Reactive Molecular Dynamics Progress Report Nov 9 2007

  • Slides: 18
Download presentation
Reactive Molecular Dynamics Progress Report Nov. 9, 2007 A. Smirnov and R. A. Carreno-Chavez

Reactive Molecular Dynamics Progress Report Nov. 9, 2007 A. Smirnov and R. A. Carreno-Chavez

Objectives • To develop the capabilities of simulating 105 and more on a workstation.

Objectives • To develop the capabilities of simulating 105 and more on a workstation.

Issues Implementation of segmentation algorithm can affect the speed of execution, such as: •

Issues Implementation of segmentation algorithm can affect the speed of execution, such as: • Memory allocation/deallocation procedure. • Various parameters: cutoff radius

Interaction Acceleration Space segmentation scheme Enables to achieve near linear dependence of execution time

Interaction Acceleration Space segmentation scheme Enables to achieve near linear dependence of execution time on the number of molecules.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are created.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are created and old links are reassigned.

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are

Looped Lined Lists Enable to avoid memory allocations and deallocations. Instead new links are created and old links are reassigned.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Looped Lined Lists Enables more efficient nested looping over neighboring particles.

Work Done • Looped linked lists implemented • Two versions of code are implemented:

Work Done • Looped linked lists implemented • Two versions of code are implemented: GUI-based and batch-job, both using the same solver module. • Performance tests completed

Performance Molecules CPU/timestep, millisec RAM MB / Iterations Simple Segmented 1 e 3 /

Performance Molecules CPU/timestep, millisec RAM MB / Iterations Simple Segmented 1 e 3 / 1000 80 3 1 e 4 / 1000 2800 34 % RAM 1 e 5 / 1000 850 24 0. 3 1 e 6 / 1000 12857 232 2. 9 1 e 7 / 100 9100 2350 28. 8

Performance (in process) Molecules / Iterations CPU/timestep [millisec] 1. 5*cutoff 1 e 3 /

Performance (in process) Molecules / Iterations CPU/timestep [millisec] 1. 5*cutoff 1 e 3 / 1000 2*cutoff 3 1 e 4 / 1000 15/? m 34/? m 39/? m 1 e 5 / 1000 501/? g 850/? g 2447/? g 1 e 6 / 1000 1 e 7 / 100 10914/? g, */: init 6 cut 1 -5. log 12857/? g 9510/2351 m 9100/? g 10914, *6338/2. 6 g: init 6. lo 9030, *6350/4. 9 g: init 7. log

One Million Initial distribution

One Million Initial distribution

One Million 10 time steps

One Million 10 time steps

Conclusions • The efficiency bursting measures are working. • Tens of millions of particles

Conclusions • The efficiency bursting measures are working. • Tens of millions of particles can be accommodated on a single workstation. • Time of execution becomes one day for 10 M particles and 1000 time-steps. • Effect of interaction cutoff should be investigated.