Parallel simulation of Carbon Nanotube based composites

1
Parallel simulation of Carbon Nanotube based
composites

• J.Kolhe1, U. Chandra2, S. Namilae3, A.
  Srinivasan1, and N. Chandra3
• Computer Science, Florida State University
• Computer Science, FAMU
• Mechanical Engineering, FSU-FAMU

Simulated using Molecular Dynamics (MD)
www.cs.fsu.edu/asriniva Partly funded by NSF-NER
program and FSU-CRC PG program
2
Outline

• Background
• Sequential algorithm
• Spatial parallelization
• Temporal parallelization
• Conclusions and future work

3
Background

• Molecular dynamics
• In each time step, forces of atoms on each other
  modeled using some potential
• Brenners potential is very accurate for CNTs
• After force is computed, update positions
• Repeat for desired number of time steps
• Time steps size 10 15 seconds, due to physical
  and numerical considerations
• Desired time range is much larger
• A million time steps are required to reach 10-9 s

4
Brenner potential features

• Molecular dynamics, using Brenners potential
• Short-range, multi-body interactions
• Neighbors can change dynamically during the
  course of the simulation
• Computational scheme
• Find force on each particle due to interactions
  with close neighbors
• Update position and velocity of each atom

5
Force computations

• Pair interactions

• Bond angles

• Dihedral

• Multibody

6
Sequential algorithm

• Initialization
• Loop over n time steps
• - Determine the neighbor list if necessary
• - For each atom
• Compute two-body forces
• Compute three-body forces
• Compute four-body forces
• Determine new positions, velocities, and its
  higher derivatives
• - Compute potential and kinetic energies
• - Apply thermostat That is, change velocities to
  keep temperature constant
• endloop
• Figure 2. Sequential Algorithm for MD simulation
  of CNT Interface

7
Optimizations

• Neighbor list computation
• Determine the neighbors of each atom
• Brenner O(n2) computation
• Cell based O(n) in practice
• Sorted projection based O(n) in practice

Cell based
Sorted projection based
8
Neighbor list time
9
Spatial parallelization

• Based on sorted projections
• Replicate force computations on boundary atoms,
  to avoid extra communication step
• Non-blocking MPI communication mode
• Thermostat with local atoms

10
Effect of optimizations

• Improvements in performance, relative to the
  original scheme. 1. Time for neighbor list
  computation on 3200 atoms, compared with
  Brenner's algorithm. The time relative to a
  cell-based algorithm is a little less than 0.5.
  2. Time for packing and unpacking, using the
  cache-aware scheme, compared with a
  non-cache-aware scheme. 3. Time using
  non-blocking sends and receives, compared with
  blocking calls

11
Speedup

• Speedup results with 10000 atoms on an IBM SP3

12
Temporal parallelization

• Each processor simulates a different time
  interval
• Based on predict-verify phases

Ideal
3.2 GHz Xeon/Linux/100Mbps Ethernet cluster
IBM SP3
Speedup on a 1000-atom CNT
13
Conclusions and future work

• Efficient parallelization is possible in the
  range of 300-500 atoms per processor, even with
  an optimized sequential code
• Thermostating with local atoms improves
  scalability
• Redundant computation at boundary atoms reduces
  the efficiency
• It may be worthwhile to perform an extra
  communication instead
• The type of potential used in MD influences the
  parallelization scheme
• Time parallelization results are promising