A research team of scientists from Georgia Tech, Oak Ridge National Laboratory (ORNL), University of California San Francisco (UCSF), and AMD have developed the first ever algorithm to run over one exaflop on a graph artificial intelligence demonstration.
Called COAST (Exascale Communication-Optimized All-Pairs Shortest Path), the algorithm could help future researchers solve medicine’s most challenging mysteries by revealing hidden connections across large bodies of research. This would revolutionize medical research by developing better treatment plans, creating more effective drugs, and improving efficiency of resource allocations.
As a result of their research, the Association of Computing Machinery (ACM) nominated the team for the 2022 Gordon Bell Prize, commonly referred to as the Nobel Prize of supercomputing.
“Within HPC (high-performance computing), COAST shows that classical ideas in algorithms and performance engineering are still critical to scaling to big machines, like ORNL’s Frontier,” said Rich Vuduc, a professor with Georgia Tech’s School of Computational Science and Engineering (CSE). “More importantly, we now have a new capability that may help speed up the search for new ideas in the biomedical domain.”
The magnitude of computing power needed to complete COAST’s computations requires access to supercomputers. Therefore, the team tested COAST on ORNL’s Frontier, the world’s first exascale supercomputer, and Summit.
With computing speeds of 1.102 exaflops per second, Frontier jumped to first place of the TOP500 list of fastest supercomputers when it became operational in May 2022. Summit currently ranks fourth on the list.
An exaflop is a measure of performance for a supercomputer that can calculate one quintillion floating-point operations per second. If a person calculated a simple math problem in one second, it would take that person about 30 billion years to complete one quintillion operations.
In their study, the team tested COAST twice on Frontier and Summit. On one test, the supercomputers carried COAST at a speed of 1.004 exaflops per second for 11.7 minutes. On the other, COAST reached a speed of 1.008 exaflops while completing calculations in 15.2 minutes.
The team applied COAST to UCSF’s Scalable Precision Medicine Open Knowledge Engine (SPOKE). SPOKE is an evolving biomedical knowledge network that integrates over 40 data sources into a graph with more than 50 million vertices and more than 100 million edges.
The SPOKE graph database facilitates discovery of new knowledge by enabling users to explore the graph’s structure and run analytical queries, like COAST, against it. The interorganizational team cited that past researchers used SPOKE to better understand Covid-19 and find treatments for the virus.
Essentially, COAST wades through these millions of vertices and edges to discover connections across the massive medical data network. On its 1.004 exaflop trial, COAST computed on a SPOKE graph segment of 7.06 million vertices drawing from 18 million publications.
As a result, this is the first scientific study on the integration of SPOKE with publication information through use of an all-pairs, shortest path algorithm.
The project also carries sentimental meaning as it brought together a small but growing community of computational scientists.
Joining Vuduc in the research group from the School of CSE is Ph.D. student Vijay Thakkar. ORNL research scientists Ramakrishnan Kannan (Ph.D. CS 16) and Piyush Sao (Ph.D. CSE 18) led the study and also have ties to the School of CSE as alumni and Vuduc’s former students.
“In CSE, we pride ourselves on being a family,” Vuduc said. “This study is a multi-generational collaboration that speaks truth that we stay close, even past graduation.”
Though it would take the human brain hundreds of billions of years to calculate the same number of computations as COAST, the HPC community won’t have to wait long to learn the winner of this year’s Gordon Bell Prize.
ACM will announce the winner Nov. 17 at International Conference for High Performance Computing, Networking, Storage, and Analysis. Commonly called SC22, the conference is held this year in Dallas Nov. 13-18.
“The 2005 strategic plan for the School of CSE predicted that HPC would be critical to data mining and analysis problems,” said Vuduc. “This nomination is the culmination of hard work done toward realizing that vision.”