on Data-Driven Decisions.”

SADC HPC Forum Adviser and University of Chicago Associate Vice President for Research Computing Hakizumwami Birali Runesha shared highlights about the Data Lifecycle Instrument, or “DaLI” (supported by the U.S. National Science Foundation, ACI 1626552).

According to Runesha, in 2016 the global research community began to realize that longstanding methods for managing data were not sustainable as 90 percent of the world’s data, at that time, had been created in the two prior years. Volumes and variety of data are being generated from simulations, instruments and observations at accelerating rates, resulting in extreme challenges with data management and computation. He argues that it is about time the community begins to adopt FAIR data principles (findable, accessible, interoperable, and reusable per Wilkinson, et al). The DaLI instrument is a platform developed for the management of the research data lifecycle that enables researchers to acquire, generate, transfer, process, compute, store, archive, share, and publish data so they are FAIR, and reproducible.

A range of projects representing several domains utilize the DaLI platform, including UChicago’s XROMM Data management Portal (also supported by NSF), The Xenon1T dark matter detector in the Gran Sasso Underground Laboratory in Italy, and the Sign and Gesture Archive (SAGA) video data library for gesture, sign and spoken language data—a collection that includes video, coding, annotations and related metadata which demonstrates DaLI’s versatility. For two years, Runesha’s team has been working on bringing standards around secure research computing and data storage for sensitive research data in HPC environments that follow the NIST1800-171 standard. This platform will leverage Globus High Assurance (HA) to securely facilitate data transfers and incorporate cloud-based resources.

John Gustafson is a professor at the National University of Singapore where his research focuses on next-generation computer arithmetic. His CHPC talk was titled, “Eliminating Weapons of Math Destruction.”

With experience as a chief product architect or principal investigator with a variety of commercial tech companies, including Sun Microsystems, Intel and AMD, Gustafson has devoted 30 years to the development of a better way to represent real numbers on computers. Gustafson, early adopters, and a growing, global research and development team are confident that a posit-based knowledge processing unit, or KPU, is “Every bit as effective for some artificial intelligence (AI) related tasks, and better at some, while more energy and storage efficient than IEEE Floating Point,” he said, and added, “KPUs do for HPC what GPUs did for graphics. Posits are simply more elegant than floats; faster, smaller silicon, and less latency.”

IEEE standard floats were designed for 1980’s engineering limitations, and some things are slow to evolve. But posits’ prospect of higher accuracy, bitwise reproducibility, and portability across architectures, are a few reasons why they are superior to IEEE floats. “For starters, posits are two orders of magnitude more accurate than floats,” said Gustafson, who begged the question, “When optimizing neural networks and machine learning workflows, what if flops aren’t what we need at all? What if we need terapops (posit operations per second) instead?”

With many AI workflows, the Sigmoid Function transition from 0-1 (as a neuron goes from on to off)—gradient minus infinity to infinity—instead of requiring 100 clock cycles, KPUs can do in less than one. KPU development teams have trained neural networks down to 8-bit precision, which reduces the time necessary to train deep learning algorithms. Among advantages, Gustafson added, “Posits don’t have the floating-point drift; a butterfly effect, of sorts, known to happen with large codes and cloud variations.” A more recent breakthrough found that very small, ultra-low power ROM performs post arithmetic without having to decode the sign-exponent-fraction. “It is holistic, and incredibly fast; the technique doesn’t work for floats because they aren’t strongly ordered, like posits,” he added.

So why not register a new IEEE standard for posits? That is exactly what some at IEEE have pressed Gustafson to do, and what he hopes to ultimately accomplish. But first, he would like to gather critical mass of KPU success stories that demonstrate the range of applications they’re best suited for.

Stillwater Computing in the U.S is commercializing the KPU-based system which is built with customized commodity parts. Hardware (FPGA AIC) is available for purchase now, and custom accelerator firmware will be available soon (Posit KPU hardware GA release Q4 2020). This is the industry’s first distributed data flow processor—a software defined parallel processor—that is especially well-suited for next-generation knowledge tasks, such as linear algebra and tensor algorithms, big data processing, deep learning and inference, and optimization.

A video overview of the conference was produced by the South African Council on Scientific and Industrial Research via, “CSIR Connect.”

The CHPC National Conference and its organizers, speakers and delegates wish to thank the generous sponsors who made this event possible: Diamond Sponsor Intel, and others, including Altair, Dell/EMC, Boston Storage, AMD, HP, Schrodinger, and Mellanox. In his opening remarks, Dr. Sithole also thanked CHPC conference planners: Dr. Werner Janse van Rensburg, Kevin Colville, Lesley Fredericks, Noxolo Moyake, Meshack Ndala, and others.

CHPC 2020 Conference: Machine Learning and its Applications in HPC

Registration is open for the 14^th annual CHPC 2020 conference to be held November 30 through December 2, 2020 at the South African CSIR International Conference Center in Pretoria. This year’s conference theme is, ”Machine Learning and its Applications in HPC.” The event will follow COVID-19 guidelines established by the SA National Department of Health. Most content will be presented online, and local provisions will be made for a limited number who prefer to attend in person.

About the Author

HPCwire Contributing Editor Elizabeth Leake is a consultant, correspondent and advocate who serves the global high performance computing (HPC) and data science industries. In 2012, she founded STEM-Trek, a global, grassroots nonprofit organization that supports workforce development opportunities for science, technology, engineering and mathematics (STEM) scholars from underserved regions and underrepresented groups.

As a program director, Leake has mentored hundreds of early-career professionals who are breaking cultural barriers in an effort to accelerate scientific and engineering discoveries. Her multinational programs have specific themes that resonate with global stakeholders, such as food security data science, blockchain for social good, cybersecurity/risk mitigation, and more. As a conference blogger and communicator, her work drew recognition when STEM-Trek received the 2016 and 2017 HPCwire Editors’ Choice Awards for Workforce Diversity Leadership.