CN104780213B - A kind of master-salve distributed figure processing system load dynamic optimization method - Google Patents
A kind of master-salve distributed figure processing system load dynamic optimization method Download PDFInfo
- Publication number
- CN104780213B CN104780213B CN201510181554.0A CN201510181554A CN104780213B CN 104780213 B CN104780213 B CN 104780213B CN 201510181554 A CN201510181554 A CN 201510181554A CN 104780213 B CN104780213 B CN 104780213B
- Authority
- CN
- China
- Prior art keywords
- node
- load
- calculate
- summit
- load capacity
- 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.)
- Active
Links
Abstract
The invention discloses a kind of master-salve distributed figure processing system to load dynamic optimization method, includes the subdivided rate-determining steps of dynamic of master computing node, load monitoring step and load transferring step in the calculate node that works.Initial division of the present invention independent of diagram data.Working node is when iteration performs, the subdivided carry out load balancing of dynamic is performed according to the instruction of host node, load monitoring step monitors the load of each calculate node, and it is sent to other each calculate nodes before the execution of each iteration terminates, load transferring step is when each iteration performs beginning simultaneously, the node overload according to where judging whether the Payload message monitored of other nodes received, and determine goal displacement node and transfer amount, when current iteration has performed, load data is transferred to destination node, so as to realize the dynamic load leveling of distributed figure processing system.Load imbalance phenomenon present in distributed figure processing system can effectively be improved by implementing the present invention.
Description
Technical field
The invention belongs to distributed diagram data process field, schemes more particularly, to the distribution based on BSP model realizations
Processing system.
Background technology
Figure is a kind of abstract data structure the most frequently used in computer science, relative to traditional relation data and XML numbers
According to storehouse, the ability to express of figure is more abundant, and therefore, the application related to figure is nearly ubiquitous.But with the big data epoch
Arrival, the scale of figure increasingly increases, and carries out distributed treatment to figure in cloud computing environment, has become new research and become
Gesture.Therefore substantial amounts of distributed figure processing system is also had, is mainly all based on the class Pregel systems of BSP model realizations.
BSP computation models are synchronous computation models, can carry out multiple iterative cycles execution, an iteration is including calculating, communicating and together
Walk three steps.BSP models are especially suitable for the successive ignition characteristic that distributed figure calculates, and therefore, Google opens according to BSP models
The internal distributed figure processing model Pregel used is sent out.Pregel employs the method centered on summit, i.e. summit is joined
With calculating, summit is divided into the process of implementation enlivens state and inactive state.Side in figure is not involved in calculating and is served only for transmitting message.
Nomography is performed and be able to could completed by successive ignition.If have received message for the summit of inactive state, can be activated.
Pregel carries out distributed treatment using host-guest architecture simultaneously, and host node is responsible for coordinating each calculate node being operated, and calculates
Node is then mainly responsible for figure task computation.
Figure division is step of crucial importance when carrying out figure processing in distributed figure processing system, effective figure division plan
Summary can greatly improve the treatment effeciency of figure.Existing figure partition strategy is the root before diagram data is loaded into calculate node mostly
According to the principle of figure division:Low connectedness between subgraph equilibrium and subgraph, carries out an initial division, is divided for this kind of figure
It is tactful that we are referred to as static map division.However, when distributed figure processing is carried out after being divided to diagram data, according to the nomography of execution
The difference of (i.e. graphic operation), the iteration feature of figure are also different (load imbalance being present in i.e. each calculate node).Because
Different nomographys in each iterative process and need not to figure in all vertex datas handle.Therefore cause different
Nomography has different load behaviors upon execution, load imbalance during so as to produce operation.But static figure division
Algorithm is difficult to the load Behavioral change in prognostic chart at the initial stage of execution, therefore the division of once static figure can not solve algorithms of different
Load imbalance during caused operation.
The content of the invention
Load imbalance problem during for operation set forth above, it is applied to distributed figure the invention provides one kind and handles
The load dynamic optimization method of scene.First, load when being performed to figure is monitored.According to the monitoring knot of each calculate node
Fruit, overload node is determined according to global average load, part load is transferred to the node that do not overload from overload node, this mistake
Journey also referred to as loads transfer.Itself it can cause certain calculating and communication overhead because dynamic is subdivided, so needing to dynamic
It is subdivided to be controlled in itself, therefore the subdivided rate-determining steps of dynamic have been also needed on the primary node.This invention can be effective
Solve the problems, such as the load imbalance as caused by nomography, make up the deficiency of static division.
The dynamic of host node is subdivided in load dynamic optimization method provided by the invention, including distributed figure processing system
Load monitoring and load transferring step in rate-determining steps, and working node.Dynamically subdivided rate-determining steps are mainly adaptive
That answers controls dynamic subdivided execution and end, to reduce dynamic division caused expense in itself.Load monitoring step and negative
Carry transfer step and be located at working node, both complement each other, and are the chief components of dynamic division.Here host node and work
It is all identical to make the physical machine performance configuration residing for node.
Described load monitoring step is used for the loading condition of each iterative process when monitoring distributed figure is handled.Work section
The load of point determines by enlivening vertex set in an iteration and enlivening side collection, wherein, movable top points (i.e. active vertex
The length of collection) it is the number for needing to call summit to calculate function in working node, enliven is in number (length of collection when enlivening)
Summit calculates function need message count to be processed in working node, and specific formula (1) is as follows:
Wherein, i is any operative node, AViTo enliven vertex set, AEiTo enliven side collection, K is calculate node number.Respectively
The load monitored is all sent to other all nodes by individual working node, and whether itself is determined for next iteration working node
The load capacity shifted is needed after overload and overload.Pay attention to, load is not sent to the reason for host node is to be calculated here
It is that the present invention is directed to the figure processing system for being all based on BSP models, the transmission of message each time only enters down after synchronization
An iteration gets to corresponding message destination.If host node carries out overload judgement and the determination for the node that diverts the aim,
Now, working node can only without it is any calculating and wait host node calculated after into next iteration, could receive
The object information that host node is sent.This will cause the waste of even more serious computing resource, so we calculate these
Carried out in each working node.
Described load transferring step is used for from overload node transfer part load summit to node is not overloaded to reach
Load balancing during operation.When each iteration performs beginning, each calculate node can receive all calculate nodes and send
Load monitoring result.Each calculate node determines global average load according to the loading condition of all calculate nodes received, sentences
Whether break itself is overload node.Here it is overload standard that can not simply set global average load, because working as each calculating
Node is more or less the same in the case of almost equilibrium, and the load capacity for also having calculate node is more than the situation of average load, so only
Having load capacity to be more than the situation in the certain section of average value can just confirm as overloading, and preferential section is to exceed average load (110%-
130%), value is too small may miss balanced node as overload node, and value is excessive may to miss some overload nodes.
If overload node, then need the load to all calculate nodes to sort and determine goal displacement node.For example, a total of K work
Node, and each working node sorts from big to small according to its load, the node of overload is certainly close to gauge outfit, position i
(1<i<=K), then the corresponding node location that diverts the aim that loads is K-i+1.According to formula (2) determine from overload node (i) to
The load transfer amount of destination node (j).
The present invention can determine to need the summit shifted, contribution letter according to the Contribution Function that summit is enlivened in calculate node
Number definition such as formula (3).
IOi,j(u)=Ii,j(u)+Oi,j(u) formula (3)
Wherein, u is to enliven summit, I in calculate node (i)i,jAnd O (u)i,j(u) it is summit u respectively from calculate node
(j) message count that receives and it is sent to the message count for calculating summit (j).Here it is possible to Contribution Function is carried out further
Optimization.For example, when iterations is s, the contribution margin that u stays in calculate node (i) isAnd it is transferred to mesh
Mark node (j) contribution margin beThe benefit and unobvious that now transfer summit u is brought, such case, I can
To consider not shifting summit u.Therefore, we can set a rate λ, the contribution margin IO only after transferi,j(u) and not
The contribution margin IO of transferi,i(u) ratio is more than λ, i.e.,We just consider to shift summit u, i.e. now contribution margin
For former IOi,j(u), such as less than λ, then contribution margin is 0.It should be noted that λ here>1, preferential section is (1.2-1.4), value
Crossing conference causes transferable number of vertex less, and value is too small to cause communication overhead to increase.Saved by above-mentioned Contribution Function to calculating
The transfer ranking enlivened the sequence of summit from big to small, obtain enlivening summit in point, will be from transferring load summit
Choose the summit with maximum contribution value and shifted in ranking top.Each load capacity for enlivening summit transfer summit reduction is by public affairs
Formula (4) calculates,
Wi(u)=1+ | AEI, j(u) | formula (4)
Wherein, | AEI, j(u) | represent and enliven summit u connections enlivens side collection length.
The subdivided rate-determining steps of described dynamic are located in host node.Host node is the control section of distributed figure calculating task
Point, it is not involved in calculating.Load information that can be according to each calculate node in the subdivided rate-determining steps of dynamic and transfer summit
Number, judges whether load has reached Optimal State, if it is subdivided then to terminate dynamic.
It is excellent for carrying out load balancing in scheming processing scene in distribution the invention provides a kind of load optimized method
Change, in methods described the overall execution of each step comprise the following steps:
(1) initialization step:The diagram data for needing to be calculated is uploaded into distributed figure processing system, it is determined that distributed
The figure that figure processing system needs to perform calculates operation;Distributed figure processing system carries out figure division, figure number to the diagram data of loading
Each calculate node is loaded into respectively according to being divided into after multiple sub-graph datas;
(2) initial load calculation procedure:Each calculate node calculates the load capacity of this section idea diagram data;In stage of communication,
Load capacity is sent to other all calculate nodes and host node;The load capacity, which equal to this node is enlivened number of vertex and added, enlivens side
Number;
(3) load capacity discriminating step:Whether each calculate node judges itself according to the other calculate node load capacity received
Overload, is to calculate this node load and gone to step before stage of communication (4);Otherwise continue to calculate this node load capacity, logical
In the letter stage, load capacity is sent to other all calculate nodes, gone to step (5);
(4) load transferring step;
(5) the subdivided step of dynamic:Host node calculates the number of vertex that all calculate nodes shift in upper once iteration, sentences
Disconnected transfer number of vertex whether be less than default transfer threshold value (threshold value here refer to dynamic it is subdivided after income and its expense
Critical value, preferential section are (0.5%-2%), and it is subdivided and reduce load optimized effect that value crosses the too early end dynamic of conference
Fruit, value it is too small can be subdivided due to long-play dynamic and cause the expense excessive):Be then host node to each calculate node
The instruction that dynamic division terminates is sent, each calculate node receives the dynamic division halt instruction of host node, stops performing load capacity
Calculate and load is shifted;Otherwise send and continue load optimized instruction, go to step (3).
Compared with prior art, the invention has the advantages that:
(1) graph data structure and figure calculating operation are relied on smaller:Original static division is largely dependent upon
Graph data structure and corresponding graphic operation, once initial static division be extremely difficult to all graphic operations in whole iteration mistake
All it is load balancing in journey.The subdivided strategy of dynamic that the present invention uses is independent of initial division, for different figure numbers
There is good balanced loaded effect according to structure and graphic operation.
(2) load optimized effect is good:By monitoring the loading condition of each calculate node in real time, positioning overload node simultaneously turns
The big summit of contribution factor is moved to the node that do not overload, dynamic is subdivided greatly to improve the load as caused by different graphic calculation judicial acts
It is unbalanced, reduce the disposed of in its entirety time that distributed figure calculates, load balance optimization positive effect.Best situation, can
Reduce 50% and calculate the time (the balanced obvious algorithm of SSSP, BSP even load), can reduce 10%-30%'s for ordinary circumstance
Calculate the time.
(3) subsidiary expense is small:Due to dynamic it is subdivided performed in distributed figure calculating process, the monitoring of load and turn
Shifting can bring certain calculating and communication overhead.Therefore, present invention employs the subdivided rate-determining steps of dynamic, supervise in the master node
The transfer number of vertex and loading condition of each calculate node are controlled, when load reaches balanced convergence, it is subdivided to stop dynamic.Separately
Outside, different from some other subdivided strategies, the present invention can't additionally increase iterations and be shifted for summit, load prison
Control and transfer all perform in same iterative calculation, so as to reduce the subdivided caused subsidiary expense of dynamic.
Brief description of the drawings
Below in conjunction with drawings and Examples, the invention will be further described, in accompanying drawing:
Fig. 1 is the overview flow chart of the present invention;
Fig. 2 is the detailed operational flow diagrams of bright load transfer;
Fig. 3 is the subdivided step workflow diagram of host node dynamic.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
The load dynamic optimization method based on the master-salve distributed figure processing system of BSP models shown in Fig. 1 performs step such as
Under:
(1) initialization step:The diagram data for needing to be calculated is uploaded into distributed figure processing system, it is determined that distributed
The figure that figure processing system needs to perform calculates operation;Distributed figure processing system carries out Hash divisions, figure to the diagram data of loading
Data are divided into after multiple sub-graph datas is loaded into each calculate node respectively;
(2) initial load calculation procedure:Each calculate node calculates the load capacity of this section idea diagram data;In stage of communication,
Load capacity is sent to other all calculate nodes and host node;The load capacity, which equal to this node is enlivened number of vertex and added, enlivens side
Number;
(3) load capacity discriminating step:Each calculate node calculates average load according to the other calculate node load capacity received
To judge whether itself overloads.If the load capacity of calculate node is more than the 110% of average load, it is determined that for overload, calculates
Gone to step after this node load before stage of communication (4);Otherwise continue to calculate this node load capacity, in stage of communication, will load
Amount is sent to other all calculate nodes, goes to step (5);
(4) load transferring step;
(5) the subdivided step of dynamic:Host node calculates the number of vertex that all calculate nodes shift in upper once iteration, sentences
Whether disconnected transfer number of vertex is less than default transfer threshold value:It is that then host node sends dynamic division end to each calculate node
Instruction, each calculate node receive the dynamic division halt instruction of host node, stop performing load capacity calculating and load transfer;Otherwise
Send and continue load optimized instruction, go to step (3);
As shown in Fig. 2 load transferring step detailed operation process step is as follows in step (4) in Fig. 1:
The load of (4-1) to all calculate nodes is ranked up, and determines goal displacement node.
(4-2) calculates transferring load amount Q according to the load capacity of goal displacement calculate nodeI, j, and willIt is assigned to interim
Variable Temp.
(4-3) is according to Contribution Function value IOi,j(u) descending sort (this reality is carried out to all summits of enlivening in calculate node i
Apply in example, the λ in Contribution Function formula is set to 1.2), obtain shifting list
(4-4) judges Temp more than 0 or transfer list also has contribution margin to enliven whether summit is true more than 0, if
Into step (4-5), otherwise load transfer and terminate.
(4-5) from transfer listMiddle taking-up top ranked enlivens summit u.
The load capacity W of reduction after (4-6) u transfersi(u) whether it is less than or equal to Temp, if into step (4-7), otherwise
Into step (4-4).
U is transferred to target computing nodes j by (4-7) from calculate node i.
The load capacity W of reduction after (4-8) shifts Temp-ui(u) Temp is assigned to, into step (4-4).
Fig. 3 is the workflow diagram of the load optimized method host node end an iteration of the embodiment of the present invention (containing dynamic
Subdivided rate-determining steps), as shown in figure 1, host node end an iteration workflow in the load optimized method and step (5) of the present invention
Comprise the following steps:
The transfer number of vertex and load information that each calculate node of last iteration that (5-1) processing receives is sent.
(5-2) calculates the summit sum that all calculate nodes shift in upper once iteration, and with threshold limit value (this implementation
Mode is arranged to 1%) being compared for figure summit total quantity.
(5-3) judges whether the transfer number of vertex of all nodes is less than threshold limit value, if into step (5-4), otherwise
Into step (5-5).
(5-4) host node sends the subdivided instruction of stopping dynamic and arrives each calculate node.
(5-5) each calculate node is iterated calculating, until all calculate nodes enter synchronous regime, current iteration knot
Beam.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, all any modification, equivalent and improvement made within the spirit and principles of the invention etc., all should be included
In the present invention.
Claims (4)
1. a kind of master-salve distributed figure processing system load dynamic optimization method based on BSP models, it is characterised in that including such as
Lower step:
(1) initialization step:The diagram data for needing to be calculated is uploaded into distributed figure processing system, it is determined that at distributed figure
The figure that reason system needs to perform calculates operation;Distributed figure processing system carries out figure division to the diagram data of loading, and diagram data is drawn
It is divided into after multiple sub-graph datas and is loaded into each calculate node respectively;
(2) initial load calculation procedure:Each calculate node calculates the load capacity of this section idea diagram data;, will be negative in stage of communication
Carrying capacity is sent to other all calculate nodes and host node;The load capacity, which equal to this node is enlivened number of vertex and added, enlivens side number;
(3) load capacity discriminating step:Each calculate node judges whether itself surpasses according to the other calculate node load capacity received
Carry, be to calculate this node load and gone to step before stage of communication (4);Otherwise continue to calculate this node load capacity, communicating
In the stage, load capacity is sent to other all calculate nodes, gone to step (5);
(4) load transferring step:Including following sub-step:
(4.1) selection target transfering node:The load capacity of all calculate nodes is sorted, generation load sequencing table, according to each section
Point load capacity determines its position in sequencing table, and load capacity is more just closer to gauge outfit, serial number of the node in sequencing table
I, then its goal displacement node ID be mapped as j, j=K-i+1, wherein K is distributed figure processing system calculate node sum;
(4.2) transferring load amount Q is calculatedI, j,
<mrow>
<msub>
<mi>Q</mi>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mrow>
<msub>
<mi>W</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>W</mi>
<mi>j</mi>
</msub>
</mrow>
<mn>2</mn>
</mfrac>
</mrow>
Wherein Wi、WjThe load capacity of operation is once calculated before expression calculate node i, j respectively, i is that node is produced in load, and j is target
Transfering node, by QI, jIt is assigned to temporary variable T;
(4.3) all contribution margin IO for enlivening summit u in calculate nodei,j(u),
IOi,j(u)=Ii,j(u)+Oi,j(u)
Wherein, Ii,jAnd O (u)i,j(u) be respectively in iteration summit u from the goal displacement node j message counts received and transmission
Message count to goal displacement node j;
WhenWhen, then summit u contribution margin is just original IOi,j(u) summit u contribution margin otherwise, is assigned to 0;
Wherein IOi,j(u) be summit u transfer after contribution margin, IOi,i(u) it is contribution margin that summit u is not shifted, λ is default transfer gate
Limit value, span 1.2-1.4;
Enliven summit to all in calculate node and sorted from big to small, obtain shifting sequencing table;
(4.4) judge whether that T is less than or equal to 0 and transfer sequencing tableIn do not have contribution margin be more than 0 element, if then will turn
The load information for moving number of vertex and this node is sent to host node in stage of communication, and load transfer terminates, and carries out step (5);
Otherwise sub-step (4.5) is carried out;
(4.5) from transfer sequencing tableMiddle taking-up top ranked enlivens summit u;Calculate the load capacity W of reduction after it is shiftedi
(u), Wi(u)=1+ | AEI, j(u) |, wherein, | AEI, j(u) | represent and enliven summit u connections enlivens side collection size;
(4.6) W is judgedi(u) whether it is less than or equal to T, then carries out sub-step (4.7), otherwise carries out sub-step (4.4);
(4.7) summit u will be enlivened and is transferred to target computing nodes j from calculate node i, by T-Wi(u) value assigns T, into step
(4.4);
(5) the subdivided step of dynamic:Host node calculates the number of vertex that all calculate nodes shift in upper once iteration, judges to turn
Move whether number of vertex is less than default transfer threshold value:It is the finger that then host node sends dynamic division end to each calculate node
Order, each calculate node receive the dynamic division halt instruction of host node, stop performing load capacity calculating and load transfer;Otherwise send out
Go out and continue load optimized instruction, go to step (3).
2. according to the method for claim 1, it is characterised in that during step (3) load capacity differentiates, calculate node judges
Whether itself overloads including following sub-step:
(1) according to each calculate node load information received, computing system average load:System average load=all meters of Σ
Operator node load capacity/calculate node sum;
(2) by this node load capacity compared with system average load, load capacity exceed average load 110%-130% and more than
For overload;Otherwise it is non-overloading.
3. according to the method for claim 1, it is characterised in that in sub-step (4.3), if summit u is transferred before
Cross, then its contribution margin is assigned to 0.
4. according to the method for claim 1, it is characterised in that the transfer number threshold value of summit described in step (5) is that figure summit is total
Several 0.5%-2%.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510181554.0A CN104780213B (en) | 2015-04-17 | 2015-04-17 | A kind of master-salve distributed figure processing system load dynamic optimization method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510181554.0A CN104780213B (en) | 2015-04-17 | 2015-04-17 | A kind of master-salve distributed figure processing system load dynamic optimization method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104780213A CN104780213A (en) | 2015-07-15 |
CN104780213B true CN104780213B (en) | 2018-02-23 |
Family
ID=53621447
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510181554.0A Active CN104780213B (en) | 2015-04-17 | 2015-04-17 | A kind of master-salve distributed figure processing system load dynamic optimization method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104780213B (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105653204B (en) * | 2015-12-24 | 2018-12-07 | 华中科技大学 | A kind of distributed figure calculation method based on disk |
CN107122248B (en) * | 2017-05-02 | 2020-01-21 | 华中科技大学 | Storage optimization distributed graph processing method |
CN112912865A (en) * | 2018-07-27 | 2021-06-04 | 浙江天猫技术有限公司 | Graph data storage method and system and electronic equipment |
CN109388733B (en) * | 2018-08-13 | 2022-01-07 | 国网浙江省电力有限公司 | Optimization method for graph-oriented data processing engine |
CN109104334B (en) * | 2018-08-23 | 2021-04-02 | 郑州云海信息技术有限公司 | Management method and device for nodes in monitoring system |
WO2020181455A1 (en) * | 2019-03-11 | 2020-09-17 | 深圳大学 | Geographically distributed graph processing method and system |
CN111859027A (en) * | 2019-04-24 | 2020-10-30 | 华为技术有限公司 | Graph calculation method and device |
CN111309976B (en) * | 2020-02-24 | 2021-06-25 | 北京工业大学 | GraphX data caching method for convergence graph application |
CN113093682B (en) * | 2021-04-09 | 2022-07-05 | 天津商业大学 | Non-centralized recursive dynamic load balancing computing system |
CN113326125B (en) * | 2021-05-20 | 2023-03-24 | 清华大学 | Large-scale distributed graph calculation end-to-end acceleration method and device |
CN115587222B (en) * | 2022-12-12 | 2023-03-17 | 阿里巴巴(中国)有限公司 | Distributed graph calculation method, system and equipment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102207891A (en) * | 2011-06-10 | 2011-10-05 | 浙江大学 | Method for achieving dynamic partitioning and load balancing of data-partitioning distributed environment |
CN104298564A (en) * | 2014-10-15 | 2015-01-21 | 中国人民解放军国防科学技术大学 | Dynamic equilibrium heterogeneous system loading computing method |
-
2015
- 2015-04-17 CN CN201510181554.0A patent/CN104780213B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102207891A (en) * | 2011-06-10 | 2011-10-05 | 浙江大学 | Method for achieving dynamic partitioning and load balancing of data-partitioning distributed environment |
CN104298564A (en) * | 2014-10-15 | 2015-01-21 | 中国人民解放军国防科学技术大学 | Dynamic equilibrium heterogeneous system loading computing method |
Non-Patent Citations (3)
Title |
---|
Dynamic Repartitioning of Adaptively Refined Meshes;Kirk Schloegel等;《IEEE》;19981113;全文 * |
Efficient Variable Partitioning and Scheduling for DSP Processors With Multiple Memory Modules;Qingfeng Zhuge,Edwin Hsing-Mean Sha,Bin Xiao等;《IEEE》;20040322;全文 * |
Load Balancing of Dynamic and Adaptive Mesh-Based Computations;Kirk Schloegel,George Karypis,Vipin Kumar;《IEEE》;19981023;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN104780213A (en) | 2015-07-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104780213B (en) | A kind of master-salve distributed figure processing system load dynamic optimization method | |
CN103401939B (en) | Load balancing method adopting mixing scheduling strategy | |
CN108965014B (en) | QoS-aware service chain backup method and system | |
CN104901898A (en) | Load balancing method and device | |
CN102868733A (en) | Method for remapping virtual network resources | |
CN105512179A (en) | Data sorting device, method and data processing chip achieved by hardware | |
CN106027410A (en) | Method and device for load balancing of controller | |
CN105634974A (en) | Route determining method and apparatus in software-defined networking | |
CN109062668A (en) | A kind of virtual network function moving method of the multipriority based on 5G access network | |
CN101803340A (en) | System and method for balancing information loads | |
CN107291544A (en) | Method and device, the distributed task scheduling execution system of task scheduling | |
CN108123891A (en) | The dynamic load balancing method realized in SDN network using distributed domain controller | |
CN108009196A (en) | Database releveling method | |
CN101651600B (en) | Method and device for updating link cost in grid network | |
Zhu et al. | Service function chain mapping with resource fragmentation avoidance | |
CN103455375A (en) | Load-monitoring-based hybrid scheduling method under Hadoop cloud platform | |
US20130067113A1 (en) | Method of optimizing routing in a cluster comprising static communication links and computer program implementing that method | |
CN108985556B (en) | Method, apparatus, equipment and the computer storage medium of flow scheduling | |
CN109635034A (en) | Training data method for resampling, device, storage medium and electronic equipment | |
CN104009904B (en) | The virtual network construction method and system of facing cloud platform big data processing | |
Penmatsa et al. | Cooperative load balancing for a network of heterogeneous computers | |
Niedermayer et al. | A simulation-based analysis of the cycle time of cluster tools in semiconductor manufacturing | |
CN107315637A (en) | The load-balancing method and device of signal processing module | |
CN107577530A (en) | Board, the method and system of balanced board memory usage | |
CN105224381A (en) | A kind of method, Apparatus and system moving virtual machine |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
EXSB | Decision made by sipo to initiate substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |