Status report on the large-scale long-run simulation on the grid - Hybrid QM/MD simulation -

1
Status report on thelarge-scale long-run
simulation on the grid- Hybrid QM/MD simulation -

• Grid Technology Research Center
• AIST
• Hiroshi Takemiya, Yoshio Tanaka

2
Goal of the experiment

• To verify the effectiveness of our programming
  approach for large-scale long-run grid
  applications
• Flexibility
• Robustness
• efficiency
• Friction simulation
• Nano-scale prober moves on the Si substrate
• Requiring hundreds of CPUs
• Requiring long simulation time over a few months
• No. of QM regions and No. of QM atoms change
  dynamically

Gridifying the application Using GridRPC MPI

• lt Initial conditiongt
• 2 QM regions with
• 72 41 QM atoms
• Totally 28598 atoms

3
Testbed for the Friction Simulation

• Used 11 clusters with totally 632 CPUs in 8
  organizations.
• PRAGMA Clusters
• SDSC (32 CPUs), KU (8 CPUs), NCSA (8 CPUs), NCHC
  (8 CPUs)
• Titech-1(8 CPUs), AIST(8 CPUs)
• AIST Super Cluster
• M64 (128 CPUs), F32-1(128 CPUs 128 CPUs)
• Japan Clusters
• U-Tokyo (128 CPUs), Tokushima-U (32 CPUs),
  Titech-2 (16 CPUs)

M64
F32
NCSA
KU
U-Tokyo
SDSC
AIST
Titech-2
NCHC
Titech-1
Tokushima-U
4
Result of the Friction Simulation

• Experiment Time 52. 5 days
• Longest Calculation Time 22 day
• Manual restart 2 times
• Execution failure 165 times
• Succeeded in recovering these failures
• Changing the No. of CPUs used 18 times
• succeeded in adjusting No. of CPUs to the No. of
  QM regions/QM atoms

5
Summary and future work

• Our approach is effective for running large-scale
  grid applications for a long time
• Need more grid services
• Getting information on available resources
• Resource reservation
• Coordinating with resource manager/scheduler
• Need cleaner MPI
• mpich quits leaving processes/IPC resources
• Using GridMPI in place of mpich

6
(No Transcript)