Morgenrot Presents Joint Research Results with JAXA on Optimizing HPC/AI Server Operations at ISC25

“Definitively Proving the Importance of CPU Utilization Visibility and the Energy Efficiency Enabled by GPUs”

Tokyo, Japan - Morgenrot Inc., a provider of infrastructure solutions that support computing power for the generative AI era, announced the outcomes of a joint research project with the Japan Aerospace Exploration Agency (JAXA) on the sound and effective use of HPC/AI servers at the ISC High Performance 2025 international supercomputing conference held in Hamburg, Germany from June 10 to June 13, 2025.

In this joint research, Morgenrot and JAXA verified a technical approach to quantitatively visualize detailed usage patterns of supercomputers on both a per-job and per-node basis, aiming to achieve efficient resource allocation and energy optimization. At the conference, Morgenrot’s Executive Officer and CTO, Hiroshi Ito, presented these results at the booth of Energy Aware Solutions, one of Morgenrot’s technology partners.

Research Theme

Sound and Effective Use of HPC/AI Servers

Background

JAXA’s supercomputing system “JSS3” boasts a high job fulfillment rate of approximately 95%, demonstrating operational stability. However, such high utilization does not necessarily equate to effective use of computing resources or energy savings. The need to visualize execution characteristics and resource consumption on a per-job basis had been identified.

Research Overview and Results

The joint research focused on quantitatively visualizing resource usage per job and node in supercomputing environments, aiming to optimize resource allocation and energy usage. A multi-faceted monitoring approach was implemented using the following technologies:

Technologies Used

• MORGENROT® Arthur - Monitoring software developed by Morgenrot

• EAR (Energy Aware Runtime) - Power consumption measurement software by Energy Aware Solutions

Key Visualization & Analysis Outcomes

• Node-level CPU usage visualization using a monthly calendar view, focusing on 48 nodes of TOKI-SORA, a main system in JSS3.

• Comparative analysis of energy consumption in molecular dynamics simulations using LAMMPS, contrasting configurations with and without GPU use.

Benchmark Results

In a sample molecular dynamics simulation:

• Through monthly analysis of CPU utilization rates, it was confirmed that CPU usage and job fulfillment rates showed different trends over time. This indicates that CPU utilization is also an effective metric for understanding the operational status of computing resources.

• In a molecular dynamics simulation case study, a configuration of 1 CPU + 1 GPU reduced simulation time to approximately one-eighth compared to a 32 CPU configuration, while also reducing power consumption by more than 75%.

• The appropriate utilization of GPU resources was quantitatively demonstrated to contribute to energy savings through reduced computation time.

－ Hiroshi Ito, Chief Technology Officer, Morgenrot commented:

“Properly understanding the operational status of large-scale HPC environments, and running energy-intensive GPUs according to optimized policies, are critical for the sound and efficient management of HPC/AI servers and data centers. We believe the results of this research offer valuable insights, especially toward managing energy consumption in future large-scale HPC infrastructures. We will continue driving efforts toward social implementation of these outcomes.”

Future Outlook

The insights gained from this collaboration will contribute to the advanced operation of HPC/AI servers. Morgenrot will continue enhancing the capabilities of its "MORGENROT Arthur" and promoting strategic partnerships like this one, with the aim of enabling enterprises and research institutions to maximize their GPU resource utilization.

Learn More about MORGENROT Arthur

Inquiry about MORGENROT Arthur