CN106897163A - A kind of algebra system method for solving and system based on KNL platforms - Google Patents
A kind of algebra system method for solving and system based on KNL platforms Download PDFInfo
- Publication number
- CN106897163A CN106897163A CN201710133888.XA CN201710133888A CN106897163A CN 106897163 A CN106897163 A CN 106897163A CN 201710133888 A CN201710133888 A CN 201710133888A CN 106897163 A CN106897163 A CN 106897163A
- Authority
- CN
- China
- Prior art keywords
- matrix
- calculating
- knl
- calculating matrix
- linear equations
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
- G06F17/12—Simultaneous equations, e.g. systems of linear equations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/54—Interprogram communication
- G06F9/546—Message passing systems or structures, e.g. queues
Abstract
This application discloses a kind of algebra system method for solving based on KNL platforms, including:The coefficient matrix and constant matrices of system of linear equations are respectively divided into L group calculating matrix;Be assigned to for L group calculating matrix N number of from process by host process, and each carries out successive ignition calculating from process to the calculating matrix for receiving, and obtains final result;Host process receives the final result that each is calculated from process.It can be seen that, the application is based on KNL platforms according to default division methods, the coefficient matrix and constant matrices of system of linear equations are respectively divided into L group calculating matrix, and L group calculating matrix being assigned to N number of from process, each carries out successive ignition calculating from process to the calculating matrix for receiving, and obtains final result, using multi-process computing advantage, system of linear equations block parallel is calculated, calculating speed is improve, the rapid computations to system of linear equations are realized.In addition, the application further correspondingly discloses a kind of algebra system solving system based on KNL platforms.
Description
Technical field
The present invention relates to high-performance computing sector, more particularly to a kind of algebra system method for solving based on KNL platforms and
System.
Background technology
With the development of science and technology the Solve problems of system of linear equations in 20 middle of century with the big molded line of the development of computer
Property equation rent solution be just possibly realized.The numerical method for solving system of linear equations can be generally divided into direct method and iterative method two
Kind.Iterative method is adapted to solve larger sparse vectors, the method for taking Approach by inchmeal, according to certain calculating lattice
Formula, constructs an infinite sequence for vector, and its limit is only the accurate solution of system of linear equations, and what limited number of time was calculated is approximate
Solution.But no matter which kind of iterative manner is substantially multiplied by vector and inner product of vectors by matrix-vector multiplication, addition of vectors (subtracting), number
Constituted Deng four kinds of computings.
KNL (Knights Landing) be Intel company release the second generation to core piece is melted by force, for high-performance simultaneously
The many-core processor that row is calculated.KNL chips can individually do central host processor, which employs the improvement of Silvermont frameworks
Customization version and 14nm new technologies, core amounts up to 64-72, each core can at most open 4 threads, at most possess 288
Individual thread, more than 3TFlops, single precision is then more than 6TFlops for double-precision floating point performance.
In the prior art, a series of numerical computation method phase such as finite difference, finite element, boundary element, non-mesh method
After birth.These numerical computation methods have a something in common:By mathematics physics model derived from practical problem by specific
Mode be separated into a linear algebraic equation systems.The problem of different field is obtained with the problem of different Numerical Methods Solves
The form or characteristic of matrix are often what is differed, and the matrix for different shape or characteristic chooses different method for solving, because
This, these mathematical physics always change into a Solve problems for linear algebra system.However, with the increasing of problem scale
Greatly, existing hardware apparatus and method, it is necessary to take considerable time when in face of complicated Solving Linear problem, it is difficult to
Calculating speed is further improved, therefore, the solution of system of linear equations is as the big bottleneck in engineering production and scientific research.
The content of the invention
In view of this, it is an object of the invention to provide a kind of algebra system method for solving and system based on KNL platforms,
To improve the calculating speed to system of linear equations.Its concrete scheme is as follows:
A kind of algebra system method for solving based on KNL platforms, including:
Host process reads system of linear equations, according to default division methods, by the coefficient matrix of the system of linear equations and
Constant matrices is respectively divided into mutually one-to-one L blocks, obtains L group calculating matrix, wherein, L is positive integer;
Be assigned to for L group calculating matrix N number of from process by the host process, and each is from process to the calculating matrix that receive
Successive ignition calculating is carried out, final result is obtained;Wherein, N is the positive integer less than or equal to L;
The host process receives the final result that each is calculated from process;
Wherein, each carries out any process of iterative calculation to the calculating matrix for receiving from process, including:From process
Calculating matrix to receiving are calculated, and obtain the first result of calculation, and first result of calculation is fed back into the master
Process, the host process receives first result of calculation from process, and first result of calculation from process is sent
To other from process.
Preferably, it is described according to default division methods, the coefficient matrix and constant matrices by system of linear equations point
The process of mutually one-to-one L blocks is not divided into, including:
By the coefficient matrix and constant matrices of the system of linear equations, divided by row is into L blocks respectively, every piece of coefficient matrix with
Every piece of constant matrices is mutually corresponded.
Preferably, the calculating matrix for receiving the are calculated process from process, including:
It is described to multiply function, inner product of vectors function and addition of vectors function using matrix-vector multiplication function, vectorial number from process
The calculating matrix to receiving are calculated.
Preferably, each interprocess communication mode is collective communication.
Preferably, also include:Be assigned to for L group calculating matrix N number of from process and host process by the host process.
Preferably, L group calculating matrix are assigned to N number of process from process by the host process, including:
Be evenly distributed to for L group calculating matrix N number of from process by the host process.
A kind of algebra system solving system based on KNL platforms, including:
Division module, reads system of linear equations, according to default division methods, by the system of linear equations for host process
Coefficient matrix and constant matrices be respectively divided into mutually one-to-one L blocks, obtain L group calculating matrix, wherein, L is just whole
Number;
, be assigned to for L group calculating matrix for the host process N number of from process by distribute module;
The calculating matrix for receiving are carried out successive ignition calculating by computing module for each from process, are most terminated
Really;Wherein, N is the positive integer less than or equal to L;
Collection module, each final result calculated from process is received for the host process;
Wherein, any iterative calculation is carried out to the calculating matrix for receiving from process to each in the computing module
Process, including:The calculating matrix for receiving are calculated from process, obtains the first result of calculation, and described first is calculated
Result feeds back to the host process, and the host process receives first result of calculation from process, and by described from process
First result of calculation is sent to other from process.
Preferably, the division module, specifically for by the coefficient matrix and constant matrices of the system of linear equations, distinguishing
Divided by row is mutually corresponded into L blocks, every piece of coefficient matrix with every piece of constant matrices.
Preferably, the computing module, including:
Computing unit, for it is described from process using matrix-vector multiplication function, vectorial number multiply function, inner product of vectors function and
The calculating matrix that addition of vectors function pair is received are calculated.
Preferably, be evenly distributed to for L group calculating matrix specifically for the host process N number of from entering by the distribute module
Cheng Zhong.
In the present invention, the algebra system method for solving based on KNL platforms, including:Host process reads system of linear equations, according to
Default division methods, mutually one-to-one L blocks are respectively divided into by the coefficient matrix and constant matrices of system of linear equations, are obtained
To L group calculating matrix, wherein, L is positive integer;Be assigned to for L group calculating matrix N number of from process by host process, and each is from process
Calculating matrix to receiving carry out successive ignition calculating, obtain final result;Wherein, N is the positive integer less than or equal to L;It is main
Process receives the final result that each is calculated from process;Wherein, each carries out any from process to the calculating matrix for receiving
The process of secondary iterative calculation, including:The calculating matrix for receiving are calculated from process, obtains the first result of calculation, and will
First result of calculation feeds back to host process, and host process receives the first result of calculation from process, and will be calculated from the first of process
Result is sent to other from process.It can be seen that, the present invention reads system of linear equations based on KNL platforms by host process, according to default
Division methods, the coefficient matrix and constant matrices of system of linear equations are respectively divided into mutually one-to-one L blocks, obtain L
Group calculating matrix, and L group calculating matrix are assigned to N number of from process, have ensured multi-process concurrent operation, and each is from process pair
The calculating matrix for receiving carry out successive ignition calculating, obtain final result, using multi-process computing advantage, by system of linear equations
Block parallel is calculated, and improves calculating speed, realizes the rapid computations to system of linear equations.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
The accompanying drawing to be used needed for having technology description is briefly described, it should be apparent that, drawings in the following description are only this
Inventive embodiment, for those of ordinary skill in the art, on the premise of not paying creative work, can also basis
The accompanying drawing of offer obtains other accompanying drawings.
Fig. 1 is a kind of algebra system method for solving schematic flow sheet based on KNL platforms provided in an embodiment of the present invention;
Fig. 2 is another algebra system method for solving schematic flow sheet based on KNL platforms provided in an embodiment of the present invention;
Fig. 3 is a kind of algebra system solving system structural representation based on KNL platforms provided in an embodiment of the present invention.
Specific embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Site preparation is described, it is clear that described embodiment is only a part of embodiment of the invention, rather than whole embodiments.It is based on
Embodiment in the present invention, it is every other that those of ordinary skill in the art are obtained under the premise of creative work is not made
Embodiment, belongs to the scope of protection of the invention.
The embodiment of the invention discloses a kind of algebra system method for solving based on KNL platforms, shown in Figure 1, the party
Method includes:
Step S11:Host process reads system of linear equations, according to default division methods, by the coefficient square of system of linear equations
Battle array and constant matrices are respectively divided into mutually one-to-one L blocks, obtain L group calculating matrix, wherein, L is positive integer.
Specifically, KNL platforms receive calculative system of linear equations, and the host process in KNL platforms reads linear equation
The information of group, for example, the information such as the ranks number of coefficient matrix and constant matrices, each numerical value in matrix, advance according to user
The division methods of setting, polylith, and coefficient matrix block and constant are divided into by the coefficient matrix and constant matrices of system of linear equations
Matrix-block is corresponded, and obtains L group calculating matrix, it is to be understood that one group of calculating matrix includes a pair of coefficient matrix blocks
With constant matrices block.
Wherein, user's division methods set in advance can include divided by row, divided by column or by block division, according to phase
The division methods answered, it is established that the corresponding relation of coefficient matrix block and constant matrices block, for example, during divided by row, coefficient matrix
The first row it is corresponding with constant matrices the first row, as first group of calculating matrix;When being divided by block, first piece of coefficient matrix with
First piece of correspondence of constant matrices, as first group of calculating matrix.
It should be noted that the selection of host process can be defaulted as first process in KNL platforms, it is also possible to by user
Specified.
Step S12:L group calculating matrix are assigned to N number of from process, each calculating from process to receiving by host process
Matrix carries out successive ignition calculating, obtains final result;Wherein, N is the positive integer less than or equal to L.
Specifically, host process L group calculating matrix are assigned to it is N number of from process, it is to be understood that when calculating matrix compared with
When few, can open be equal to calculating matrix quantity from process, it is adjustable due to opening quantity from process, while to avoid system
The waste of resource, therefore be not in the situation from number of processes more than calculating matrix quantity, but due to being drawn to system of linear equations
Divide method difference or system of linear equations scale excessive, and process sum is limited to hardware condition, therefore, it is possible to calculating square occur
Battle array quantity is more than the situation from number of processes.
Wherein, each carries out successive ignition calculating from process to the calculating matrix for receiving, and obtains final result, due to,
Coefficient matrix and constant matrices have been divided into L group calculating matrix, and each process once-through operation can only obtain a current calculating matrix
Middle coefficient matrix and constant matrices are solved accordingly, i.e. local solution, and cannot obtain global solution, therefore, each process needs docking
The calculating matrix for receiving carry out successive ignition calculating, so as to obtain final result;Each is from process to the calculating matrix that receive
Any process of iterative calculation is carried out, including:The calculating matrix for receiving are calculated from process, obtains the first calculating knot
Really, and by the first result of calculation host process is fed back to, host process receives the first result of calculation from process, and by from the of process
One result of calculation is sent to other from process, i.e. each is shared to from the first result of calculation of process all from process;From
After sharing the first result of calculation between process, will again be calculated using the first result of calculation and calculating matrix, obtain new calculating knot
Really, new work of laying equal stress on is the first result of calculation, is shared, and until calculating all of solution, obtains final result.
Step S13:Host process receives the final result that each is calculated from process.
It can be seen that, the embodiment of the present invention is based on KNL platforms and reads system of linear equations by host process, according to default division side
Method, mutually one-to-one L blocks are respectively divided into by the coefficient matrix and constant matrices of system of linear equations, are obtained L groups and are calculated square
Battle array, and L group calculating matrix are assigned to N number of from process, ensured multi-process concurrent operation, and each is from process to receiving
Calculating matrix carry out successive ignition calculating, obtain final result, using multi-process computing advantage, by system of linear equations block parallel
Calculate, improve calculating speed, realize the rapid computations to system of linear equations.
The embodiment of the invention discloses a kind of specific algebra system method for solving based on KNL platforms, relative to upper one
Embodiment, the present embodiment has made further instruction and optimization to technical scheme.It is shown in Figure 2, specifically:
Step S21:Host process reads system of linear equations, and the coefficient matrix and constant matrices of system of linear equations are pressed respectively
Row is divided into L blocks, and every piece of coefficient matrix is mutually corresponded with every piece of constant matrices, obtains L group calculating matrix, wherein, L is for just
Integer.
Specifically, it is shown in Figure 3, by the coefficient matrix and constant matrices of system of linear equations, distinguish divided by row into L
Block, every piece of coefficient matrix is mutually corresponded with every piece of constant matrices, L group calculating matrix is obtained, for example, being in system of linear equations
Matrix number and constant matrices are respectively the matrix of 10 rows, and the first row of coefficient matrix and constant matrices to the third line is divided into one
Block, fourth line and fifth line are divided into one piece, and the 6th row and the 7th row are respectively two pieces, and the 8th row to one piece of the tenth behavior will
Coefficient matrix and constant matrices are divided into 5 pieces in correspondence with each other.
Further, system of linear equations expression formula can be:Ax=b;
In formula, A is coefficient matrix, and x is solution vector matrix, and b is constant matrices.
For example, opening L process P0, P1..., PL-1, can be by coefficient matrices A and constant matrices b divided by row into L blocks, i.e. A
=[A0 T,A1 T,…,AL-1 T]T, b=[b0 T,b1 T,…,bL-1 T]T, T representing matrixs in formula.By data block A0~AL-1And b0~bL-1
It is respectively allocated to process P0~PL-1, and vector x is shared for all processes.Therefore, each process only calculates n/L unit in x
Element.
It should be noted that after dividing system of linear equations, count matrix quantity can be less than or equal to the hardware thread of KNL
Number, to avoid the occurrence of the process with competitive relation, further to improve arithmetic speed, for example, hardware thread is up to
288, then the count matrix quantity for dividing should be less than being equal to 288 groups, such as 288 groups.
Step S22:Be evenly distributed to for L group calculating matrix N number of from process and host process by host process, and each is from process pair
The calculating matrix for receiving carry out successive ignition calculating, obtain final result;Wherein, N is the positive integer less than or equal to L;
It is understood that being all the process in KNL platforms due to host process and from process, the two is substantially without area
Not, meanwhile, in order to give full play to multi-process advantage and utilize system resource, can by L group calculating matrix be evenly distributed to it is N number of from
In process and host process, host process can also participate in follow-up calculating, meanwhile, it also is responsible for the collection and distribution to result of calculation.
Further, function, inner product of vectors function and vector can be multiplied using matrix-vector multiplication function, vectorial number from process
The calculating matrix that receive of function pair are added to be calculated, can be completed by way of calling subfunction four kinds of matrixes of the above to
Amount is calculated, and is utilized the mode of " #pragma omp " speech to complete kernel and is accelerated design.
Step S23:Host process receives the final result that each is calculated from process.
It is understood that each interprocess communication mode can be collective communication, by the message passing library letter for calling MPI
Number realization, including:MPI_Reduce, MPI_ALLReduce, MPI_Bcast and MPI_Allgatherv.
Further, in order to improve calculating speed, can be by the intermediate variable in iterative process, for example, solution vector
Matrix, is saved in MCDRAM high bandwidth internal memories;At the same time it can also make KNL platforms include multiple KNL chips, while parallel
Calculating system of linear equations, further improve calculating speed.
Accordingly, the embodiment of the invention also discloses a kind of algebra system solving system based on KNL platforms, referring to Fig. 3
Shown, the system includes:
Division module 11, reads system of linear equations, according to default division methods, by system of linear equations for host process
Coefficient matrix and constant matrices are respectively divided into mutually one-to-one L blocks, obtain L group calculating matrix, wherein, L is positive integer;
, be assigned to for L group calculating matrix for host process N number of from process by distribute module 12;
The calculating matrix for receiving are carried out successive ignition calculating by computing module 13 for each from process, obtain final
As a result;Wherein, N is the positive integer less than or equal to L;
Collection module 14, each final result calculated from process is received for host process;
Wherein, any mistake of iterative calculation is carried out to the calculating matrix for receiving from process to each in computing module 13
Journey, including:The calculating matrix for receiving are calculated from process, obtains the first result of calculation, and the first result of calculation is anti-
Feed host process, host process receives the first result of calculation from process, and will be sent to other from the first result of calculation of process
From process.
It can be seen that, the embodiment of the present invention is based on KNL platforms and reads system of linear equations by host process, according to default division side
Method, mutually one-to-one L blocks are respectively divided into by the coefficient matrix and constant matrices of system of linear equations, are obtained L groups and are calculated square
Battle array, and L group calculating matrix are assigned to N number of from process, ensured multi-process concurrent operation, and each is from process to receiving
Calculating matrix carry out successive ignition calculating, obtain final result, using multi-process computing advantage, by system of linear equations block parallel
Calculate, improve calculating speed, realize the rapid computations to system of linear equations.
Further, above-mentioned division module 11, specifically for by the coefficient matrix and constant matrices of system of linear equations, distinguishing
Divided by row is mutually corresponded into L blocks, every piece of coefficient matrix with every piece of constant matrices.
Above-mentioned computing module 13, including computing unit;Wherein,
Computing unit, for multiplying function, inner product of vectors function and vector using matrix-vector multiplication function, vectorial number from process
The calculating matrix that addition function pair is received are calculated.
, be evenly distributed to for L group calculating matrix specifically for host process N number of from process by above-mentioned distribute module 12.
In the embodiment of the present invention, above-mentioned distribute module 12 can also include allocation unit;Wherein,
, be assigned to for L group calculating matrix for host process N number of from process and host process by allocation unit.
It should be noted that in the embodiment of the present invention, each interprocess communication mode can be collective communication.
Finally, in addition it is also necessary to explanation, herein, such as first and second or the like relational terms be used merely to by
One entity or operation make a distinction with another entity or operation, and not necessarily require or imply these entities or operation
Between there is any this actual relation or order.And, term " including ", "comprising" or its any other variant meaning
Covering including for nonexcludability, so that process, method, article or equipment including a series of key elements not only include that
A little key elements, but also other key elements including being not expressly set out, or also include for this process, method, article or
The intrinsic key element of equipment.In the absence of more restrictions, the key element limited by sentence "including a ...", does not arrange
Except also there is other identical element in the process including the key element, method, article or equipment.
A kind of algebra system method for solving and system based on KNL platforms provided by the present invention have been carried out in detail above
Introduce, specific case used herein is set forth to principle of the invention and implementation method, the explanation of above example
It is only intended to help and understands the method for the present invention and its core concept;Simultaneously for those of ordinary skill in the art, according to this
The thought of invention, be will change in specific embodiments and applications, and in sum, this specification content should not
It is interpreted as limitation of the present invention.
Claims (10)
1. a kind of algebra system method for solving based on KNL platforms, it is characterised in that including:
Host process reads system of linear equations, according to default division methods, by the coefficient matrix and constant of the system of linear equations
Matrix is respectively divided into mutually one-to-one L blocks, obtains L group calculating matrix, wherein, L is positive integer;
Be assigned to for L group calculating matrix N number of from process by the host process, and each is carried out from process to the calculating matrix for receiving
Successive ignition is calculated, and obtains final result;Wherein, N is the positive integer less than or equal to L;
The host process receives the final result that each is calculated from process;
Wherein, each carries out any process of iterative calculation to the calculating matrix for receiving from process, including:From process docking
The calculating matrix for receiving are calculated, and obtain the first result of calculation, and first result of calculation is fed back into the host process,
The host process receives first result of calculation from process, and is sent to other from the first result of calculation of process by described
From process.
2. the algebra system method for solving based on KNL platforms according to claim 1, it is characterised in that described according to pre-
If division methods, it is described that the coefficient matrix and constant matrices of system of linear equations are respectively divided into mutually one-to-one L blocks
Process, including:
By the coefficient matrix and constant matrices of the system of linear equations, divided by row is into L blocks respectively, every piece of coefficient matrix with every piece
Constant matrices is mutually corresponded.
3. the algebra system method for solving based on KNL platforms according to claim 1, it is characterised in that described from process
The process that the calculating matrix for receiving are calculated, including:
It is described to multiply the docking of function, inner product of vectors function and addition of vectors function using matrix-vector multiplication function, vectorial number from process
The calculating matrix for receiving are calculated.
4. the algebra system method for solving based on KNL platforms according to claim 1, it is characterised in that lead between each process
Letter mode is collective communication.
5. the algebra system method for solving based on KNL platforms according to claim 1, it is characterised in that also include:
Be assigned to for L group calculating matrix N number of from process and host process by the host process.
6. the algebra system method for solving based on KNL platforms according to any one of claim 1 to 5, it is characterised in that institute
State the host process and L group calculating matrix are assigned to N number of process from process, including:
Be evenly distributed to for L group calculating matrix N number of from process by the host process.
7. a kind of algebra system solving system based on KNL platforms, it is characterised in that including:
Division module, system of linear equations is read for host process, according to default division methods, is by the system of linear equations
Matrix number and constant matrices are respectively divided into mutually one-to-one L blocks, obtain L group calculating matrix, wherein, L is positive integer;
, be assigned to for L group calculating matrix for the host process N number of from process by distribute module;
The calculating matrix for receiving are carried out successive ignition calculating by computing module for each from process, obtain final result;Its
In, N is the positive integer less than or equal to L;
Collection module, each final result calculated from process is received for the host process;
Wherein, any mistake of iterative calculation is carried out to the calculating matrix for receiving from process to each in the computing module
Journey, including:The calculating matrix for receiving are calculated from process, obtains the first result of calculation, and knot is calculated by described first
Fruit feeds back to the host process, and the host process receives first result of calculation from process, and by described from the of process
One result of calculation is sent to other from process.
8. the algebra system solving system based on KNL platforms according to claim 7, it is characterised in that the division mould
Block, specifically for by the coefficient matrix and constant matrices of the system of linear equations, distinguishing divided by row into L blocks, every piece of coefficient square
Battle array is mutually corresponded with every piece of constant matrices.
9. the algebra system solving system based on KNL platforms according to claim 7, it is characterised in that the calculating mould
Block, including:
Computing unit, function, inner product of vectors function and vector are multiplied for described from process using matrix-vector multiplication function, vectorial number
The calculating matrix that addition function pair is received are calculated.
10. the algebra system solving system based on KNL platforms according to any one of claim 7 to 9, it is characterised in that
, be evenly distributed to for L group calculating matrix specifically for the host process N number of from process by the distribute module.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710133888.XA CN106897163A (en) | 2017-03-08 | 2017-03-08 | A kind of algebra system method for solving and system based on KNL platforms |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710133888.XA CN106897163A (en) | 2017-03-08 | 2017-03-08 | A kind of algebra system method for solving and system based on KNL platforms |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106897163A true CN106897163A (en) | 2017-06-27 |
Family
ID=59185834
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710133888.XA Pending CN106897163A (en) | 2017-03-08 | 2017-03-08 | A kind of algebra system method for solving and system based on KNL platforms |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106897163A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110780842A (en) * | 2019-10-25 | 2020-02-11 | 无锡恒鼎超级计算中心有限公司 | Parallel optimization method for ship three-dimensional acoustic-elastic simulation calculation based on Shenwei architecture |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2006164307A (en) * | 2006-02-08 | 2006-06-22 | Fujitsu Ltd | Parallel processing device and method of simultaneous equation using various matrix storage methods |
CN105045768A (en) * | 2015-09-01 | 2015-11-11 | 浪潮(北京)电子信息产业有限公司 | Method and system for achieving GMRES algorithm |
CN105260342A (en) * | 2015-09-22 | 2016-01-20 | 浪潮(北京)电子信息产业有限公司 | Solving method and system for symmetric positive definite linear equation set |
CN106407561A (en) * | 2016-09-19 | 2017-02-15 | 复旦大学 | A division method of the parallel GPDT algorithm on a multi-core SOC |
-
2017
- 2017-03-08 CN CN201710133888.XA patent/CN106897163A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2006164307A (en) * | 2006-02-08 | 2006-06-22 | Fujitsu Ltd | Parallel processing device and method of simultaneous equation using various matrix storage methods |
CN105045768A (en) * | 2015-09-01 | 2015-11-11 | 浪潮(北京)电子信息产业有限公司 | Method and system for achieving GMRES algorithm |
CN105260342A (en) * | 2015-09-22 | 2016-01-20 | 浪潮(北京)电子信息产业有限公司 | Solving method and system for symmetric positive definite linear equation set |
CN106407561A (en) * | 2016-09-19 | 2017-02-15 | 复旦大学 | A division method of the parallel GPDT algorithm on a multi-core SOC |
Non-Patent Citations (2)
Title |
---|
DOUNIA KHALDI 等: "Towards Automatic HBM Allocation using LLVM: A Case Study with Knights Landing", 《2016 THIRD WORKSHOP ON THE LLVM COMPILER INFRASTRUCTURE IN HPC》 * |
蔺小林 等: "线性矩阵方程的迭代求解方法", 《陕西科技大学学报》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110780842A (en) * | 2019-10-25 | 2020-02-11 | 无锡恒鼎超级计算中心有限公司 | Parallel optimization method for ship three-dimensional acoustic-elastic simulation calculation based on Shenwei architecture |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN101706741B (en) | Method for partitioning dynamic tasks of CPU and GPU based on load balance | |
CN111626414B (en) | Dynamic multi-precision neural network acceleration unit | |
Bailey | Extra high speed matrix multiplication on the Cray-2 | |
CN105528191B (en) | Data accumulation apparatus and method, and digital signal processing device | |
CN103049241B (en) | A kind of method improving CPU+GPU isomery device calculated performance | |
Aharonson et al. | Counting networks with arbitrary fan-out | |
CN103617150A (en) | GPU (graphic processing unit) based parallel power flow calculation system and method for large-scale power system | |
CN103927231B (en) | The energy optimization data set distribution method that a kind of data-oriented processes | |
GB1537504A (en) | Network computer system | |
CN102110079B (en) | Tuning calculation method of distributed conjugate gradient method based on MPI | |
CN106681688A (en) | Set similarity calculation method and system based on minhash | |
CN114201287B (en) | Method for cooperatively processing data based on CPU + GPU heterogeneous platform | |
CN102567080A (en) | Virtual machine position selection system facing load balance in cloud computation environment | |
CN105045670A (en) | Method and system for balancing loads of central processing units and graphic processing units | |
CN104572587A (en) | Data matrix multiplying acceleration computing method and device | |
CN112418396A (en) | Sparse activation perception type neural network accelerator based on FPGA | |
CN104484234A (en) | Multi-front load flow calculation method and system based on GPU (graphics processing unit) | |
WO2021036729A1 (en) | Matrix computation method, computation device, and processor | |
CN105260342A (en) | Solving method and system for symmetric positive definite linear equation set | |
CN106528490A (en) | FPGA (Field Programmable Gate Array) heterogeneous accelerated computing device and system | |
CN111428192A (en) | Method and system for optimizing high performance computational architecture sparse matrix vector multiplication | |
CN106897163A (en) | A kind of algebra system method for solving and system based on KNL platforms | |
CN104156271A (en) | Method and system for balancing cooperative computing cluster load | |
CN112560356A (en) | Sparse matrix vector multiply many-core optimization method for many-core architecture | |
CN103455518A (en) | Data processing method and device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170627 |