Graph500 benchmarking problem includes the mapping of a Kronecker graph with 16 trillion relationships

The Graph500 is a list published twice a year that benchmarks the speed at which a computer performs graph operations.

For the first time ever and using only a fraction of its processing power, an Oak Ridge National Laboratory (ORNL) supercomputer has entered the Graph500 ranking, a list published twice a year that benchmarks the speed at which a computer performs graph operations.

The Summit supercomputer, located at the Oak Ridge Leadership Computing Facility (OLCF)—a US Department of Energy (DOE) Office of Science User Facility at ORNL—placed fourth in the ranking as announced during the 2019 International Conference for High Performance Computing, Networking, Storage, and Analysis (SC19).

“Graphs are present in everything we do,” said Ramki Kannan, team lead for Computational Artificial Intelligence and Machine Learning in the Computer Science and Mathematics Division (CSMD) and one of the researchers working on the benchmark. “Relationships between the elements of transportation systems, the reach certain people can have in social networks, and the steps involved in commercial transactions are all examples of graph problems that are part of our daily lives,” he said.

Graph theory is also applicable to many other areas, such as in the mapping of biological networks of viruses and bacteria. An example of this would be using graphs to help understand the spread of COVID-19 by contact-tracing.

Fewer resources, big results

To enter the ranking, supercomputers must solve a specific type of synthetically generated graph, called a Kronecker graph. The graphs generated by this model satisfy many of the properties found in real-world networks, including networks that are hierarchically organized into communities. In Kronecker graphs, every node is described by a sequence of attributes, and the probability of a relationship between them depends on such characteristics.

The graph Summit had to resolve to rank on Graph500 included over 1 trillion vertices with approximately 16 trillion relationships among them.

ORNL’s Summit used a total of 86,016 CPU cores to solve the problem, which amounts to 45 percent of its CPU computing power, which is a far smaller fraction of its total compute power because GPUs were not used for this benchmark.

In contrast, the supercomputer that ranked first in solving this task—China’s Sunway TaihuLight—used 10,599,680 CPU cores to make the same solution happen.

According to the ORNL Graph500 team—which also included scientists Hao Lu of the National Center for Computational Sciences (NCCS), Kamesh Madduri and Piyush Sao of CSMD, and Michael Matheson and Drew Schmidt of NCCS—the result achieved by Summit is a good example of computing power used efficiently.

“Most teams look at this ranking and think ‘How can we optimize and grow our systems to rank first?’ But we looked at it from a different perspective. We wanted to know how fast we could solve this graph problem by using fewer resources. The fact that we made it in fourth place under that philosophy is a truly remarkable thing,” Kannan said.

Although other teams had been preparing to submit their metrics for years, the ORNL researchers had only 9 months to measure and submit Summit’s performance.

Summit’s place in the ranking is based on a metric called Giga Traversal Edges Per Second (GTEPS), but the system ranks first in other common metrics such as GTEPS per node, Mega TEPS per core, and GTEPS to peak node memory bandwidth, according to observers of the ranking.

Graph500 complements the Top500 list, a biannual ranking of the world’s fastest supercomputers. Summit has placed first in the Top500 since June 2018.

UT-Battelle LLC manages Oak Ridge National Laboratory for DOE’s Office of Science, the single largest supporter of basic research in the physical sciences in the United States. DOE’s Office of Science is working to address some of the most pressing challenges of our time. For more information, visit