Mapping to Irregular Torus Topologies and Other Techniques for Petascale Biomolecular Simulation

Area of Science:

• High-performance computing
• Computational science
• Network topology optimization

Background:

• Petascale supercomputers utilize multi-dimensional toroidal networks.
• Topology-agnostic codes face network contention on large node counts (20,000+ nodes).
• Irregular node allocations complicate topology adaptation.

Purpose of the Study:

• To develop topology-aware mapping methods for NAMD on supercomputers.
• To enable large-scale (hundred-million atom) molecular dynamics simulations.
• To optimize domain decomposition for efficient communication.

Main Methods:

• Mapping periodic 3-D spatial decomposition domains to 3-D Cray Gemini and 5-D IBM Blue Gene/Q networks.
• Utilizing topology awareness to mitigate network contention.
• Partitioning node allocations into compact domains for smaller simulations with multiple-copy algorithms.

Main Results:

• Enabled hundred-million atom full machine simulations using NAMD.
• Demonstrated improved performance through topology adaptation.
• Reported performance across multiple supercomputing platforms (NCSA Blue Waters, ORNL Titan, ANL Mira, TACC Stampede, NERSC Edison).

Conclusions:

• Topology-aware mapping is crucial for efficient large-scale simulations on modern supercomputers.
• The presented methods enhance the performance of communication-intensive codes like NAMD.
• Optimized domain decomposition strategies are vital for maximizing supercomputer utilization.