CN108228970A - The explicit asynchronous long parallel calculating method of structural dynamical model - Google Patents
The explicit asynchronous long parallel calculating method of structural dynamical model Download PDFInfo
- Publication number
- CN108228970A CN108228970A CN201711308881.3A CN201711308881A CN108228970A CN 108228970 A CN108228970 A CN 108228970A CN 201711308881 A CN201711308881 A CN 201711308881A CN 108228970 A CN108228970 A CN 108228970A
- Authority
- CN
- China
- Prior art keywords
- subregion
- explicit
- boundary
- node
- length
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/06—Power analysis or power optimisation
Abstract
The present invention provides a kind of structural dynamical model explicit asynchronous long parallel calculating method, including four steps:(1) by subregion multiple knot strategy generating parallel computation needed for data file;(2) each subregion is selected time step to carry out explicit prediction step and is calculated according to zoning unit characteristic;(3) small step-length subregion is successively successively decreased using explicit subcycle process asks for asking for internal and boundary node data using main time step with boundary node data, big step-length subregion inside subregion;(4) each subregion receives parallel required boundary node data, regeneration block information, and according to calculating strain, stress is required, exports partitioned nodes data.The program determination if main time the end of the step, otherwise again since (2).The present invention can select different time steps during extensive dynamics parallel parsing according to zoning unit characteristic, realize the coupling of subregion asynchronous computing using multiple knot strategy, effectively increase parallel efficiency.
Description
Technical field
The present invention relates to a kind of design method of supercomputer applied technical field, specifically a kind of structural dynamic credits
The explicit asynchronous long parallel calculating method of analysis.
Background technology
Theoretical, experiment is to push the carriage drawn by a team of three horses of development in science and technology with calculating.It is ground recently as China in supercomputer
Continuous input in system develops with technology, and it is super for representative with " Milky Way 2 ", " light in martial prowess Taihu Lake " to have emerged a batch
Computing system, continuously in occupation of the strong super computer system umber ones of world TOP500.Along with hardware technology and numerical technique
Fast development, computing technique have become qualified engineering sharp weapon, Space Equipment, deep-sea technology, nano material and
The fields such as biomedicine play the role of vital.
The opposite of one and another gratifying breakthrough is obtained with super computer system hardware technology, in high-performance calculation
Software development is with applying the upper development for much lagging behind computing hardware system.Compared with external numerical simulation developed country, no matter
In the application aspect of simulation algorithm theoretical research, the development of high-performance calculation software or high-performance calculation, China also exist compared with
Big gap.With the development of Technology for Modern Equipment system complexity, to high-precision, extensive, overall process and multiple dimensioned numerical simulation
Technology proposes the requirement of more high-tech.
In order to meet the needs of domestic scientific research and project analysis, China will be external large scale business software every year
Pay expensive cost of use.In finite element analysis field, large scale business software such as Ansys, the Abaqus of foreign well-known,
Ls-Dyna and Adina etc. has issued parallel computation version.But they can manipulate most Chinese user for various reasons
Big Parallel implementation check figure has carried out a series of limitations, and the business finite element software dynamics calculation ability that China introduces substantially maintains
In million free measurement levels, cause the efficient calculated performance of China's supercomputer cannot be fully played.At the same time,
Also existing on some supercomputers by the completely independent research and development in China from the business software that foreign countries introduce cannot be normal
The problem of installation and operation.
For supercomputer, corresponding application can not just be carried out by lacking suitable application software, also can not just be inhaled
Reference family efficiently solves the large-scale calculations problem faced in science and engineering using supercomputer.For this problem,
The existing farsighted scholar in part in the country, has started correlative study work, has been taken in parallel software development and system application field
Obtain achievement centainly.But existing achievement, mostly based on Region Decomposition strategy, solves complicated fine numerical simulation energy in technological layer
Power is insufficient.Therefore, how efficiently to develop finite element analysis parallel computation software with effectively widen supercomputer big
Scale, fine, multiple dimensioned field are using extremely urgent.
It finds by prior art documents:Liu Cheng etc. is in Chinese science:Physics mechanics astronomy, 2017,47
(10):It publishes an article on 104603 " the flexible multi-body system High Efficient Parallel Algorithms based on Region Decomposition ", this article is based on FETI
(Finite Element Tearing and Interconnecting) Region Decomposition technology, subzone boundaries use fore condition
Conjugate gradient (Preconditioned Conjugate Gradients, PCG) iterative algorithm Parallel implementation linearisation after
Equation group improves the efficiency of parallel computation.However when using this method progress large-scale parallel structural analysis, system in parallel effect
Rate is not high.This is because for extensive problem, first with the scale and item that increase interface equation of subregion number
Number of packages, which also sharply increases, causes it to need more iterations that could restrain when solving, and it is total thus to have increased considerably system
The traffic;Next, all subregions use the time step of same size in this method, and flexible multi-body system is needed to part
The fine grid division of local can cause entirety that must use smaller time step, the time required to greatly increasing calculating;In addition,
This method does not consider the characteristics of multinuclear distributed environment, its all subregion process when solving interface equation should participate in office
Portion's communication participates in global communication again.Due in node with inter-node communication delay difference, this result in interprocess communication and
Synchronization overhead can increase much with increasing for subregion and significantly.
Invention content
In view of the deficiencies of the prior art, the present invention proposes a kind of explicit asynchronous long parallel computation sides of structural dynamical model
Method.The characteristics of this method binding domain decomposition method is with subcycle method, can be according to each in the emulation of extensive fine structure
Zoning unit characteristic selects different integration steps, so as to effectively increase parallel efficiency calculation.
According to a kind of explicit asynchronous long parallel calculating method of structural dynamical model provided by the invention, including walking as follows
Suddenly:
The first step passes through the data file needed for subregion multiple knot strategy generating parallel computation;
Second step, each partitioned process form alone subregion stiffness matrix and subregion mass matrix, according to multiblock technique category
Property using explicit integration time step carry out prediction step calculate, obtain small step-length subregion and big step-length subregion;
Wherein, small step-length subregion refers to:The structure minimum critical step-length subregion limited by unit size and stability;Big step
Long subregion refers to:With minimum critical step-length into integral multiple (being more than 1) than the step-length subregion of row;
Third walks, and small step-length subregion is successively successively decreased using explicit subcycle process according to time of integration step-length asks for internal node
Data and boundary node data, big step-length subregion directly ask for internal node data using the explicit main time step of subcycle process
With boundary node data, and with small step-length subregion transmit boundary node data;
4th step, each subregion according to transmit come boundary node data, regeneration block nodal information, and according to requiring to count
Calculate zoning unit strain and stress, the displacement of output partitioned nodes, speed and acceleration information;
The program determination if main time the end of the step, otherwise again since second step.
Preferably, in the first step, the subregion multiple knot strategy is to realize subregion process by node allocation:Structure
Finite element grid is M subregion first by parallel subdivision, and each partitioned nodes are divided into internal node and (are not involved in the section of data transfer
Point) and boundary node (subregion participates in the node of data transfer).Previously given explicit product is actually needed according to structure in each subregion
Divide time step.It is after subregion the result is that all nodes are divided into given subregion, different time of integration step-lengths according to number of partitions
Adjacent subregion realizes subregion asynchronous parallel by being overlapped multiple boundary node.
Preferably, in the first step, the data file refers to each height point by subregion multiple knot strategy generating
The partition information file in area and boundary information file, each partition information file include the unit of the child partition, node, load,
Boundary constraint information, each boundary information file include boundary node number, boundary node number, the affiliated subregion of node, adjacent
The information such as the number of partitions and partition number, boundary information file are mainly used for the data communication during Asynchronous parallel computation.
Preferably, in the second step, multiblock technique attribute refers to influence the grid category of subregion minimum time step-length
Property, material properties selected including cell dimension, cell node number, sizing grid and subregion etc..
Preferably, in the second step, prediction step, which calculates, to be integrated by the explicit Newmark based on prediction correction form
Form is realized:Explicit Newmark is divided into prediction and walks with correcting two steps of step, wherein prediction walks to calculate partitioned nodes speed
With the discreet value of displacement, correction step corrects partitioned nodes speed and displacement according to the partitioned nodes acceleration actual value calculated,
Obtain partitioned nodes speed and displacement true value;Subregion mass matrix uses lumped mass matrix, thus subregion mass matrix only has
Diagonal element, solution procedure are not related to the process of matrix inversion, are a kind of finite element explicit solution form.
Preferably, in the third step, when the explicit subcycle process refers to that a time step includes two different
Between step-length, that is, main time step and subcycle time step Dynamic solving process.During Asynchronous parallel computation, each subregion can
To select subregion time of integration step-length according to subregion minimum time step-length.Main time step is the integral multiple n of subcycle time step.
In the iteration step of one time of integration step-length, main time step performs primary, subcycle step execution n times.
Preferably, in the third step, the boundary node data that need to be transmitted refer to the boundary information file in data file
In given information such as boundary node number, boundary node number, the affiliated subregion of boundary node and partition number.Explicit solution
After the iteration step of a complete time of integration step-length, subregion solves to obtain boundary node data and internal node data respectively, i.e., complete
Into the calculating of explicit correction step.The boundary section corresponding to given partition boundaries node serial number transmitted from subregion internal extraction needs
Boundary node according to the partition number that boundary node data are received in boundary information file, is numbered corresponding boundary by point data
Node data is passed on.Meanwhile it receives from the corresponding boundary node data of given partition number.
Preferably, in the 4th step, partitioned nodes information includes modal displacement, velocity and acceleration;Internal node is compiled
Number information is stored in the partition information file of data file, transmits and boundary node volume is contained in the boundary node data come
Number, wherein, internal node number refers to it is all be not involved in partition data transmission node serial numbers, be used to form structure finite element from
Scattered kinetics equation, boundary node number are transmitted for adjacent sectors boundary node information.It is newer in partitioned nodes information
Data are boundary node data.Each subregion corresponds to an output file, and output content can be that unit should according to program requirements
Power, strain either modal displacement are, velocity and acceleration.
Compared with prior art, the present invention has following advantageous effect:
(1) form is solved using the explicit Newmark based on prediction correction, subregion mass matrix uses lumped mass matrix
Form, the finite element structure kinetics equation freedom decoupling of time discrete form, it is only necessary to according to the degree of freedom on a node basis side of being formed
Journey, direct solution are not related to the solution procedure of Large Scale Sparse matrix.Meanwhile explicit solution method does not need to traditional minor structure
Interface equation is solved in method and then the process of internal node data is asked in back substitution.Boundary information transmission is only related with adjacent sectors,
Improve parallel computation procedural information transmission efficiency.
(2) structure is divided into several subregions.Single subregion corresponds to a partition information file and boundary information file, subregion
Message file save mesh, node, external force and boundary constraint information, the storage of boundary information file and partition information exchange files
Relevant information.Multifile parallel computation is suitble to extensive and ultra-large degree of freedom Structural Dynamics simulation calculation.
(3) each subregion can select dynamics time of integration step-length according to zoning unit characteristic, ensure partitioned nodes
Under the premise of computational accuracy, there is larger flexibility.It was calculated between different subregion step-lengths by the guarantee of subcycle flow
Cheng Tongyi solves the problems, such as that subregion step-length is inconsistent caused by grid local fine divides.
(4) using the subarea management strategy of multiple knot, the subregion of adjacent sectors is set as to couple subregion.Subregion walks
Long ratio is different, and the unit number of layers of overlapping is different.During explicit subcycle, boundary node data are passed according to explicit algorithm information
Characteristic solution is passed, is not related to Interpolation Process, solving precision is higher.
Description of the drawings
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon:
Fig. 1 is the partition of nodes strategy schematic diagram in the embodiment of the present invention;
Fig. 2 is the multiple knot schematic diagram in the embodiment of the present invention;
Fig. 3 is the explicit subcycle flow diagram in the embodiment of the present invention;
Fig. 4 is the data transfer schematic diagram in the embodiment of the present invention.
Specific embodiment
With reference to specific embodiment, the present invention is described in detail.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill to this field
For personnel, without departing from the inventive concept of the premise, various modifications and improvements can be made.These belong to the present invention
Protection domain.
Embodiment
Present embodiments provide a kind of explicit asynchronous long parallel calculating method of structural dynamical model.
The explicit asynchronous long parallel calculating method of the structural dynamical model, overall flow include:
Data file needed for first by subregion multiple knot strategy generating parallel computation, then each partitioned process is alone
Subregion stiffness matrix and subregion mass matrix are formed, prediction step is carried out using explicit integration time step according to multiblock technique attribute
Flow;Then small step-length subregion successively successively decreased according to time of integration step-length using explicit subcycle process ask for internal node data with
Boundary node data, big step-length subregion directly ask for internal node data and boundary node data, and pass using main time step
Boundary node data are passed to small step-length subregion;Last each subregion according to transmit come boundary node information, regeneration block information,
And according to calculating strain, stress is required, export partitioned nodes displacement, speed and acceleration information.The program if time the end of the step
It terminates, otherwise re-starts the explicit prediction step of each subregion.
With reference to specific example, the present embodiment is described in further detail.
Parallel calculating method provided in this embodiment is applied into " Milky Way 2 " supercomputing in Guangzhou supercomputing center
It is calculated on machine." Milky Way 2 " supercomputer is a typical large-scale distributed parallel computer.The Milky Way 2 by
16000 calculate node compositions.Each node has 2 based on 2692 processors of Ivy Bridge-E Xeon E5 and 3 Xeon
Phi coprocessors, it is accumulative to share 32000 Ivy Bridge processors and 48000 Xeon Phi.Each processor clock frequency
Rate is 2.2GHz, is 12 cores.Coprocessor clock frequency is 1.1GHz, and each coprocessor uses 57 in 61 cores
It is a.Add up to 3,120,000 calculating cores.Each node possesses 64GB main memories, and in the onboard 8GB of each Xeon Phi coprocessors
It deposits, therefore per the common 88GB memories of node.It is connected between operation node using TH Express-2 trunks topological structure network.
It is described in detail successively below as the sequencing of system operatio:
The first step passes through the data file needed for subregion multiple knot strategy generating parallel computation.METIS is internationally famous
It increases income serial subregion software.ParMETIS is the parallel version of increasing income of METIS, particularly suitable for large-scale structure grid parallel computation point
Area.This example realizes sectoring function using ParMETIS.
Structure finite element grid is n sub-regions by parallel subdivision first by ParMETIS, after ParMETIS is divided
The result is that structure finite element grid node belongs to different subregions, a partition information file and boundary are corresponded to per sub-regions
Message file.It needs that structured grid information is written to different partition information file and boundary information by subregion multiple knot strategy
File.Its flow is as shown in Figure 1.Structure finite element mesh generation completes output according to grid number order and node serial number sequence
File.
Grid data includes grid number and grid internal node number.Node data includes node serial number and node
Coordinate value.
It is numbered according to grid and reads in grid internal node number, according to node allocation data, search affiliated point of corresponding node
Area, judges whether the affiliated node of grid fully belongs to a subregion, if grid number and its node serial number then are written node
The corresponding partitioned file of affiliated partition number.If the affiliated different subregions of grid internal node, by all points of grid number write-in
Area code, the corresponding subregion of node write-in need whether decision node has existed the subregion, and there is no be then written.External force and boundary
Constraint is added according to node serial number.Multiple knot partitioning strategies are as shown in Figure 2.
Second step, each partitioned process form alone subregion stiffness matrix and mass matrix, are made according to multiblock technique attribute
Prediction step is carried out with explicit integration step-length to calculate." Milky Way 2 " supercomputer determines to calculate required node when task is submitted
With into number of passes, explicit Asynchronous parallel computation all calculates core using CPU, each node contains 2 CPU, and every CPU contains 12
A calculating core, all application calculating contains 24 cores to each node.Each core corresponds to a partitioned file and boundary text
Part.Each process reads in grid data, generation unit stiffness matrix and mass matrix, mass matrix according to grid sequence from information
Using lumped mass form, generating structure stiffness matrix and mass matrix are superimposed according to the degree of freedom on a node basis.Explicit time-integration lattice
Formula, minimum time step-length are limited by CFL condition
In formula:ΔtcrFor unit critical time step, μ is proportionality coefficient, and l is element characteristic length, ρ, ν and E generation respectively
Density, Poisson's ratio and the elasticity modulus of table material therefor.The time step that then may be used for a subregion is all lists
First minimum time step-length.
In formula:Δ t is zone time step-length,Represent critical step-length of the n-th in the n-th time step.Different subregions can
It is different with the step-length of use, implement algorithm and be the need to ensure that vectorial node step-length ratio is integral multiple.
Explicit Newmark Integration Schemes prediction step based on prediction correction form can be expressed as
In formula:Subscript n is time step mark, for zone time step-length, un、WithThe n-th time step subregion section is represented respectively
Point displacement, velocity and acceleration vector,WithRepresent the (n+1)th time step partitioned nodes displacement and rate predictions, β and γ
It is two adjustable parameters of explicit Newmark Integration Schemes.Adjustment β and γ can change filter of the algorithm for different frequency waveform
Wave property.
Third walk, small step-length subregion successively successively decrease according to time step using explicit subcycle process ask for subregion internal node and
Boundary node data, big step-length subregion directly asks for internal, boundary node data using main time step, and transmits boundary node
Data are to small step-length subregion.Subcycle flow is as shown in Figure 3.Given adjacent sectors step-length ratio is m
In formula:Δ T is big step-length subregion step-length, and Δ t is small step-length subregion step-length.Small step-length partitioned nodes are by Newmark
Prediction step calculate node speed and displacement prediction value.Structural Dynamics time discrete acceleration solution formula
In formula:M represents partitioned organization mass matrix, and C is partitioned organization damping matrix, and K is structural division stiffness matrix,The outer force vector of node is represented, subscript n+1 represents node in the corresponding variate-value of the (n+1)th time step.
In subcycle time step, small step-length Regionalization basis explicit algorithm communication process, with the increase of subcycle time step,
Successively successively decrease and ask for the acceleration value of internal node and boundary node, internal node acceleration figure can be expressed as
In formula:MiWith KiRepresent i-th of subregion mass matrix and stiffness matrix,WithTable
Show subregion mass matrix with stiffness matrix by internal node degree of freedom BIMatrix in block form.SubscriptRepresent n-th of main time step
+ 1 subcycle time step moment of jth is to variable, miRepresent subregion i step-length ratios.Represent boundary node displacement prediction value vector.Represent internal node displacement prediction value.fext,iRepresent that internal node corresponds to the outer force vector applied.Boundary node degree of freedom Bc
Vector.
The acceleration formula of boundary node can be expressed as
In formula:WithRepresent boundary node degree of freedom BcCorresponding piecemeal mass matrix,
Stiffness matrix.fext,cOuter force vector corresponding with boundary node degree of freedom,Represent boundary node vector acceleration.∑ expression is asked
And symbol.
The acceleration value of egress is calculated, displacement can be walked by Newmark integral corrections with speed and be calculated
In formula:WithRepresent n-th of main time step jth+1 sub- step moment circulation time, the displacement of node with
Velocity correction vector.Equation (9) is the correction equation based on the explicit Newmark forms of prediction correction with equation (10), with side
Time step is consistent used by journey (3), (4).Displacement, the speed of small step-length subregion internal node can be calculated by this step
With acceleration information.
Big step-length subregion carries out prediction step using main time step Δ T and calculates, and predicts step displacement calculation formula square journey
(3), time step Δ t in equation (3) and (4) is substituted for Δ T, Ran Houji by prediction of speed accounting equation square journey (4)
Calculate the acceleration value of big step-length subregion internal node
In formula:Subscript n+1 represents main time step n+1 to dependent variable,For internal node vector acceleration.
Calculate the acceleration value of internal node, the speed and shift value of node, using the explicit correction format meters of Newmark
It calculates, speed can be write as with displacement correction formula
In formula:WithFor the (n+1)th main time step modal displacement and velocity vector.Calculate the inside of big step-length subregion
Node data needs to send the data to information of the adjacent sectors as the boundary node of adjacent sectors, to participate in next master
The calculating of time step.
4th step, regeneration block information, and according to calculating strain, stress is required, export partitioned nodes displacement, speed and add
Speed data.The program determination if time the end of the step, the explicit prediction step for otherwise re-starting each subregion calculate.
Each subregion is in a main time step, and the data of internal node can correctly be obtained, and the data of boundary node can not
Correctly it is obtained.Using the partitioning strategies for overlapping node, for adjacent sectors, the boundary node of a subregion is exactly another point
The internal node in area.The information such as node serial number and partition number that subregion needs transmit are recorded in boundary information file.Number
According to transmitting, schematic diagram is as shown in Figure 4.
Partition data, which is transmitted, to carry out data transmission using MPI communication protocols and receives, and data transfer is dead in order to prevent
Subregion is divided into odd number and even number subregion by lock, and odd number subregion sends internal node data, even number after first receiving boundary node data
Subregion first sends internal node data and is followed by by boundary node data.Each subregion updates complete partitioned nodes data, according to need
Output partitioned nodes displacement, speed or acceleration value, computing unit strain and stress value can be selected, it as needed can be with
Output.
Main time step, which counts, increases by a step, calculates total time, program stopped if reaching, no person is calculated from algorithm second step
Start to continue to calculate
A kind of explicit asynchronous long parallel calculating method of structural dynamical model provided in this embodiment, passes through subregion multiple knot
Data file needed for strategy generating parallel computation;Each subregion selects time step explicitly to be predicted according to zoning unit characteristic
Step calculates;Small step-length subregion is successively successively decreased using explicit subcycle process asks for subregion inside and boundary node data, big step-length
Subregion asks for internal and boundary node data using main time step;Each subregion receives parallel required boundary node data, update
Partition information, and according to calculating strain, stress is required, export partitioned nodes data.The program determination if main time the end of the step, it is no
Then again since second step.The present embodiment can be according to zoning unit characteristic during extensive dynamics parallel parsing
Different time steps is selected, the coupling of subregion asynchronous computing is realized using multiple knot strategy, effectively increases parallel efficiency.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited in above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring the substantive content of the present invention.
Claims (7)
1. the explicit asynchronous long parallel calculating method of a kind of structural dynamical model, which is characterized in that include the following steps:
The first step passes through the data file needed for subregion multiple knot strategy generating parallel computation;
Second step, each partitioned process form alone subregion stiffness matrix and subregion mass matrix, and according to multiblock technique attribute
Prediction step is carried out using explicit integration time step to calculate, and subregion is divided into small step-length point according to subregion time of integration step sizes
Area and big step-length subregion;
Wherein, small step-length subregion refers to:The structure minimum critical step-length subregion limited by unit size and stability;Big step-length point
Area refers to:Step-length subregion with minimum critical step-length into the integral multiple more than 1 than row;
Third walks, and small step-length subregion is successively successively decreased using explicit subcycle process according to time of integration step-length asks for internal node data
With boundary node data;Big step-length subregion directly asks for internal node data and side using the explicit main time step of subcycle process
Boundary's node data, and transmit boundary node data with small step-length subregion;4th step, each subregion is according to the boundary node number of reception
According to, regeneration block nodal information, and according to requiring to calculate zoning unit strain and stress, the displacement of output partitioned nodes, speed and
Acceleration information;
Otherwise the program determination if main time the end of the step returns to second step and re-executes second step to the 4th step.
2. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 1, which is characterized in that described
In the first step, the subregion multiple knot strategy is to realize subregion process by node allocation:
Structure finite element grid by node allocation by parallel subdivision be M subregion, each partitioned nodes be divided into internal node and
Boundary node;Each subregion presets explicit integration time step according to multiblock technique characteristic;Different time of integration step-length phases
Adjacent each subregion realizes the asynchronous long parallel computation of subregion by being overlapped multiple boundary node.
3. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 1 or 2, which is characterized in that
In the first step, the data file refer to by the partition information file of each subregion of subregion multiple knot strategy generating with
Boundary information file;Each partition information file includes unit, node, load and the boundary constraint information of the subregion, each
Boundary information file includes boundary node number, boundary node number, the affiliated subregion of node, adjacent sectors number and partition number
Information;The boundary information file is used for the data communication during asynchronous long parallel computation.
4. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 3, which is characterized in that described
In second step, the prediction step, which calculates, to be realized by the explicit Newmark Integration Schemes based on prediction correction form:Explicitly
Newmark is divided into prediction and walks with correcting step, wherein prediction walks the discreet value for calculating partitioned nodes speed and displacement, school
Stride according to the partitioned nodes acceleration actual value correction partitioned nodes speed and displacement calculated, obtain partitioned nodes speed with
Displacement true value;Subregion mass matrix uses lumped mass matrix, which only has diagonal element, and solution procedure does not relate to
And the process of matrix inversion, it is a kind of finite element explicit solution form.
5. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 4, which is characterized in that described
In third step, the explicit subcycle process refers to that a time of integration step-length includes the i.e. main time step of two kinds of different time step-lengths
With the Dynamic solving process of subcycle time step;During asynchronous long parallel computation, each subregion is according to subregion minimum time
Step-length selects subregion time of integration step-length, and main time step is the integral multiple n of subcycle time step;In time of integration step-length
In iteration step, main time step performs primary, subcycle time step execution n times.
6. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 5, which is characterized in that described
In third step, the boundary node data refer to the boundary node given in the boundary information file of data file number, side
The affiliated subregion of boundary's interstitial content, boundary node and partition number information;The iteration of the complete time of integration step-length of explicit solution
After step, subregion solves to obtain boundary node data and internal node data respectively, that is, completes the calculating of explicit correction step;From subregion
The corresponding boundary node data of given partition boundaries node serial number that internal extraction needs transmit, connect according in boundary information file
The given partition number of boundary node data is received, boundary node is numbered into corresponding boundary node data transfer to specified partition,
Meanwhile it receives and numbers corresponding boundary node data from specified partition.
7. the explicit asynchronous long parallel calculating method of structural dynamical model according to claim 1, which is characterized in that described
In 4th step, the partitioned nodes information includes partitioned nodes displacement, velocity and acceleration;Internal node number information is stored in
In the partition information file of data file, transmit and boundary node number is contained in the boundary node data come, wherein, inside section
Point number refers to all node serial numbers for being not involved in partition data transmission, is used to form the discrete dynamics side of structure finite element
Journey, boundary node number are transmitted for adjacent sectors boundary node information;The newer data of institute are boundary in partitioned nodes information
Node data, each subregion correspond to an output file, the content of output file according to program requirements for zoning unit stress with
Strain either partitioned nodes displacement, velocity and acceleration.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711308881.3A CN108228970B (en) | 2017-12-11 | 2017-12-11 | Structural dynamics analysis explicit different step length parallel computing method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711308881.3A CN108228970B (en) | 2017-12-11 | 2017-12-11 | Structural dynamics analysis explicit different step length parallel computing method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108228970A true CN108228970A (en) | 2018-06-29 |
CN108228970B CN108228970B (en) | 2020-07-14 |
Family
ID=62649359
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711308881.3A Active CN108228970B (en) | 2017-12-11 | 2017-12-11 | Structural dynamics analysis explicit different step length parallel computing method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108228970B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109101708A (en) * | 2018-07-25 | 2018-12-28 | 福建工程学院 | The implicit finite element parallel method decomposed based on level-2 area |
CN109325260A (en) * | 2018-08-21 | 2019-02-12 | 上海交通大学 | A kind of kinetic model shows hidden mixed asynchronous long stagger scheme calculation method |
CN109359427A (en) * | 2018-11-26 | 2019-02-19 | 东南大学 | A kind of real-time hybrid analog-digital simulation test method of more actuator for space structure |
CN109726354A (en) * | 2018-12-18 | 2019-05-07 | 山东交通学院 | A kind of parallel cross-correlation calculation method towards mass data |
CN113836766A (en) * | 2021-09-18 | 2021-12-24 | 北京理工大学 | METIS-based spherical hinge-containing multi-body system diagram dividing method |
CN117709129A (en) * | 2024-02-05 | 2024-03-15 | 国家超级计算天津中心 | Multi-scale simulation method, device, equipment and storage medium |
CN117709129B (en) * | 2024-02-05 | 2024-05-14 | 国家超级计算天津中心 | Multi-scale simulation method, device, equipment and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101095132A (en) * | 2004-12-31 | 2007-12-26 | 英特尔公司 | Data partitioning and critical section reduction for bayesian network structure learning |
US20120222034A1 (en) * | 2011-02-25 | 2012-08-30 | International Business Machines Corporation | Asynchronous checkpoint acqusition and recovery from the checkpoint in parallel computer calculation in iteration method |
CN103324850A (en) * | 2013-06-21 | 2013-09-25 | 上海交通大学 | Finite element two-stage partition and twice polycondensation parallel method based on multiple document flows |
CN103914428A (en) * | 2014-03-25 | 2014-07-09 | 上海交通大学 | Efficient communication method of structural analysis under multi-core distributed computing environment |
CN104572109A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Two-stage partitioned two-time polycondensation parallel computing system development method and parallel computing system |
CN107239586A (en) * | 2016-03-29 | 2017-10-10 | 南京理工大学 | The domain decomposition parallel method effective to unconditionally stable FDTD method |
-
2017
- 2017-12-11 CN CN201711308881.3A patent/CN108228970B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101095132A (en) * | 2004-12-31 | 2007-12-26 | 英特尔公司 | Data partitioning and critical section reduction for bayesian network structure learning |
US20120222034A1 (en) * | 2011-02-25 | 2012-08-30 | International Business Machines Corporation | Asynchronous checkpoint acqusition and recovery from the checkpoint in parallel computer calculation in iteration method |
CN103324850A (en) * | 2013-06-21 | 2013-09-25 | 上海交通大学 | Finite element two-stage partition and twice polycondensation parallel method based on multiple document flows |
CN103914428A (en) * | 2014-03-25 | 2014-07-09 | 上海交通大学 | Efficient communication method of structural analysis under multi-core distributed computing environment |
CN104572109A (en) * | 2015-01-19 | 2015-04-29 | 上海交通大学 | Two-stage partitioned two-time polycondensation parallel computing system development method and parallel computing system |
CN107239586A (en) * | 2016-03-29 | 2017-10-10 | 南京理工大学 | The domain decomposition parallel method effective to unconditionally stable FDTD method |
Non-Patent Citations (1)
Title |
---|
张伟伟: "大型正交异性结构动力学分析的空间—时域多尺度方法及应用研究", 《中国博士学位论文全文数据库 基础科学辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109101708A (en) * | 2018-07-25 | 2018-12-28 | 福建工程学院 | The implicit finite element parallel method decomposed based on level-2 area |
CN109101708B (en) * | 2018-07-25 | 2022-06-28 | 福建工程学院 | Implicit finite element parallel method based on two-stage region decomposition |
CN109325260A (en) * | 2018-08-21 | 2019-02-12 | 上海交通大学 | A kind of kinetic model shows hidden mixed asynchronous long stagger scheme calculation method |
CN109359427A (en) * | 2018-11-26 | 2019-02-19 | 东南大学 | A kind of real-time hybrid analog-digital simulation test method of more actuator for space structure |
CN109359427B (en) * | 2018-11-26 | 2021-12-07 | 东南大学 | Multi-actuator real-time hybrid simulation test method for spatial structure |
CN109726354A (en) * | 2018-12-18 | 2019-05-07 | 山东交通学院 | A kind of parallel cross-correlation calculation method towards mass data |
CN109726354B (en) * | 2018-12-18 | 2022-12-20 | 山东交通学院 | Parallel cross-correlation computing method for mass data |
CN113836766A (en) * | 2021-09-18 | 2021-12-24 | 北京理工大学 | METIS-based spherical hinge-containing multi-body system diagram dividing method |
CN113836766B (en) * | 2021-09-18 | 2024-04-26 | 北京理工大学 | METIS-based spherical hinge-containing multi-body system diagram dividing method |
CN117709129A (en) * | 2024-02-05 | 2024-03-15 | 国家超级计算天津中心 | Multi-scale simulation method, device, equipment and storage medium |
CN117709129B (en) * | 2024-02-05 | 2024-05-14 | 国家超级计算天津中心 | Multi-scale simulation method, device, equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN108228970B (en) | 2020-07-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108228970A (en) | The explicit asynchronous long parallel calculating method of structural dynamical model | |
CN103324850B (en) | Twice polycondensation parallel method of finite element two-stage subregion based on multifile stream | |
Barnard et al. | Fast multilevel implementation of recursive spectral bisection for partitioning unstructured problems | |
Bender et al. | Cache-adaptive algorithms | |
Benner et al. | A mixed-precision algorithm for the solution of Lyapunov equations on hybrid CPU–GPU platforms | |
CN103345580A (en) | Parallel CFD method based on lattice Boltzmann method | |
JP2023506343A (en) | Vector reduction using shared scratchpad memory | |
Minervini et al. | Vitruvius+: an area-efficient RISC-V decoupled vector coprocessor for high performance computing applications | |
Stone et al. | GPGPU parallel algorithms for structured-grid CFD codes | |
Bartezzaghi et al. | An explicit dynamics GPU structural solver for thin shell finite elements | |
Munch et al. | hyper. deal: An efficient, matrix-free finite-element library for high-dimensional partial differential equations | |
Daxini et al. | Numerical shape optimization based on meshless method and stochastic optimization technique | |
Kavouklis et al. | Parallel adaptation of general three-dimensional hybrid meshes | |
CN106777065A (en) | The method and system that a kind of Frequent tree mining is excavated | |
Meister et al. | A software concept for cache-efficient simulation on dynamically adaptive structured triangular grids | |
CN102779025A (en) | Parallel PLSA (Probabilistic Latent Semantic Analysis) method based on Hadoop | |
CN116303219A (en) | Grid file acquisition method and device and electronic equipment | |
Wermelinger et al. | An efficient compressible multicomponent flow solver for heterogeneous CPU/GPU architectures | |
Nagy et al. | Accelerating unstructured finite volume computations on field‐programmable gate arrays | |
CN109522127A (en) | A kind of fluid machinery simulated program isomery accelerated method based on GPU | |
Wei et al. | GPU acceleration of a 2D compressible Euler solver on CUDA-based block-structured Cartesian meshes | |
CN104793955A (en) | Distributed vector computing frame | |
WO2022026549A1 (en) | Optimization processing unit utilizing digital oscillators | |
CN104424123A (en) | Lock-free data buffer and usage thereof | |
CN103530183B (en) | In large scale scale heterogeneous calculating system, task computation measurer has the dispatching method of randomness |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |