文稿案例讲稿expl-l09largemodels

1、Lesson content:IntroductionSimplifying the Model Parallel ExecutionTechniques for Reducing CPU TimeTipsLesson 9: Managing Large Models: Part 190 minutesIntroduction (1/2)What is a Large Model?Difficult to defineAbsolute size changes as computers e more powerful.Rough measures for large models based

2、on number of variables:1980: Models with 100,000 variables 1990: Models with 1,000,000 variables2000: Models with 5,000,000 variablesTurn-around time is a universal measureThis measure includes jobs with relatively few variables and a huge number of increments.Models that take at least 12 hours (ove

3、rnight) to run are generally considered “large.Introduction (2/2)Large models often require:Simplification of the model geometry, physics, element formulation, etc. Intelligent use of system resources:disk space, available processors, memoryMinimization of CPU time, number of increments, etc. These

4、techniques will be discussed in this lecture Several techniques are available to assist in the management of large models:Submodeling Restart Parts and assemblies These techniques will be discussed in the next lecture. Simplifying the Model (1/14)There are a number of techniques that can be used to

5、simplify a model:Geometry simplificationModeling fewer detailsRigid body prototypingSolid element formulation Small-strain shellsSimplifying the Model (2/14)Simplify geometryWhen building a complex model, start simple (and small), then add complexity gradually.Making careful choices regarding your g

6、eometry and mesh:Take advantage of symmetry axial, planar, quarter, half, cyclic,Defeature the geometryAvoid small features They result in small elements which in turn reduce the stable time increment.Defeaturing a mechanical componentSimplifying the Model (3/14)Model fewer detailsConsider what you

7、want from the analysis and reduce the level of modeling detail in regions that are less critical to the analysis goals.For example:If you have two parts connected to each other and if you are not interested in detailed stresses of the connection, use connector elements.Use beams or shells instead of

8、 solids when appropriate. Focus mesh in regions of interest and use a coarser mesh elsewhere.Simplify the physics. Can a stiff portion of an assembly be replaced with a rigid body or a boundary condition?Do you need all interactions?Simplifying the Model (4/14)Example: Simplified model of a sports b

9、all for use in an impact analysis.The ball consists of a solid core and a shell cover.The detailed behavior of the ball is not of interest.The ball is reduced to the following:Analytical rigid surface for ball exterior contact surfaceLumped mass and rotary inertia for coreLumped mass and rotary iner

10、tia for coverConnector elements for stiffness and damping between:Core and coverCover and rigid exterior surfacecorecoverSchematic of modeled ball with a quarter of the cover cutawayBall ModelSimplifying the Model (5/14)Prototype with rigid bodiesMake deformable parts rigid with a rigid body constra

11、int for faster initial runs of the analysis.Once you have verified the model definition, remove the rigid body constraints.Simplifying the Model (6/14)Example: Drop Test Simulation of a Cordless MouseRigid body constraints are used so that initial versions of the analyses run quickly.*RIGID BODY, EL

12、SET=RIGID-BASE, REF NODE=BASE-RP, POSITION=CENTER OF MASSMouse baseMouseRIGID-BASE(white elements)Deformable elements (red)Simplifying the Model (7/14)Example (contd): Drop Test Simulation of a Cordless MouseFinal deformed shapesDeformable regions in colorfully deformable mouserigid prototype mouseM

13、ouse type Portion of mouse that is deformableRelative CPU timerigid prototype23%0.35fully deformable100%1Simplifying the Model (8/14)Solid element formulationAbaqus/Explicit offers three alternative kinematic formulations for the C3D8R solid element:The average strain formulation (default)Most expen

14、sive; highest level of accuracyThe orthogonal formulation Provides a balance between computational speed and accuracy i.e., less expensive than the average strain formulation and more accurate than the centroid formulationThe centroid formulation Least expensive; lowest level of accuracyC3D8RSimplif

15、ying the Model (9/14)Average strain kinematic formulation for hexahedral solid elementsThe default average strain kinematic formulation for solid elements is based on the uniform strain operator and orthogonal hourglass shape vectors. The resulting elements:Pass the constant strain patch test for a

16、general configuration.Have zero strain under large rigid body rotation.The default kinematic formulation can be very expensive, especially for three-dimensional problems. Simplifying the Model (10/14)Orthogonal kinematic formulation for hexahedral solid elementsThe orthogonal formulation is based on

17、 the centroidal strain operator and a slight modification to the hourglass shape vectors. The centroidal strain operator requires 3 fewer floating point operations than the uniform strain operator. These elements pass the patch test only for rectangular or parallelepiped element configurations.Howev

18、er, they converge well for general element configurations as the mesh is refined. This formulation is mended for all analyses except those involving:highly distorted elements, very coarse meshes, or high confinement. Simplifying the Model (11/14)Example: Copper rod impact*SOLID SECTION, ELSET=ROD, M

19、ATERIAL=COPPER, CONTROL=ORTHOGONAL*SECTION CONTROLS, KINEMATIC SPLIT=ORTHOGONAL, BINED, NAME=ORTHOGONALundeformed shapedeformed shapeC3D8R formulationShortening (mm)Widening (mm)Relative CPU timedefault (average strain)-13.085.4731orthogonal kinematics-13.095.4720.63Simplifying the Model (12/14)Cent

20、roid kinematic formulation for hexahedral solid elementsThe centroid formulation is the most economical. The centroid formulation is based on the centroidal strain operator and simple hourglass base vectors. Using the simple hourglass base vectors instead of the orthogonal hourglass shape vectors re

21、duces hourglass mode computations by a factor of three.The hourglass base vectors are not orthogonal to rigid body rotation for general element configurations. Therefore, hourglass strain may be generated with large rigid body rotations with this formulation. This formulation is mended only for prob

22、lems with little rigid body rotation and reasonable mesh refinement. Simplifying the Model (13/14)Example: Copper rod impact (contd)*SOLID SECTION, ELSET=ROD, MATERIAL=COPPER, CONTROL=ORTHOGONAL*SECTION CONTROLS, KINEMATIC SPLIT=CENTROID, BINED, NAME=ORTHOGONALdeformed shapeundeformed shapeC3D8R for

23、mulationShortening (mm)Widening (mm)Relative CPU timedefault (average strain)-13.085.4731orthogonal kinematics-13.095.4720.63centroidal kinematics-13.075.2310.49Simplifying the Model (14/14)Small-strain shell elementsSmall-strain shell elements were discussed in Lecture 2, Elements.Small-strain shel

24、l elements are more computationally efficient than their large-strain counterparts.They are appropriate for analyses involving small membrane strains and arbitrarily large rotationsundeformed shapedeformed shapePipe whip simulation analyzed with various shell element typesElement typeRelative CPU ti

25、meS4R1.0S4RSW0.78S4RS0.66Parallel Execution (1/13)Parallel execution: reduces run time for analyses that require a large number of increments and/or contain a large number of nodes and elementsproduces analysis results that are independent of the number of processors used provided the number of doma

26、ins is the same for each job 45,785 elements 370,086 DOF 3 ms simulation 87,369 incrementsFor the latest performance data visit Support Certified hardware SIMULIA System InformationParallel Execution (2/13)Activating parallel processing with the default options: Parallelization in Abaqus/Explicit is

27、 available for:shared memory architecture platforms using a thread-based loop level or thread-based domain position implementationboth shared and distributed memory architecture platforms using an MPI-based domain position parallel implementationabaqus job=phone cpus=4Parallel Execution (3/13)Shared

28、 memorySeparate processors share data on the same memory spacei.e., these are multiple processors of a single machine Data shared via pointers (thread-based) or passed via an interface (MPI)Distributed memory Processors have independent memory space i.e., these are multiple processors on separate ma

29、chines (clusters). Data passed via an interface (MPI)shared memory architectureCPU 1CPU 2Memorydistributed memory architecture (clusters)CPU 1CPU 2NetworkMemoryMemoryParallel Execution (4/13)shared memory architectureCPU 1CPU 2MemoryCommunication between processorsThe thread-based mode:data are shar

30、ed between processors via pointersavailable only on shared memory architecture platformsis available on all supported shared memory systemsThread - basedParallel Execution (5/13)shared memory architectureCPU 1CPU 2Memorydistributed memory architecture (clusters)NetworkCPU 1MemoryCPU 2MemoryCommunica

31、tion between processorsThe MPI (Message Passing Interface) mode: is available for both shared and distributed memory architecture platformsThe MPI facilitates messaging between processorsrequires that the MPI components be installedMPIMPIParallel Execution (6/13)Parallelization in Abaqus/Explicit is

32、 implemented in two ways: domain level loop levelDomain-level position methodWith this method Abaqus automatically breaks the model up into topological domains and assigns each domain to a processor. Domains are distributed evenly among the available processors. The analysis is then carried out inde

33、pendently in each domain.However, information must be passed between the domains in each increment because the domains share common boundaries. 8 domains1Parallel Execution (7/13)Domain-level position method (contd)The domain-level method is the default.It is mended for most large analyses.The domai

34、n-level speedup factor may be significantly more than what can be achieved with loop-level parallelization (discussed later). Both MPI- and thread-based communication schemes are supported with the domain-level method.The MPI mode is required for distributed memory systems (clusters).During the anal

35、ysis, separate selected results (.sel) and state (.abq) files are created for each processor.At the end of the analysis the individual files are merged automatically.There are some restrictions when using the domain-level method. For example, general contact is mended over contact pairs Refer to the

36、 user guides for more information.18 domainsParallel Execution (8/13)Domain-level position method (contd)Execution procedure (at the command line):abaqus job=phone cpus=8 parallel=domain domains=8 mp_mode=mpiDefaults can be modified using abaqus_v6.env.RequiredDefaults 18 domainsParallel Execution (

37、9/13)Domain-level position method (contd)Execution procedure (at the command line):abaqus job=phone cpus=4 parallel=domain domains= mp_mode=mpiRequired18 dynamic_load_balancing The number of domains must be evenly divisible by the number of CPUsDynamic load balancing is available when the number of

38、domains is higher than the number of CPUsThe efficiency of the dynamic load balancing scheme depends onthe load imbalance inherent to the problem, on the degree of position, mendation: number of domains 4 * number CPUsand on the efficiency of the hardware8 domainsParallel Execution (10/13)Domain-lev

39、el position method (contd)Applications most likely to benefit from dynamic load balancing are problems with a strongly time-dependent and/or spatially varying computational load. 1Airbag deploymentCoupled Eulerian-Lagrangian (CEL) bird strike analysis Eulerian domaintarget panels (Lagrangian)birds (

40、Eulerian)Parallel Execution (11/13)Domain-level position method (contd)Abaqus automatically creates element and node sets for each domain.8 domains1Parallel Execution (12/13)Loop-level methodThe loop-level method parallelizes low-level loops that are responsible for most of the computational cost. T

41、he element, node, and contact pair operations account for the majority of the low-level parallelized routines. Only the thread-based communication scheme is supported with the loop-level method.Distributed memory systems (clusters) are not supported.There are no additional restrictions when running

42、with this method;however, the speedup factor may be significantly less than what can be achieved with domain-level parallelization.The speedup factor will vary depending on the features included in the analysis.Execution procedure (at the command line):abaqus job =job-name cpus=n parallel=loop2Paral

43、lel Execution (13/13)Parallelization options summary for Abaqus/Explicit:MPI-based Loop-levelparallelizationDomain-levelparallelizationShared memory architectureDistributed memory architecture (clusters)Thread-basedThread-basedMPI-based Domain-levelparallelization= Default (when available) Technique

44、s for Reducing CPU Time (1/13)Features that can be used to reduce analysis run time include:Initial conditionsStatic initialization and importMonitoring for extreme values of output variablesMonitoring for steady stateMass scalingSelective subcyclingIncreasing load ratesTechniques for Reducing CPU T

45、ime (2/13)Initial conditions can be used to reduce the length of some analyses.You can define initial conditions (stress fields, hardening conditions, velocities,) rather than modeling the history that lead to the initial conditions.Example: Drop test of a cordless mouseApply initial velocity condit

46、ions consistent with a 1m free fall.V3Mouse initial velocity*INITIAL CONDITIONS, TYPE=VELOCITY ALL_NODES, 3, 4500Techniques for Reducing CPU Time (3/13)Static initialization and importAbaqus provides a capability to transfer the deformed mesh and associated material state between an Abaqus/Standard

47、analysis and an Abaqus/Explicit analysis (and vice-versa).Reduce the analysis run time by solving portions of an analysis with Abaqus/Standard.Examples:Simulate a transient event in Abaqus/Explicit after steady-state conditions have been simulated in Abaqus/StandardSimulate springback in Abaqus/Stan

48、dard after forming analysis in Abaqus/Explicit This functionality is discussed in Lecture 12, Importing and Transferring Results.steady-state spinning solution in Abaqus/Standarddynamic blade release analysis in Abaqus/ExplicitTechniques for Reducing CPU Time (4/13)Monitoring for extreme values of o

49、utput variablesMonitoring the values of critical variables and halting the analysis when those variables exceed a given criterion can increase modeling efficiency.For example, in a hydroforming simulation the time needed to model the completion of the physical process may be unknown. It may depend o

50、n the magnitude of the displacement of a node or a group of nodes in the model.Terminating the analysis when that displacement is reached may save many CPU hours.fluid pressurepunch loadTechniques for Reducing CPU Time (5/13)Use output filters to record bounding output values over time and halt an a

51、nalysis when a critical value has been reachedThe bounding value operators areMaximum valueMinimum valueAbsolute maximum valueThese can be applied to field output and history outputThey can be applied to digitally filtered or unfiltered data For more information on digital filters see Lecture 11 Out

52、put FilteringFor tensor and vector field output requests, Abaqus applies the bounding value operations to each component independently. Techniques for Reducing CPU Time (6/13)Example: Hydroforming of a boxStop the analysis before the MONITOR node pulls out from between the holder and the draw cap. D

53、efine the filter*FILTER, NAME=MIN-53, OPERATOR=MIN, LIMIT=-53, HALTExploded view of initial configuration draw capblankBlank holderpunchhydroforming pressure loadpunch load13253 mm in the negative 2-directionMONITOR nodeUse the Operator type to find the minimum without digitally filtering the result

54、1Techniques for Reducing CPU Time (7/13)Example: Hydroforming of a boxStop the analysis before the MONITOR node pulls out from between the holder and the draw cap. Define the output request*Output, history, filter=Min-53*Node Output, nset=MONITOR U2,2Techniques for Reducing CPU Time (8/13)Example (c

55、ontd): Hydroforming of a boxThe end of the status (.sta) file:Final deformed shape of the blankMONITOR (node 2)Techniques for Reducing CPU Time (9/13)Monitoring for steady stateSteady-state detection can be used to terminate an Abaqus/Explicit analysis when specified steady-state criteria are met.Th

56、is feature is available for quasi-static unidirectional processes.For example rolling, wire drawing, and extrusion processes The following steady-state detection norms are available for the specification of steady-state detection criteria:Normalized cross-section plastic strain at a specified cuttin

57、g planeLargest of the area moments of inertia of the plane cross sectionAverage force magnitude at a rigid body reference nodeAverage torque magnitude at a rigid body reference nodeTerminating the analysis when steady state is reached may save many CPU hours.Techniques for Reducing CPU Time (10/13)E

58、xample: Steady-state flat rolling simulationSteady-state detection is used to terminate a flat rolling simulation.A complete discussion of this example is included in Appendix 3 Advanced Adaptive Meshing.Request sampling at uniform intervals for an Eulerian adaptive meshing analysis. Cutting plane A

59、nalysis terminates when the steady state is detected at the cutting plane.outflowinflowsteady-state criteria definitionOnly when all of the criteria specified have been satisfied will the analysis be considered to have reached steady state. Techniques for Reducing CPU Time (11/13)Mass scaling speeds

60、 up an analysis by increasing the stable time increment size.Artificially increasing the material density by a factor of f 2 increases the stable time increment by a factor of f.Mass scaling for quasi-static analyses was discussed in Lecture 4.Mass scaling for high-speed dynamic events was discussed