CN110262995A - It executes body creation system and executes body creation method - Google Patents
It executes body creation system and executes body creation method Download PDFInfo
- Publication number
- CN110262995A CN110262995A CN201910633638.1A CN201910633638A CN110262995A CN 110262995 A CN110262995 A CN 110262995A CN 201910633638 A CN201910633638 A CN 201910633638A CN 110262995 A CN110262995 A CN 110262995A
- Authority
- CN
- China
- Prior art keywords
- task
- node
- data
- execution
- executes
- 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 OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/16—Combinations of two or more digital computers each having at least an arithmetic unit, a program unit and a register, e.g. for a simultaneous processing of several programs
- G06F15/163—Interprocessor communication
- G06F15/173—Interprocessor communication using an interconnection network, e.g. matrix, shuffle, pyramid, star, snowflake
- G06F15/17306—Intercommunication techniques
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/16—Combinations of two or more digital computers each having at least an arithmetic unit, a program unit and a register, e.g. for a simultaneous processing of several programs
- G06F15/163—Interprocessor communication
- G06F15/173—Interprocessor communication using an interconnection network, e.g. matrix, shuffle, pyramid, star, snowflake
- G06F15/17356—Indirect interconnection networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- Computer Hardware Design (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Neurology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present disclosure discloses a kind of execution bodies to create system, it include: task topological diagram formation component, for the breakdown of operation of required completion to be performed task by executing body in isomery framework to be a series of, and while breakdown of operation, based on the intrinsic relationship between being decomposed for task, generate task nexus topological diagram, the task node of the task topological diagram contains the whole nodal communities for executing corresponding required by task, the task topological diagram formation component includes specified constituent element node selecting unit, for having the specific tasks node of the specified constituent element comprising additional node attributes according to specific tasks selection during generating task nexus topological diagram, the additional node attributes change the mode that the specific tasks node executes corresponding task；And it executes body and creates component, whole nodal communities of task based access control relationship topology figure and each task node or whole nodal communities containing specified constituent element are that each task node creates corresponding execution body, and makes specific executions body corresponding with there is the specific tasks node of additional node attributes according to the corresponding task of additional node attributes execution in computing resource.
Description
Technical field
This disclosure relates to the system that memory space carries out fine-grained management and hardware optimization in a kind of pair of data processing network
And more specifically method is related to executing body creation system in a kind of data processing network and executes body creation method.
Background technique
With the development and artificial neural network of machine learning research gradually deeply, the concept of deep learning obtains
Extensive concern and application.Deep learning is a kind of special machine learning, it is learnt using netted hierarchical structure to express
Object, abstract concept is combined by simple concept, realized by simple concept calculating abstract concept express.Mesh
Before, deep learning has considerable progress in image recognition, speech recognition and natural language processing field.Deep learning is related to
Model parameter it is more, cause calculation amount huge, and the scale of training data is big, it is therefore desirable to consume more computing resource etc.
Feature.
As big data calculating and the rise of deep learning, various coprocessors are normally used for sharing the data of CPU
Processing function.Such as GPU (Graphic Processing Unit), APU etc..GPU has high parallel organization (highly
Parallel structure), so GPU possesses efficiency more higher than CPU in terms of processing graph data and complicated algorithm.
When CPU executes calculating task, a moment only handles a data, parallel there is no truly, and GPU have it is multiple
Processor core, can be with the multiple data of parallel processing a moment.Compared with CPU, GPU possesses more ALU (Arithmetic
Logic Unit, logical operation execute body) it is used for data processing, rather than data high-speed caches and flow control.Such structure is non-
It is very suitable for large-scale data for type high unity, mutually without dependence and does not need the pure calculating ring being interrupted
Border.
Existing big data calculate and deep learning network system by make a reservation for each operation execute the operating function of body come
Continuous data processing is carried out, therefore, network once launches into the runtime, executes body and just needs according to predetermined operation to data
Block is operated and handled, and the operation for executing running body and data block all immobilizes.It will lead to and held certain in this way
The wasting of resources in the treatment process of data block at row body results even in operating process since upstream-downstream relationship is to data block
It is vacant excessively it is slow influence downstream execute body operation.
Accordingly it is desirable to have a kind of data processing network can to the specific relationship executed between body and data block into
Row is adjusted, to improve the data processing speed of overall data process network and improve the recycling rate of waterused for executing body.
Summary of the invention
Since big data and deep learning are since handled data are with type high unity type, to provide one
Kind can eliminate the above-mentioned problems in the prior art and provide possibility.The disclosure is designed to provide a kind of execution body wound
Build system, comprising: task topological diagram formation component, for being a series of by being held in isomery framework by the breakdown of operation of required completion
Row body is performed task, and while breakdown of operation, based on the intrinsic relationship between being decomposed for task, generates task and closes
It is topological diagram, the task node of the task topological diagram contains the whole nodal communities for executing corresponding required by task, the task
Topological diagram formation component includes specified constituent element node selecting unit, is used for during generating task nexus topological diagram according to specific
Task choosing has the specific tasks node of the specified constituent element comprising additional node attributes, and the additional node attributes change the tool
Body task node executes the mode of corresponding task；And execute body and create component, task based access control relationship topology figure and each appoint
Whole nodal communities of business node or whole nodal communities containing specified constituent element are each task node wound in computing resource
Corresponding execution body is built, and makes specific execution body corresponding with there is the specific tasks node of additional node attributes according to building-out section
Point attribute executes corresponding task.
System is created according to the execution body of the disclosure, wherein the additional node attributes that the designated groups member includes are data behaviour
Make nodal community, for so that the specific execution body created only executes operation to the header data of data processed.
System is created according to the execution body of the disclosure, wherein the additional node attributes that the designated groups member includes are deposited for data
Nodal community is stored up, for so that the specific execution body created only stores the header data of data processed.
System is created according to the execution body of the disclosure, wherein the additional node attributes that the designated groups member includes are repaired for data
Change nodal community, for enabling the specific execution body created to modify the header data of data processed.
A kind of execution body creation method another aspect of the present disclosure provides, comprising: by the operation of required completion
Be decomposed into it is a series of be performed task by executing body in isomery framework, and while breakdown of operation, appointed based on what is decomposed
Intrinsic relationship between business, generates task nexus topological diagram, and the task node of the task topological diagram contains the corresponding task of execution
Required whole nodal communities；During generating task nexus topological diagram, is had according to specific tasks selection and include building-out section
The specific tasks node of the specified constituent element of point attribute, the additional node attributes change the specific tasks node and execute corresponding task
Mode；And task based access control relationship topology figure and each task node whole nodal communities or contain the complete of specified constituent element
Portion's nodal community is that each task node creates corresponding execution body, and makes and have additional node attributes in computing resource
The corresponding specific execution body of specific tasks node executes corresponding task according to additional node attributes.
According to the execution body creation method of the disclosure, wherein the additional node attributes that the designated groups member includes are data behaviour
Make nodal community, for so that the specific execution body created only executes operation to the header data of data processed.
According to the execution body creation method of the disclosure, wherein the additional node attributes that the designated groups member includes are deposited for data
Nodal community is stored up, for so that the specific execution body created only stores the header data of data processed.
According to the execution body creation method of the disclosure, wherein the additional node attributes that the designated groups member includes are repaired for data
Change nodal community, for enabling the specific execution body created to modify the header data of data processed.
By specifying the selection of constituent element node selecting unit that there are the specific tasks of the specified constituent element comprising additional node attributes
Node enables to execute body to the operation model of data block based on whole nodal communities creation comprising node adeditive attribute
Enclose, mode of operation and storage mode make variation corresponding to nodal community so that executing body has more flexible operation
Function to realize the fine-grained management in static data processing network to memory space, and makes the execution body of each node
Hardware is optimized.By specifying the selection of constituent element node selecting unit to have, specified constituent element includes plus the specific of nodal community is appointed
Business node allows corresponding specific execution body to modify the header data of data block, so that executing body can surpass
The build-in attribute of data block itself carries out the operation processing of richer multiplicity to data block out, handles net so as to growth data
Diversified processing requirement of the network to data block.
A part of the further advantage of the disclosure, target and feature will be emerged from by following explanation, another part
It will be understood by the person skilled in the art by the research and practice to the disclosure.
Detailed description of the invention
Shown in FIG. 1 is according to the disclosure for the structural representation for executing body creation system in data processing network
Figure；
Shown in Fig. 2 is to execute to specify the selection of constituent element node in body creation system in the data processing network according to the disclosure
Unit is on the specific schematic illustration for executing body and influencing.
Fig. 3 is execution body that the execution body creation system in data processing network according to the disclosure is created shown in being
The structural schematic diagram of network.
Specific embodiment
The disclosure is described in further detail below with reference to embodiment and attached drawing, to enable those skilled in the art's reference
Specification word can be implemented accordingly.
Example embodiments are described in detail here, and the example is illustrated in the accompanying drawings.Following description is related to
When attached drawing, unless otherwise indicated, the same numbers in different drawings indicate the same or similar elements.Following exemplary embodiment
Described in embodiment do not represent all implementations consistent with this disclosure.On the contrary, they be only with it is such as appended
The example of the consistent device and method of some aspects be described in detail in claims, the disclosure.
It is only to be not intended to be limiting and originally open merely for for the purpose of describing particular embodiments in the term that the disclosure uses.?
The "an" of singular used in disclosure and the accompanying claims book, " described " and "the" are also intended to including most shapes
Formula, unless the context clearly indicates other meaning.It is also understood that term "and/or" used herein refers to and includes
One or more associated any or all of project listed may combine.
It will be appreciated that though various information, but this may be described using term first, second, third, etc. in the disclosure
A little information should not necessarily be limited by these terms.These terms are only used to for same type of information being distinguished from each other out.For example, not departing from
In the case where disclosure range, hereinafter, one of two possible equipment can be referred to as the first execution body or be referred to as
Second executes body, and similarly, another of two possible equipment can be referred to as the second execution body or be referred to as first and hold
Row body.Depending on context, word as used in this " if " can be construed to " ... when " or " when ... "
Or " in response to determination ".
In order to make those skilled in the art more fully understand the disclosure, with reference to the accompanying drawings and detailed description to this public affairs
It opens and is described in further detail.
Shown in FIG. 1 is according to the disclosure for the structural representation for executing body creation system in data processing network
Figure.As shown in Figure 1, execute body creation system be arranged in isomery framework, the isomery framework by CPU00 and CPU01 respectively with CPU00
Connected GPU00, GPU01 and GPU2 and GPU10, GPU11 and GPU12 for being connected with CPU01 are constituted.Although herein only
Show two CPU and six GPU, but isomery framework may include more CPU, and the GPU being connected with each CPU
Can be more or less, this can be determined based on actual needs.
System 100 is created according to the execution body of the disclosure to be deployed in isomery framework shown in FIG. 1.What although Fig. 1 was shown
The composition part for executing body creation system is shown separately in except each CPU and GPU, this is to highlight and facilitate description
The processing of progress.The composition part of actually execution body creation system is all distributed among CPU and/or GPU.
As shown in Figure 1, the execution body creation system 100 includes task topological diagram formation component 120 and execution body wound
Build component 130.Executing volume grid component 140 is the creation result networking component for executing body creation component 130.
As shown in Figure 1, task topological diagram formation component 120 is used to the breakdown of operation of required completion be a series of by isomery
Body is executed in framework and is performed task, and while carrying out breakdown of operation, based on the intrinsic pass between being decomposed for task
System generates task nexus topological diagram.Executing body creation system 100 is set up to handle the work data of predefined type,
In order to continuously continuously handle the data of same type, need to become breakdown of operation into the arithmetic element for being suitble to CPU or GPU
Execute the simple task of operation or other operations.Specifically, being exactly that breakdown of operation is become being associated with each other for task.It is described
Task topological diagram formation component 120 includes decomposition to data block and to the decomposition of data processing model to the decomposition of operation,
It to the decomposition of operation is arranged to by the isolation of work data to be processed.Specifically, being wanted according to job task
The description asked carries out hierarchicabstract decomposition into multilayer neural network structure to operation according to by process to be processed.One operation
(Job) a series of tasks (Task) to interdepend are broken down into, this dependence usually uses directed acyclic graph (Directed
Acyclic graph, DAG) describe, each node indicates a task, the connecting line between node indicate data according to
Rely relationship (producers and consumers' relationship).The situation of task relational graph after breakdown of operation is not specifically described herein.
While gradually apportioned effort, task topological diagram formation component 120 also successively forms task nexus topological diagram.By
In breakdown of operation formed each task between there are intrinsic logical relations, therefore, with operation be broken down into it is different
Task, on different task layers, task topological diagram formation component 120 is also subsequently formed task nexus topological diagram, these tasks
Relationship topology figure forms the neural network between decomposed task.In the case where operation complexity, task nexus topological diagram
Include multilayer, therefore also forms multilayer task neural network.Every layer of neural network had both included the nerve of corresponding specific tasks
First node, also comprising relationship between each neuron, and both comprising appointing for the following processing that will be used for fragment data
The data parallel network of business also includes the model parallel network that will be used for the task of fragment model.Selectively, these nerves
It can also only include data parallel network in network.Whether simultaneously comprising data parallel network and model parallel network, Ke Yigen
It is configured according to actual needs.
In order to disposably execute body to the arbitrary node creation of task topological diagram in the subsequent body creation component that executes, according to
The task topological diagram formation component 120 of the disclosure each node for generating task topological diagram simultaneously, assign each node and execute
Whole nodal communities of corresponding required by task.The whole nodal community, which contains, such as indicates required by task corresponding to node
Resource Resource Properties and trigger task execution the conditional attribute of trigger condition etc..The whole nodal community also wraps
The additional node attributes that will be mentioned below are contained.Whole is contained just because of each node in the task topological diagram of the disclosure
Nodal community, therefore it has all resources of execution task and all properties when subsequent creation executes body automatically immediately,
In complete configuration status, do not need to carry out such as carrying out dynamic point to environmental resource when executing specific tasks to specific data
Match and dynamic configuration trigger condition etc..For the task topological diagram based on the disclosure and the section containing whole nodal communities
It for each execution body that point is created, is treated in journey to specific data, itself is in static state, variation
The only difference of input data.The node of neural network for being currently used for the execution body creation system of deep learning included
Nodal community is considerably less or does not have substantially, thus node correspond to task execution needs temporarily pushed away in specific tasks implementation procedure
Attribute needed for export completes corresponding task to dynamic acquisition corresponding attribute.And this attribute needle temporarily derived
Same task is required temporarily to derive every time, therefore may require that a large amount of computing overhead.
It should be pointed out that task topological diagram formation component 120, which exists, successively forms task nexus topological diagram simultaneously, need
The task nexus topological diagram formed is optimized.Therefore it is also wrapped according to the task topological diagram formation component 120 of the disclosure
Include topological diagram optimization component 121.The topological diagram optimization component 121 includes various optimization units, such as redundant node eliminates list
Member 1211, obstruction node eliminate equivalent subgraph converter unit 1213 as unit 1212 and specified constituent element node selecting unit
And other are used to optimize the unit 1214 of topological diagram.Although display contains above three unit in Fig. 1 of the disclosure,
It is that be not offered as the disclosure must all include these units.The realization of the disclosure does not need centainly to include above topology optimization group
Part 121.The presence of topological optimization component 121, will so that the task topological diagram generated of task topological diagram formation component 120 more
Rationally, and more smooth by what is run in subsequent data handling procedure, treatment effeciency is higher.
Specifically, may exist and be directed to during task topological diagram formation component 120 generates task topological diagram
A certain task duplicates the case where generating corresponding node.For example, in a neural network subgraph, it is possible that two
Node arranged side by side, the two upstream node having the same and identical downstream node, and the corresponding same task.Such section
Point is exactly redundant node situation.Calculation resources in the presence meeting repeat consumption isomery framework of this redundant node, so that neural
Network complicates.Therefore this redundant node is the node for needing to remove.It is opened up in 120 generation task of task topological diagram formation component
If it find that this duplicate node, redundant node, which eliminates unit 1211, can know the presence of this node simultaneously during flutterring figure
The redundant node is directly deleted, so that being only associated with the upstream and downstream node of the redundant node and to deleted redundant node phase
The upstream and downstream node of same node (node of same task is executed with redundant node).In addition, in task topological diagram formation component
During 120 generate task topological diagram, may exist for the interaction between certain tasks, it can be due to task processing
Occurs downstream node stopping state not in time, so that the congestion situations that will lead to the node being blocked are conducted forward.For this purpose, in office
Be engaged in topological diagram formation component 120 generate task topological diagram during if it find that this obstruction node, obstruction node are eliminated
Unit 1212 can eliminate the node for causing operation to block in task topological diagram.Specifically, be exactly change obstruction node with it is upper
The connection side between node is swum, one or more nodes are increased, eliminates conduction of the obstruction to upstream at obstruction node.
During task topological diagram formation component 120 generates task topological diagram, in order to execute body with more flexible
Operating function, to realize fine-grained management to memory space in static data processing network, and make each node
It executes body hardware to be optimized, specifies constituent element node selecting unit 1213 during generating task nexus topological diagram according to specific
Task choosing has the specific tasks node of the specified constituent element comprising additional node attributes, and the additional node attributes change the tool
Body task node executes the mode of corresponding task, for example, the range for executing body processing task data of the corresponding task node of change,
Change the range of output data storage, eliminate primary unnecessary backward operation, change output data modify attribute and
Change and executes the frequency etc. that body executes task.
Although front only list in detail such as redundant node eliminate unit 1211, obstruction node eliminate unit 1212 and
The unit of three kinds of topological diagrams optimization of specified constituent element node selecting unit 1213, but optimize unit for the topological diagram of the disclosure
It is very more, it does not describe one by one herein.In addition, during task topological diagram formation component 120 generates task topological diagram, it may
There can be the situation more complicated or inefficient for the network subgraph of certain associated tasks generation.In order to obtain more efficient task
Topological diagram, task topological diagram formation component 120 are understood to the multiple network subgraphs generated to certain associated tasks, thus, it is desirable to
Equivalent transformation is carried out to each drawing of seeds in topological diagram optimization component 121, thus from multiple sons that can complete identical operation function
The subgraph network of highest operation efficiency is selected to substitute current subgraph network in figure network.Although open elaborate above topology figure
The various optimization units of optimization component 121, it is also possible to include other any optimization units, such as shown in Fig. 1 other
Unit 1214.
After task topological diagram formation component 120 generates each layer task neural network topological diagram, executes body and create component
130 task based access control relationship topology figures are the corresponding execution body of each task creation in the computing resource that isomery framework is included.
Specifically, according to hardware resource needed for task description, whole nodal communities based on each node are each task
It specifies the arithmetic element of respective numbers and corresponding storage unit to constitute in isomery framework and executes body to execute corresponding task.
The execution body created includes the various resources in the computing resource in isomery framework, for example, storage unit, message send or
Receiving unit, arithmetic element etc..The arithmetic element for executing body may may also be comprising multiple, as long as it can be completed for one
Specified task.Body is being executed after being created, the executing specified always of the task will not be changed, held except non-required
Capable task disappears, such as isomery framework belonging to the execution body is applied to the processing of other types operation again.It is created
Execution body between the cyberrelationship that is formed it is corresponding with the relationship between each neural network node in task topological diagram from
And form execution volume grid component 140 shown in Fig. 1.The each execution body for constituting execution volume grid component 140 is distributed in
In one or more CPU of composition isomery framework and the coprocessor being connected with CPU.Coprocessor GPU, TPU etc..Such as figure
Shown schematically in 1, replaced in each execution body for executing volume grid component 140 using various small circles.Some small circles
It is cascaded to form data handling path each other by dotted line.There can be some branches in one data processing path.Two
Or more can have the data handling path for crossing each other to form an increasingly complex relationship between data processing path.These
Data handling path will remain constant on the isomery framework.
Volume grid component 140 is executed when receiving actual job data, actual job data fragmentation is become into task data,
The task data is by continuous input data processing path, to complete the processing of task data.It is continuous to input for image
Data in homogeneous data fragment will be fixed and be input in same data handling path, the data fragmentation inputted is to flowing water
It equally inputs, the data for sequentially entering the Data entries of same data handling path, and generating after treatment are also sent out automatically
It is sent to back to back downstream in data handling path and executes body, until flowing through entire data handling path.
Shown in Fig. 2 is to execute to specify the selection of constituent element node in body creation system in the data processing network according to the disclosure
Unit is on the specific schematic illustration for executing body and influencing.As shown in Fig. 2, calculating the system with deep learning for big data
In, data processing network is made of various execution bodies, for the convenience of description, only showing 12 in Fig. 2, is marked respectively
For 21,22 ... 32.In practical application scene, executes body and be based on needing be any amount.These execute body to input
Data block executes scheduled operation.Although carrying out showing three data blocks in Fig. 2, in actual scene, the quantity of data block is
Magnanimity.
As shown in Fig. 2, executing scheduled operation to externally input data block 1-3 in the execution body 21-29 of ellipse
Processing.Executing body is usually the arithmetic element in data processing equipment, such as arithmetic element or an operation in GPU
Component.Executing one of body 21-29 and can only receiving has a data input, also can receive multiple data inputs.Some are executed
Body can not need to input any content-data.
In deep learning data processing network, execute body the initial stage based on determining data processing type respectively by
Assign scheduled processing task.It is each to execute body in fixed reception from upstream execution body output with the inflow of data block
Generated data block is simultaneously exported or is output to and execute body downstream by data block.
In the data processing network of the disclosure, data block includes header data and content-data, and following table 1 gives one
Kind block data structure table.
Table 1
As shown in Table 1, header data has determined that the metadata contained by header data describes the content number of data block
According to particular content, and guided the specific location in given memory space of content-data.By default,
Execution body is to the operation that the operation of data block is to header data and content-data entirety in data block.But specific
In data handling procedure, in order to realize a variety of different purposes and eliminate problem in some static datas processing networks, need pair
Data block carries out selective operation, changes to the mode of operation of data block or modify to data block itself.
In specific data handling procedure, due to specified constituent element node selecting unit 1213 can for scheduled data block or
Task section of the scheduled specific tasks node executed in topological diagram corresponding to body 20 from node repository selection containing specified constituent element
Point, therefore these task nodes just have some specified nodal communities.Hereinafter, in the case where not particularly pointing out, appoint
Business node is usually corresponding with body is executed.If not needing clearly to execute body to be specifically to execute which of body 21-29
When executing body, it is used uniformly the replacement of label 20, unless it is necessary to particularly point out.
In a kind of scene, specify constituent element node selecting unit 1213 that can select tool for a certain task node for executing body
There is the task node of specified constituent element or additional node attributes, so that the corresponding execution body of the task node specifies the execution body
Use the range of data block.For example, data block used in the disclosure includes header data or content-data.Specified constituent element node
The selected task node of selecting unit 1213 has so that the data for executing the generation of body 22 are only used only in the predetermined body 25 that executes
The header data of block 1-2.In data processing network, since specified constituent element node selecting unit 1213 has selected and executes body 25
Corresponding task node has the specified constituent element that header data is only used only, will be from scheduled routine operation side thereby executing body 25
Formula is changed to the new mode of operation after obtaining specified constituent element, such as the head for executing the data block 1-2 that body 22 generates only is used only
Portion's data.Stated differently, since corresponding task node has specified constituent element, therefore that execute the consumption of body 25 is data block 1-2
Header data.Therefore, the occurrence that body 25 only needs the header data of data block 1-2 is executed, it is not necessary that obtain and execute the guarantor of body 22
The content-data of deposit data block 1-2.This will obviously accelerate data from body 22 is executed to the flowing for executing body 25.
In static data processing network, the consumer that body is data is executed, may simultaneously be also the producer of data.Cause
This it is multiple execute bodies constitute data processing networks in, data block flows between different executions bodies, through execution body handle or
Become a new data block after consumption, thus the operation subscribed for next execution body.As described in below for Fig. 3,
Downstream executes body after consuming or having used upstream to execute the data block that body exports, and can upstream execute body feedback and use completion
Message new need to handle to receive so that upstream executes memory space occupied by vacant the exported data block of body
Data block.In actual scene, it may need multiple input blocks that could execute since some execute the operation that body executes
Complete scheduled operation.This is likely to occur, after receiving the first input block, due to second or third data block
There are no receiving therefore, the case where which cannot be immediately performed operation, therefore the output of the first input block executes
Body executes the feedback message of body by that cannot obtain multi input quickly, so that the output execution body of the first input block cannot be vacant
For storing the memory space (such as the output data caching for executing body) of first input block, thus also cannot be into one
It walks to upper level and executes body sending feedback message.This blocking situation that will lead to data flow go ahead level-one conduction.It is specific and
Speech, as shown in Fig. 2, needing to obtain the data block 2-2 from the output of body 26 is executed when execution operates for executing body 29
With the data block 3-3 exported via body 27 and execution body 28 is executed.Therefore, in executing operating process, may exist and execute
Body 29 is having received the data block 3-3 for executing body 28 and exporting at the first time, but it is defeated never to receive upstream execution body 26
Data block 2-3 out, therefore scheduled operation cannot be executed.Even if therefore executing body 29 obtains data block 3-3, due to executing
Body 29 is not carried out predetermined operation and has consumed data block 3-3, therefore upstream execution body 28 can not obtain the feedback for executing body 29
Message, also cannot memory space (cannot be written over state) occupied by vacant owned data block 3-3.In this way,
Cause to execute the confirmation message that body 28 cannot also be finished to its upstream execution sending of body 27 always, this also causes to execute body 27
Memory space occupied by middle data block 3 can not be vacant, therefore eventually leads to and execute fixed storage space bound in body 27
New data block cannot be received and carry out continuous data processing.Accordingly even when execute body 27 downstream execute body 30,31 and
In the case that 32 have completed scheduled operation, also new data block progress subsequent processing can not be obtained from body 27 is executed.To
Execution body 30,31 and 32 is caused to be also at stagnation wait state.The execution body 26 of coming is received therefore, because executing body 29 and being in
The delay of output block cause its operation to be waited for, therefore result in data processing network it is some with execute body
There is the case where data blocking on 29 relevant task nodes.In order to eliminate existing this data flow in data processing
Congestion situations.Specified constituent element node selecting unit 1213 or blocking task node, which eliminate unit, can will appear data thus to be estimated
Selection one has the task node for the specified constituent element for keeping header data before the correspondence task node of the execution body 29 of flow blocked,
Such as task node corresponding to body 28 is executed, so that the execution body 28 changes institute in the case where having the specified constituent element to place
Operative relationship between the data block of reason executes body 28 to the scheduled mode of operation of data block to change, for example, will processing
The data storage method of data block afterwards changes into new data storage method, for example, only for obtained data block
Keep its header data.Equally, head number is only kept to cooperate selected task node corresponding with body 28 is executed to have
According to specified constituent element, need the selected task node corresponding with execution body 29 of specified constituent element node selecting unit 1213 to have
The specified constituent element for executing the header data of the data generated of body 28 is only used only.Head is only saved due to executing body 28 in this way
Data, thus in the case where executing body 29 and being not carried out operation also due to specified constituent element in the presence of " only be used only header data ",
Therefore, it executes in the output data caching of body 28 and is partially in blank state for storage content data, so that executing body 28
Can to execute 27 feedback message of body, allow execute body 27 make its own output data cache in blank state so as to
In the state for being able to carry out the new data block of next round operation acquisition.Therefore, the downstream execution body 30-33 for executing body 27 can
The processing of next data block is timely entered, executes the blocking of the data flow at body 29 to the related data for executing body to eliminate
The influence of processing operation.
As shown in Fig. 2, in order to enable data block meet in by operating process it is some execute bodies operation requirements, for example,
Specified constituent element node selecting unit 1213 can issue specified constituent element to data block 2, to modify its data attribute.Such as the attribute
It has been shown that, the header data of data block 2 inform that the execution body operated on it, the execution body that can be subsequently received are repaired
Change, such as body 24 can be performed and modified, to meet the operational requirements for executing body 24.Generally, this is to input number
According to the change of the header data of block.
The task node with specified constituent element may be selected in specified constituent element node selecting unit 1213, executes body pair to change
Opereating specification, mode of operation and the storage mode of data block, this eliminates execution body and needs temporarily to be changed behaviour at runtime
Make mode.I.e. by (not needing operation when progress using priori manner selection complete task node in the task node choice phase
Modification), to eliminate overhead when operation.On the other hand, as shown in Fig. 2, in order to enable the data block of output meets downstream
The operation for executing body needs, and specified constituent element node selecting unit 1213 can choose corresponding with body is executed with can modify head
The task node of the specified constituent element of data so that the corresponding body and data block of executing is influenced by specified constituent element, such as to hold
The data block 2-3 that row body 26 exports changes its header data content, to meet the operation needs that downstream executes body 29.
In addition, as shown in Fig. 2, in order to eliminate unnecessary backward operation, specify constituent element node selecting unit 1213 can be with
Selection with execute body 22 it is corresponding have eliminate after to operation specified constituent element task node so that execution body 22 at this
Under the influence of specified constituent element, eliminates and executing at body 23 for the rear to operation of execution body 22.It is being related to deep learning system
In data processing network, to operation and contrary operation before existing mostly, these operations all defaults need to be implemented body to execute.For
Determine the forward direction operation for not needing contrary operation, such as since the specified constituent element node selecting unit 1213 of the disclosure selects and holds
Row body 22 is corresponding with, to the task node of the specified constituent element of operation, elimination executes the contrary operation that body 23 carries out after eliminating.
To be greatly saved treatment process and expense.
It can be by selecting particular task node to execute according to the specified constituent element node selecting unit 1213 of the disclosure
Body has more flexible operating function, thus realize the fine-grained management in static data processing network to memory space, and
So that the execution body hardware of each task node is optimized.It is corresponding execution body by specified constituent element node selecting unit 1213
Selection can have the task node for the specified constituent element modified to the header data of data block, so that executing body can be with logarithm
It modifies according to the header data of block, so that the build-in attribute of data block itself can be exceeded to data block progress by executing body
The operation processing of richer multiplicity, so as to growth data processing network to the diversified processing requirement of data block.Due to referring to
The presence for determining constituent element node selecting unit 1213 also brings more the writing for program for being applied to run in data processing network
More conveniences.
Although will execute body in data processing network in the description of the disclosure creates some execution bodies that system is created
It is described as two independent individuals, but selectively, being not meant to that the two separation exists is necessary to realizing the disclosure
It arranges, but can combine both.
Big data technology and deep learning are used for when executing body creation system in the data processing network according to the disclosure
Field and when constituting distributed system, the fluency of data processing seems extremely important.At the data of a task node
When blocked state, the data processing that will lead to other parts will appear pause, to will lead to whole system in data processing
Aspect pause, so that data flowing process is waited for.Basis is used in big data calculating and deep learning
Body is executed in the data processing network of the disclosure and creates system, since data block includes content-data and header data, is led to
Cross the specified selection of constituent element node selecting unit 1213 and have so that execution body operative relationship and execute body and data block
Operative relationship receives the task node of specified constituent element, thus it is possible to vary the predetermined operation relationship between body and data block is executed, thus
Fine-grained management can be carried out to memory space associated with body is executed, some memory spaces are efficiently utilized,
The hardware performance that optimization executes body improves the efficiency for executing body continuous processing data block.
Fig. 3 is execution body that the execution body creation system in data processing network according to the disclosure is created shown in being
The structural schematic diagram of network.As shown in figure 3, big dotted line frame represents an execution body.Execution volume grid component shown in Fig. 3
In 140, in order to illustrate conveniently, five execution bodies are only shown.In fact, corresponding to task topological diagram, neural network has more
Few task node there is how many execution bodies in executing volume grid component 140, therefore in the lower left side of Fig. 3 using continuous
Small closed square executes body to represent unshowned other.Fig. 3 principle shows the structure for constituting each execution body of the disclosure
At it includes have message storehouse, finite state machine, processing component and output data caching.From figure 3, it can be seen that each holding
Row body seems all to include an input data caching, but uses broken line representation.Actually this is the composition of an imagination
Component, this will be explained in detail below.Each execution body in data handling path, such as the second execution body in Fig. 3,
A task node in the neural network of task based access control topological diagram is established, and is based on complete nodal community, formed this second
Execute body and its upstream and downstream execute body topological relation, message storehouse, finite state machine and (processing component) processing mode and
Generate the cache location (output data caching) of data.Specifically, second executes body when executing data processing, task is needed
The third of its upstream is wanted to execute the output data of body.When third executes the data that body generation will be output to the second execution body, example
When such as generating third data, third executes message the disappearing to the second execution body that body will issue DSR to the second execution body
Storehouse is ceased, informs that the second execution body third data have been in third and have executed in the output data caching of body and in retrievable
State, so that the second execution body can execute the reading of the third data at any time.Second executes the finite state machine of body in message
Its state is modified after the message of storehouse acquisition third execution body.If the processing component of the second execution body has once executed operation upper
Data, such as the second data are produced after task, and are buffered in its output data caching, and the downstream for executing body to second
Body is executed, such as two first execution bodies (the first execution body A and the first execution body B) sendings can read disappearing for the second data
Breath.
When first execute body A and first execute body B read the second data and using complete after, can be held respectively to second
Row body feedback message informs that the second execution body has used second data, and therefore, the output data caching of the second execution body is in
Blank state.The finite state machine of the second execution body can also modify its state at this time.
In this way, when the state change of finite state machine reaches scheduled state, such as the execution operation of the second execution body
Required input data (such as third data) is in when can obtain state and its output data caching in blank state, then
It informs that processing component reads the third data in the output data caching of third execution body, and executes specified processor active task, from
And the output data of the execution body is generated, such as the second new data, and be stored in the output data caching of the second execution body.
Equally, after second, which executes body, completes specified processor active task, finite state machine revert to its original state, etc.
Change to next next state and recycle, while the second execution body is executed body feedback to third and arrived to third data using the message of completion
Third executes the message storehouse of body and sends the message for having generated the second data to the first execution body A and the first execution body B, accuses
Know that the first execution body A and first executes body B, the second data have been in the state that can be read.
After third, which executes body acquisition the second execution body, has used the message of third data, so that third executes the output of body
Data buffer storage is in blank state.Equally, the second execution body obtains the first execution body A and the first execution body B has used the second number
According to message after so that second execute body output data caching be in blank state.
The process of the above-mentioned execution task of second execution body executes in body at other equally to be occurred.Therefore, in each execution
Under the control of finite state machine in body, the output of body is executed as a result, the same generic task of circular treatment based on upstream.To each hold
Row body is not required to like the post personnel of the pinned task in a data processing path to form the pipeline processes of data
Want any other external instruction.
It should be pointed out that as shown in figure 3, the first execution body A and the first execution body B, the second execution body and third are held
Specified constituent element (such as the first designated groups selected by constituent element node selecting unit 1213 in task node are all designated in row body
Member, the second specified constituent element and third specify constituent element) it is influenced, to execute specified operation for specified data.Citing and
Speech, first, which executes body A and first, which executes body B, (although being exemplified as two first execution bodies herein, can only only have one
The first execution of a first execution body or more body) in the first specified constituent element be so that the first execution body is only used only on certain
Trip executes the specified constituent element of the header data of the output data of body, and it is so that third is held that third, which executes the third in body and specifies constituent element,
The 4th specified constituent element that row body can modify the specified constituent element of the header data of data and the 4th executes in body is so that the 4th
Execute the specified constituent element after body does not require to operation.It needs to specialize, although solve using the second specified constituent element herein
The characteristic of the second execution body is released, second herein executes the specified function of body not necessarily by specified constituent element node selection in fact
What unit 1213 directly determined, but by specifying the selected task section with specified constituent element of constituent element node selecting unit 1213
Determined by point, i.e., a kind of special duty node directly formed during forming task node, the task node itself
Attribute contains its function of only saving header data, so that the corresponding execution body execution of the task node is only kept
The function of header data.Other execute body.In other words, contain that corresponding task node included in body specified is executed
One of nodal community described in constituent element.But for convenience, still using the side of specified constituent element in each execution body
Formula describes this special to execute body itself contained particular community or function.
In the data storage method of the disclosure, since header data and content-data are separation storages.As above it is directed to
Described in Fig. 3, since the first execution body A and first executes the specified constituent element that body B has the header data that data are only used only,
Therefore, the second execution body of upstream only keeps the header data of data.Third is only carried due to the second execution body to hold
(header data is not stored in output data caching but deposits the header data of data in the output data caching of row body
Storage is in the CPU that coprocessor is connected, i.e., the header data separates storage with the content-data in output data caching), because
This second execution, which is known from experience to third execution body feedback, uses the message finished, therefore third execution to the content-data of third data
Space in body output data caching for content-data is vacant.Third executes only to be needed to be the in the output data caching of body
Two execution bodies keep the header data of its second data.Third execution body can be also obtained to which the upstream that third executes body executes body
Feedback message, execute body the phenomenon that blocking to eliminate third and execute the upstream of body.
First executes possessed first specified constituent element in the execution of body A and first body B executes for certain upstreams are only used only
The specified constituent element of the header data of the output data of body, thus the data in its output data caching for the second execution body,
Form the Consumption relation that header data is only used only.When the first execution body A and the first execution body B have read the second execution body
After header data in output caching, body feedback message is executed to second, so that the second execution body will be another in next cycle
One header data is transported in its output data caching.
It is the specified constituent element that can modify the header data of data that third, which executes the third in body and specifies constituent element, in order to enable
The data block of output meets the operation needs that third executes the downstream execution body (such as second execution body) of body, specifies constituent element node
Selecting unit 1213 can make third execute between body and data block, and there is third to specify constituent element, so that the head number of data block
It include " usable " metadata in, so that third, which executes body, to repair the data of its upstream execution body output
Change.
Although describing the basic composition for executing body above for Fig. 3 and the operation directly between upstream and downstream execution body being closed
System, but perhaps some processing components for executing body do not execute actual operation, but only data are moved, change
The position of parameter evidence, that is, a kind of simple carry execute body.Such as second execute body processing component only by its from its
The header data that third executes the data that body obtains is transported in its output data caching without executing acquired in body to third
Data carry out any transformation (transformation).This presence for carrying execution body, which can eliminate some execution bodies, to be caused
Execute blocking caused by blocking upstream conduct so as to cause whole data handling path upstream blocking and other branches
Processing pause.
By first execute body A and first execute body B before setting only save header data second execute a body, first
Executing the execution of body A and first body B can be known based on the message from the second execution body, when being operated to the data,
Only read the header data that the second execution body is carried and is stored in its output data caching.As the first execution body A
One of body B is executed in the case where wait other input datas with first, and second executes body, and it is only necessary to keep its head number
According to without actual storage content-data.Therefore third execute body can after the feedback message for being connected to the second execution body incite somebody to action
Third execute body output data caching is placed in blank state, thus third execute body can downstream first execute body A and
First execution body B is in when waiting blocked state and can still handle new data.Body is executed if there is no second,
Body directly is executed with third and is connected by the first execution body A and the first execution body B, then third executes body and executes body A and the first
One executes body B waiting time, it is impossible to so that (content-data is very big, can occupy very in blank state for its output data caching
Big memory), this, which causes its upstream execution body that will also stay cool, cannot execute operation.Therefore, by specifying constituent element
The selection of node selecting unit 1213 executes body A and first with first and executes the corresponding wherein comprising head number is only used only of body B
According to the first specified constituent element task node, and first execute body A and first execute body B corresponding task node before
It includes the task node for only keeping the second specified constituent element of header data that it is corresponding, which to execute body, with second for increase, so that
First, which executes body A and first, executes body B when carrying out real data processing, only obtains and executes body output data immediately upstream
Header data is in so that all upstream datas of the second execution body can make output data caching vacant in time
The state of data processing can be executed.By the component according to this acceleration data processing of the disclosure, data processing road is eliminated
Blocking in diameter, when so that there is the execution body task node being waited in the paths, upstream executes the normal of body
Run it is unaffected, thus accelerate data flowing.
Executing the execution of body A and first body B first can be generated by other execution body acquisitions and third execution body
The similar content-data of data completes its final calculating.As shown in figure 3, the 4th executes body as first and executes body A and the
One upstream for executing body B executes body, can export and execute the similar content-data of body data generated with third, in this way can be with
Solution is provided when the first execution body A and the first execution body B need content-data.In deep learning neural network often
The task node that will appear some teammate's types, the execution body for creating and describing based on the task node be it is the same, in Fig. 3
Shown in third execute body and the 4th execution body just belong to teammate's type in deep learning neural network execute body task node.This
Belong to state of the art, herein without specifically describing.
As previously mentioned, each execution body shows to include an input data caching in Fig. 3, do not include actually
, because each execution body does not need any caching to store data to be used, but used needed for only obtaining
Data are in the state that can be read.Therefore, each execution body data to be used is not in specifically in execution body
When the state of execution, data are still stored thereon trip and execute in the output data caching of body.Therefore, for image display, often
Input data caching in a execution body, which is adopted, to be represented by dashed line, and is not present in really actually and is executed in body.In other words, on
The output data caching of trip execution body is exactly the virtual input data caching that downstream executes body.Therefore, in Fig. 3, to input number
Broken line representation is used according to caching.
Referring back to Fig. 1.As shown in Figure 1, the execution body creation system for isomery framework according to the disclosure further includes
Job description component 110 is used to describe operation neural network model, the neural network number of plies and every layer of mind based on homework type
Quantity through neuron in network.Specifically, calculation resources and needs needed for job description component 110 describes operation are held
Which kind of capable operation.For example, job description is used to illustrate that the operation to be classified for image classification or speech recognition, needed for
Neural network the number of plies, every layer of task node quantity, connection between layers, execute data handling procedure in input
The storage place of data.Describing operation is a kind of prior art.The job description component 110 of the disclosure uses separation expression side
The object of required description is split into several relevant dimensions by method, and in terms of several or dimension distinguishes description, and describes
Orthogonality relation between several dimensions.Due to being retouched according to the separate mode that is distinguished from each other from different dimensions between described dimension
Operation is stated, is in orthogonality relation each other, therefore each dimension does not interfere with each other each other, to the description of task without the concern for dimension
Between association, therefore the program code run in the execution body creation system of the isomery framework of the disclosure can be substantially reducing at
Complexity, therefore also significant mitigate the intelligence burden for writing the programmer of these program codes.Although showing work in Fig. 1
Industry describes component 110.But the purpose of the disclosure also may be implemented using existing job description component.
Although as Fig. 1 show according to the disclosure for isomery framework include one or more central processing unit and
At least one coupled coprocessor device end, but in the system shown in figure 1 may include the gateway group between CPU
Part may also comprise coprocessor, such as GPU, between direct communication component, example uses dotted line to be connected to two GPU as shown in figure 1
Between biggish circle.
In addition, according to another aspect of the disclosure, component as described above is also performed for accelerating coprocessor
The method of data flowing in data handling path.Multiple execute in body has the second designated groups in the data handling path
The output number that second execution body of member will only be stored in the output data caching for the third execution body for executing body as its upstream
Local output data caching, and one or more tools into the multiple execution body are transported to according to the corresponding header data of block
There is the first execution body of the first specified constituent element to issue the message that can read data.Multiple execution bodies in the data handling path
In one or more of multiple execution bodies first execute body based on executing body from second with the first specified constituent element
Message reads the header data from the output data caching of the second execution body, and executes scheduled operation.Described second holds
Row body executes body transmission to the third of the second execution body and disappears while header data is transported to local output data caching
Breath, so that third, which executes body, is placed in blank state for the output data caching of its own.Described first executes body from disposed thereon
It is similar with the third execution content-data of body data block generated that the 4th of trip executes acquisition in the output data caching of body
Data are to execute scheduled operation.Second execution body receive it is all first execute bodies feedback message after by its own
Output data caching is placed in blank state.
The basic principle of the disclosure is described in conjunction with specific embodiments above, however, it is desirable to, it is noted that this field
For those of ordinary skill, it is to be understood that the whole or any steps or component of disclosed method and device, Ke Yi
Any computing device (including processor, storage medium etc.) perhaps in the network of computing device with hardware, firmware, software or
Their combination is realized that this is that those of ordinary skill in the art use them in the case where having read the explanation of the disclosure
Basic programming skill can be achieved with.
Therefore, the purpose of the disclosure can also by run on any computing device a program or batch processing come
It realizes.The computing device can be well known fexible unit.Therefore, the purpose of the disclosure can also include only by offer
The program product of the program code of the method or device is realized to realize.That is, such program product is also constituted
The disclosure, and the storage medium for being stored with such program product also constitutes the disclosure.Obviously, the storage medium can be
Any well known storage medium or any storage medium that developed in the future.
It may also be noted that in the device and method of the disclosure, it is clear that each component or each step are can to decompose
And/or reconfigure.These decompose and/or reconfigure the equivalent scheme that should be regarded as the disclosure.Also, execute above-mentioned series
The step of processing, can execute according to the sequence of explanation in chronological order naturally, but not need centainly sequentially in time
It executes.Certain steps can execute parallel or independently of one another.
Above-mentioned specific embodiment does not constitute the limitation to disclosure protection scope.Those skilled in the art should be bright
It is white, design requirement and other factors are depended on, various modifications, combination, sub-portfolio and substitution can occur.It is any
Made modifications, equivalent substitutions and improvements etc., should be included in disclosure protection scope within the spirit and principle of the disclosure
Within.
Claims (8)
1. a kind of execution body creates system, comprising:
Task topological diagram formation component, for holding the breakdown of operation of required completion by executing body in isomery framework to be a series of
Capable task, and while breakdown of operation, based on the intrinsic relationship between being decomposed for task, generate task nexus topology
Figure, the task node of the task topological diagram contain the whole nodal communities for executing corresponding required by task, the task topological diagram
Formation component includes specified constituent element node selecting unit, for being selected during generating task nexus topological diagram according to specific tasks
The specific tasks node with the specified constituent element comprising additional node attributes is selected, the additional node attributes change the specific tasks
Node executes the mode of corresponding task；And
Whole nodal communities of execution body creation component, task based access control relationship topology figure and each task node contain specified
Whole nodal communities of constituent element are that each task node creates corresponding execution body, and makes and have building-out section in computing resource
The corresponding specific execution body of specific tasks node of point attribute executes corresponding task according to additional node attributes.
2. executing body as described in claim 1 creates system, wherein the additional node attributes that the designated groups member includes are number
According to running node attribute, for so that the specific execution body created only executes operation to the header data of data processed.
3. executing body as described in claim 1 creates system, wherein the additional node attributes that the designated groups member includes are number
According to memory node attribute, for so that the specific execution body created only stores the header data of data processed.
4. executing body as described in claim 1 creates system, wherein the additional node attributes that the designated groups member includes are number
According to modification nodal community, for enabling the specific execution body created to modify the header data of data processed.
5. a kind of execution body creation method, comprising:
The breakdown of operation of required completion is performed task by executing body in isomery framework to be a series of, and in breakdown of operation
Meanwhile based on the intrinsic relationship between being decomposed for task, task nexus topological diagram, the task section of the task topological diagram are generated
Point contains the whole nodal communities for executing corresponding required by task；
During generating task nexus topological diagram, the specified constituent element comprising additional node attributes is had according to specific tasks selection
Specific tasks node, the additional node attributes change the mode that the specific tasks node executes corresponding task；And
Whole nodal communities of task based access control relationship topology figure and each task node or whole nodes containing specified constituent element
Attribute is that each task node creates corresponding execution body, and makes and have the specific of additional node attributes to appoint in computing resource
The corresponding specific execution body of node be engaged according to the corresponding task of additional node attributes execution.
6. body creation method is executed as claimed in claim 5, wherein the additional node attributes that the designated groups member includes are number
According to running node attribute, for so that the specific execution body created only executes operation to the header data of data processed.
7. body creation method is executed as claimed in claim 5, wherein the additional node attributes that the designated groups member includes are number
According to memory node attribute, for so that the specific execution body created only stores the header data of data processed.
8. body creation method is executed as claimed in claim 5, wherein the additional node attributes that the designated groups member includes are number
According to modification nodal community, for enabling the specific execution body created to modify the header data of data processed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910633638.1A CN110262995A (en) | 2019-07-15 | 2019-07-15 | It executes body creation system and executes body creation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910633638.1A CN110262995A (en) | 2019-07-15 | 2019-07-15 | It executes body creation system and executes body creation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110262995A true CN110262995A (en) | 2019-09-20 |
Family
ID=67926099
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910633638.1A Pending CN110262995A (en) | 2019-07-15 | 2019-07-15 | It executes body creation system and executes body creation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110262995A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245108A (en) * | 2019-07-15 | 2019-09-17 | 北京一流科技有限公司 | It executes body creation system and executes body creation method |
| CN110955734A (en) * | 2020-02-13 | 2020-04-03 | 北京一流科技有限公司 | Distributed signature decision system and method for logic node |
| CN110955511A (en) * | 2020-02-13 | 2020-04-03 | 北京一流科技有限公司 | Executive body and data processing method thereof |
| CN111666151A (en) * | 2020-02-13 | 2020-09-15 | 北京一流科技有限公司 | Topological graph conversion system and method |
| WO2021147876A1 (en) * | 2020-01-20 | 2021-07-29 | 北京一流科技有限公司 | Memory resource in-situ sharing decision-making system and method |
| WO2021213076A1 (en) * | 2020-04-24 | 2021-10-28 | 中科寒武纪科技股份有限公司 | Method and device for constructing communication topology structure on basis of multiple processing nodes |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6983324B1 (en) * | 2000-10-23 | 2006-01-03 | International Business Machines Corporation | Dynamic modification of cluster communication parameters in clustered computer system |
| US20090219835A1 (en) * | 2008-02-29 | 2009-09-03 | International Business Machines Corporation | Optimizing A Physical Data Communications Topology Between A Plurality Of Computing Nodes |
| US20090300154A1 (en) * | 2008-05-29 | 2009-12-03 | International Business Machines Corporation | Managing performance of a job performed in a distributed computing system |
| US20090307703A1 (en) * | 2008-06-09 | 2009-12-10 | International Business Machines Corporation | Scheduling Applications For Execution On A Plurality Of Compute Nodes Of A Parallel Computer To Manage temperature of the nodes during execution |
| US20120278365A1 (en) * | 2011-04-28 | 2012-11-01 | Intuit Inc. | Graph databases for storing multidimensional models of softwqare offerings |
| US20150205888A1 (en) * | 2014-01-17 | 2015-07-23 | International Business Machines Corporation | Simulation of high performance computing (hpc) application environment using virtual nodes |
| CN106648859A (en) * | 2016-12-01 | 2017-05-10 | 北京奇虎科技有限公司 | Task scheduling method and device |
| CN106681820A (en) * | 2016-12-30 | 2017-05-17 | 西北工业大学 | Message combination based extensible big data computing method |
| CN107005422A (en) * | 2014-09-30 | 2017-08-01 | 慧与发展有限责任合伙企业 | Topology-Based Management for Day-Day Operations |
| CN107005421A (en) * | 2014-09-30 | 2017-08-01 | 慧与发展有限责任合伙企业 | Utilize the management based on topology of stage and version policy |
| CN108121330A (en) * | 2016-11-26 | 2018-06-05 | 沈阳新松机器人自动化股份有限公司 | A kind of dispatching method, scheduling system and map path planing method |
| CN108388474A (en) * | 2018-02-06 | 2018-08-10 | 北京易沃特科技有限公司 | Intelligent distributed management of computing system and method based on DAG |
| US20180262392A1 (en) * | 2017-03-08 | 2018-09-13 | Linkedin Corporation | Automatically detecting roles of nodes in layered network topologies |
| CN110245108A (en) * | 2019-07-15 | 2019-09-17 | 北京一流科技有限公司 | It executes body creation system and executes body creation method |
-
2019
- 2019-07-15 CN CN201910633638.1A patent/CN110262995A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6983324B1 (en) * | 2000-10-23 | 2006-01-03 | International Business Machines Corporation | Dynamic modification of cluster communication parameters in clustered computer system |
| US20090219835A1 (en) * | 2008-02-29 | 2009-09-03 | International Business Machines Corporation | Optimizing A Physical Data Communications Topology Between A Plurality Of Computing Nodes |
| US20090300154A1 (en) * | 2008-05-29 | 2009-12-03 | International Business Machines Corporation | Managing performance of a job performed in a distributed computing system |
| US20090307703A1 (en) * | 2008-06-09 | 2009-12-10 | International Business Machines Corporation | Scheduling Applications For Execution On A Plurality Of Compute Nodes Of A Parallel Computer To Manage temperature of the nodes during execution |
| US20120278365A1 (en) * | 2011-04-28 | 2012-11-01 | Intuit Inc. | Graph databases for storing multidimensional models of softwqare offerings |
| US20150205888A1 (en) * | 2014-01-17 | 2015-07-23 | International Business Machines Corporation | Simulation of high performance computing (hpc) application environment using virtual nodes |
| CN107005421A (en) * | 2014-09-30 | 2017-08-01 | 慧与发展有限责任合伙企业 | Utilize the management based on topology of stage and version policy |
| CN107005422A (en) * | 2014-09-30 | 2017-08-01 | 慧与发展有限责任合伙企业 | Topology-Based Management for Day-Day Operations |
| CN108121330A (en) * | 2016-11-26 | 2018-06-05 | 沈阳新松机器人自动化股份有限公司 | A kind of dispatching method, scheduling system and map path planing method |
| CN106648859A (en) * | 2016-12-01 | 2017-05-10 | 北京奇虎科技有限公司 | Task scheduling method and device |
| CN106681820A (en) * | 2016-12-30 | 2017-05-17 | 西北工业大学 | Message combination based extensible big data computing method |
| US20180262392A1 (en) * | 2017-03-08 | 2018-09-13 | Linkedin Corporation | Automatically detecting roles of nodes in layered network topologies |
| CN108388474A (en) * | 2018-02-06 | 2018-08-10 | 北京易沃特科技有限公司 | Intelligent distributed management of computing system and method based on DAG |
| CN110245108A (en) * | 2019-07-15 | 2019-09-17 | 北京一流科技有限公司 | It executes body creation system and executes body creation method |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245108A (en) * | 2019-07-15 | 2019-09-17 | 北京一流科技有限公司 | It executes body creation system and executes body creation method |
| WO2021147876A1 (en) * | 2020-01-20 | 2021-07-29 | 北京一流科技有限公司 | Memory resource in-situ sharing decision-making system and method |
| CN110955734A (en) * | 2020-02-13 | 2020-04-03 | 北京一流科技有限公司 | Distributed signature decision system and method for logic node |
| CN110955511A (en) * | 2020-02-13 | 2020-04-03 | 北京一流科技有限公司 | Executive body and data processing method thereof |
| CN110955511B (en) * | 2020-02-13 | 2020-08-18 | 北京一流科技有限公司 | Executive body and data processing method thereof |
| CN111666151A (en) * | 2020-02-13 | 2020-09-15 | 北京一流科技有限公司 | Topological graph conversion system and method |
| WO2021159928A1 (en) * | 2020-02-13 | 2021-08-19 | 北京一流科技有限公司 | Distributed signature decision-making system and method for logical nodes |
| CN111666151B (en) * | 2020-02-13 | 2023-11-03 | 北京一流科技有限公司 | Topological graph conversion system and method thereof |
| US11818231B2 (en) | 2020-02-13 | 2023-11-14 | Beijing Oneflow Technology Co., Ltd | Logical node distributed signature decision system and a method thereof |
| WO2021213076A1 (en) * | 2020-04-24 | 2021-10-28 | 中科寒武纪科技股份有限公司 | Method and device for constructing communication topology structure on basis of multiple processing nodes |
| US12050545B2 (en) | 2020-04-24 | 2024-07-30 | Cambricon (Xi'an) Semiconductor Co., Ltd. | Method and device for constructing communication topology structure on basis of multiple processing nodes |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110262995A (en) | It executes body creation system and executes body creation method | |
| CN110209629A (en) | Data flowing acceleration means and its method in the data handling path of coprocessor | |
| CN110245108A (en) | It executes body creation system and executes body creation method | |
| US12147829B2 (en) | Data processing system and method for heterogeneous architecture | |
| CN113010302A (en) | Multi-task scheduling method and system under quantum-classical hybrid architecture and quantum computer system architecture | |
| Maturana et al. | Autonomous operator management for evolutionary algorithms | |
| US12602348B2 (en) | Data actor and data processing method thereof | |
| CN109144374A (en) | Method for processing business, system and relevant device based on visualization regulation engine | |
| CN108388474A (en) | Intelligent distributed management of computing system and method based on DAG | |
| CN110428054A (en) | Business Rule Engine pattern matching system and implementation method based on technique of compiling | |
| Zhang et al. | A hybrid intelligent algorithm and rescheduling technique for job shop scheduling problems with disruptions | |
| Ahmad et al. | Data-intensive workflow optimization based on application task graph partitioning in heterogeneous computing systems | |
| Chattopadhyay et al. | QSCAS: QoS aware web service composition algorithms with stochastic parameters | |
| WO2025055556A1 (en) | Method and apparatus for executing plurality of computing tasks | |
| Xu et al. | Living with artificial intelligence: A paradigm shift toward future network traffic control | |
| CN114240632A (en) | Batch job execution method, apparatus, apparatus, medium and product | |
| CN107797852A (en) | The processing unit and processing method of data iteration | |
| Ashcroft | Dataflow and education: Data-driven and demand-driven distributed computation | |
| Ghooshchi et al. | Visualisation of compliant declarative business processes | |
| Glover et al. | Basis exchange characterizations for the simplex SON algorithm for LP/embedded networks | |
| Oliveira et al. | IMCReo: interactive Markov chains for stochastic Reo | |
| CN109636299A (en) | Node time limit mechanism based on workflow band | |
| Wang | Analysis of Interval Flow Calculation Algorithm of Multi-Loop Complex Power Network Based on Knowledge Mapping | |
| Cheng et al. | Researches on asynchronous pipeline Web service composition method | |
| Idan | Effective Reward Schemes for Tardiness Optimization |
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 |