TOP    Events & Outreach    News & Announcements    Fugaku Tops Graph500 Ranking - Achieving the Highest Evaluation in Graph Analysis Crucial for Big Data Processing -

A collaborative research group consisting of RIKEN, Tokyo Institute of Technology, Fixstars Corporation, Nippon Telegraph and Telephone Corporation (NTT), and Fujitsu Limited has significantly improved the measurement results using the supercomputer Fugaku and achieved the world’s top position in the BFS (Breadth-First Search) category of the international supercomputer performance ranking "Graph500" for large-scale graph analysis. This marks the ninth consecutive win for Fugaku.

The ranking was announced by the Graph500 Committee on their website Graph500The webpage will open in a new tab. in conjunction with the international conference on high-performance computing (HPC), "ISC High Performance 2024," currently being held at the Congress Center Hamburg in Germany and online.

Performance in large-scale graph analysis is a crucial metric for big data analysis, which requires processing vast and complex datasets. The collaborative research group will continue to advance the development of large-scale graph processing technology using Fugaku.

Comment from Technical Scientist Dr. Masahiro Nakao

The first time we submitted the Graph500 results using all of Fugaku's computing units (158,976 units) was in November 2020, achieving a performance of 102,955 GTEPS[1]. By June 2023, we had improved performance to 137,096 GTEPS through the development of a feature that automatically determines optimal algorithm parameters[2]. Furthermore, this time, we enhanced performance to 166,029 GTEPS by developing a feature that removes unnecessary vertices. Additionally, we have significantly reduced the amount of memory used by developing new graph data compression techniques. This demonstrates that even with ultra-large systems like Fugaku, it is possible to achieve both performance improvement and resource efficiency through algorithmic innovations.

Executing this project required meticulous and time-consuming preparation to conduct benchmark measurements using Fugaku almost exclusively. Specifically, this involved developing the new features mentioned above, verifying their operation when combined, optimizing the system at the OS level, and measuring power consumption, among other tasks. We also established a robust system monitoring and support structure to quickly address any hardware failures that might occur during benchmark measurements. This comprehensive setup was made possible not only by the members of RIKEN's Operations and Computer Technologies Division but also with the cooperation of Fujitsu Limited and the Research Organization for Information Science and Technology (RIST). I would like to take this opportunity to express my deepest gratitude to all involved.

Notes
  • [1] TEPS (Traversed Edges Per Second): The number of graph edges traversed per second. GTEPS stands for Giga TEPS, which is the number of TEPS divided by one billion.
  • [2] Results after June 2023 utilized 152,064 units, approximately 4.3% fewer than the total number of computing units, as part of Fugaku was used for other purposes.

Related Link

(May 13, 2024)