Visualizing Network Traffic to Understand the Performance of Massively Parallel Simulations.

Area of Science:

• High-performance computing
• Computer science
• Network visualization

Background:

• Massively parallel application performance is often limited by communication costs.
• Supercomputer network management is complex due to increasing scale and intricacy.
• Understanding network activity is crucial for optimizing parallel applications.

Purpose of the Study:

• To introduce a visualization tool for exploring packet flow in supercomputer interconnects.
• To aid parallel application developers in diagnosing network-related performance issues.
• To enhance the understanding of network behavior in large-scale computing environments.

Main Methods:

• Development of a visualization tool employing two linked views: a 2D network structure projection and a 3D physical topology view.
• Utilizing visualization to enable detailed exploration of packet flow through hardware interconnects.
• Case study application using the pF3D multi-physics code on an IBM Blue Gene/P system.

Main Results:

• The visualization tool provides valuable insights into network activity.
• The tool facilitates the identification of trends and patterns in network traffic.
• The case study demonstrated the tool's effectiveness in explaining and optimizing application performance.

Conclusions:

• Visualization techniques are effective for understanding and optimizing supercomputer network performance.
• The developed tool offers a practical solution for developers to analyze complex network data.
• Improved network understanding leads to enhanced performance for massively parallel applications.