Parallelisation of Nonlinear Structural Analysis using Dual Partition Super-Elements
G.A. Jokhio and B.A. Izzuddin
Overview
•Introduction
•Proposed Structural Decomposition Approach
•Implementation by Parallelisation
•Verification
•Conclusions
Introduction
• A new approach of structural domain partitioning
• Parts of a structure replaced by partition super elements
• Removed parts modelled separately as child structures
• Dual partition super elements wrap the partition boundary of respective
child structures
Introduction
Purpose
• Reduction in computational time
• By using parallel processing
• By using reduced dimensional elements for non-critical parts
• Simplification of modelling
• By modular modelling
• Freedom
• To choose partitions
• To use different integration schemes
• To consider multi-physics
Introduction
•Implementation
• Parallelisation of ADAPTIC
• MPI
•Verification Examples
Case 2: A parent and a child partition (parent also models a part of structure made with other elements)
Case 2: A parent and a child partition (parent also models a part of structure made with other elements)
Case 3: A parent and a child partition (Same as case 2 but parent and child roles reversed)
Case 3: A parent and a child partition (Same as case 2 but parent and child roles reversed)
Structural Domain Partitioning
Partition super element on parent side
Dual partition super element on child side
Case 1: A parent and 2 child partitions (parent has only partition super elements)
Case 1: A parent and 2 child partitions (parent has only partition super elements)
2nd Level of Partitioning2nd Level of Partitioning1st Level of Partitioning1st Level of Partitioning
Structural Domain Partitioning
Hierarchical Approach
Partition super element on parent side
Dual partition super element on child side
Rationale
• Partitions ‘represented’ by partition super elements
• Any solution procedure based on monolithic approach ‘will do’
• The child partition boundary is analogous to Essential Boundary Conditions
• Compatibility: Parent sends these BCs to partitions
• Equilibrium: Parent receives the tangent stiffness and resistance forces from
partitions
Procedure
•Load/Time Steps
• Step wise loading/time-stepping controlled by the parent
• The load factor/time is sent to the partitions
•Equilibrium Convergence
• Convergence at parent structure level
• Convergence at partition level
•Iteration
• Iterative corrections to displacements sent to the partitions
• Resulting forces and tangent stiffness received by the parent
•The Frontal Method is most suitable – other methods can be used
Implementation
ADAPTIC
• Source code available
• Written in Fortran – Most widely used language in HPC
• Analysis types:• Static Proportional Loading• Static Time History• Dynamic• Eigenvalue
• A wide range of structural elements and material models available
• Supports advanced adaptive techniques
Other structural analysis programs can
also be used
Implementation
Start, Initialize MPIStart, Initialize MPI
Send Load Factor, Incremental Displacements etc. to the
Partitions`
Send Load Factor, Incremental Displacements etc. to the
Partitions`
Get Resistance Forces and Tangent Stiffness from the
Partitions
Get Resistance Forces and Tangent Stiffness from the
Partitions
Check ConvergenceCheck Convergence
Next Iteration or Load Step?Next Iteration or Load Step?
EndEnd
Overview of the
Parent Structure
Algorithm
Implementation
Start, Initialize MPIStart, Initialize MPI
Receive Instruction and the relevant data from the ParentReceive Instruction and the
relevant data from the Parent
Overview of the
Partition/Child
Algorithm
Perform the relevant TaskPerform the relevant Task
Continue?Continue?
EndEnd
Send the Outcome of the relevant task to the parent
Send the Outcome of the relevant task to the parent
Verification
Example 1: A 4 storey Structure
6
3
3
3
4
3
3
3
4
6
Original Structure Partition No. 1 Partition No. 2
Node 42
x
y
Parent Structure
Verification
Example 2: I-Beam using 3D Brick Elements
Time taken for monolithic analysis 188 secondsTime taken for partitioned analysis with 3 partitions
80 seconds
Time saved 57.4 %
Verification
Example 3: 3D Framed Structure
Load steps completed (in 72 hours) for:
Initial Loading Time-history Total
Monolithic analysis
1 9 10
Partitioned analysis with 11 partitions
1 120 121
Conclusions
• A new structural decomposition approach for partitioned analysis using
parallel processing has been proposed
• Uses dual partition-super elements
• Can be used for the parallelisation of existing monolithic analysis codes
• Has been implemented with ADAPTIC using a parallel MPI scheme
• The results match exactly with those obtained from conventional
monolithic analysis
• Significant computational savings arise in the analysis of large structures,
with great speedups achieved
• The proposed partitioning approach can also simplify the modelling
process through the use of modular partitions
Acknowledgements
• Higher Education Commission of Pakistan
• High Performance Computing (HPC) Services, Imperial College London
Questions?