A Hybrid in Situ Approach for Cost Efficient Image Database Generation.

Area of Science:

• High-performance computing
• Scientific visualization
• Computational science

Background:

• In situ visualization is crucial for analyzing large-scale simulations.
• Current methods face challenges like idle time and resource inefficiency.
• Optimizing visualization processes on supercomputers is essential for cost savings.

Purpose of the Study:

• To present a novel hybrid approach for in situ image database generation.
• To reduce computational costs on supercomputers through efficient visualization.
• To minimize idle time in extreme-scale computing environments.

Main Methods:

• A hybrid technique combining inline and in transit visualization.
• Dynamic distribution of visualization tasks between simulation and visualization nodes.
• Probing-based estimation of rendering costs to optimize task allocation.

Main Results:

• The hybrid approach demonstrated superiority over traditional inline and in transit methods.
• Significant reduction in idle time due to variability, scalability, overhead, and rightsizing inefficiencies.
• Effective performance across various configurations, including large-scale simulations exceeding 19,000 cores.

Conclusions:

• The proposed hybrid in situ visualization technique offers substantial cost savings on supercomputers.
• This method can lead to significant reclamation of computing cycles as in situ visualization becomes more widespread.
• The approach effectively addresses fundamental inefficiencies in large-scale visualization workflows.