CN109547263A - Network-on-chip optimization method based on approximate calculation - Google Patents
Network-on-chip optimization method based on approximate calculation Download PDFInfo
- Publication number
- CN109547263A CN109547263A CN201811537135.6A CN201811537135A CN109547263A CN 109547263 A CN109547263 A CN 109547263A CN 201811537135 A CN201811537135 A CN 201811537135A CN 109547263 A CN109547263 A CN 109547263A
- Authority
- CN
- China
- Prior art keywords
- data
- node
- congestion
- network
- error budget
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/147—Network analysis or design for predicting network behaviour
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0803—Configuration setting
- H04L41/0823—Configuration setting characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L43/00—Arrangements for monitoring or testing data switching networks
- H04L43/08—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
- H04L43/0876—Network utilisation, e.g. volume of load or congestion level
- H04L43/0882—Utilisation of link capacity
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/10—Flow control; Congestion control
- H04L47/32—Flow control; Congestion control by discarding or delaying data units, e.g. packets or frames
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L49/00—Packet switching elements
- H04L49/10—Packet switching elements characterised by the switching fabric construction
- H04L49/109—Integrated on microchip, e.g. switch-on-chip
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/142—Network analysis or design using statistical or mathematical methods
Abstract
The present invention discloses a kind of network-on-chip optimization method based on approximate calculation comprising data cropping tool, data recoverer, volume forecasting device, global controller and local control；The data cropping tool is cut out data before data packet injects network interface, shortens data packet length；Data recoverer restores the data of loss after receiving the data packet cut；Volume forecasting device predicts the data traffic in next regulation interval according to past node communication situation；Quality requirement of the global controller for based on global information and user under global visual angle calculates the allocation optimum of each node approximate calculation, and sends control information to each node；Local control carries out the configuration of data loss rate to each data packet for waiting injection network according to the control information received.This method can use lower cost, under the premise of not violating requirement of the user to output quality, optimize to performance, the power consumption of network-on-chip.
Description
Technical field
The present invention relates to network optimisation techniques fields, and in particular to a kind of network-on-chip optimization side based on approximate calculation
Method.
Background technique
Under the background that semiconductor technology upgrading tends towards stability, many-core chip is a kind of raising system performance-power dissipation ratio
Design effectively.In face of many-core chip in various requirements such as communication bandwidth, low-power consumption, expansibility, network-on-chip is considered
It is a kind of very promising technology.Such as image procossing of many current prevalences, the application in machine learning field have number simultaneously
According to the traffic greatly and the characteristics of can tolerate certain output resultant error, therefore the approximate calculation computer new as one kind is set
Thought is counted, by relaxing the requirement to result accuracy, the performance of Lai Tigao system or the energy consumption for saving system.However, traditional
Network-on-chip design in do not have or rarely have using the error tolerance applied, and in the environment of network-on-chip, application is defeated
The control of mass is a very complicated problem.
One kind " two based on particle swarm algorithm is disclosed in the application for a patent for invention that application publication number is CN108183860A
Tie up network-on-chip adaptive routing method ", particle swarm algorithm is applied in the patent application, wherein particle swarm algorithm is based on
The optimal path for calculating data packet routing achievees the purpose that reduce network delay, balance network load.Above-mentioned patent application is by changing
Become data transfer path to carry out the network optimization, does not tolerate characteristic using the error of application program.
A kind of " calculate based on simulated annealing is improved is disclosed in the application for a patent for invention that application publication number is CN108173760A
The NoC mapping method of method " passes through enhanced simulated annealing in the patent application and calculates IP kernel to network-on-chip section
The optimal mapping of point, carrys out the power consumption of optimization system.Position of the above-mentioned patent application only by adjustment IP kernel in network-on-chip
Relationship optimizes network, not using application program to the tolerance of error in output.
Summary of the invention
The network-on-chip optimization based on approximate calculation that it is an object of the invention to overcome the deficiencies of the prior art and provide a kind of
Method, this method can maximize under the premise of not violating user to output quality requirement to error tolerance in application
It utilizes, the performance of Lai Youhua network-on-chip.
The technical solution that the present invention solves above-mentioned technical problem is as follows.
A kind of network-on-chip optimization method based on approximate calculation, this method include volume forecasting device, data cropping tool, number
According to restorer, global controller and the several parts of local control；Wherein, the volume forecasting device is run on each node,
In the beginning that each regulation is spaced according to the past operating condition of node, prediction regulates and controls the data traffic in interval, each node
Traffic prediction value is sent to global controller.
Each node of network-on-chip respectively configures a data cropping tool and a data recoverer, and data cropping tool is in number
It injects before network interface according to packet and it is cut according to the data loss rate of setting, data recoverer completes cropped data packet
The data wherein lost are restored after reception.
Global controller is assembled on main controlled node, according to global flow information and user to computation outcome quality
Requirement be that each node distributes an allocation optimum, and corresponding node is sent by network by the control information of generation；
Flow information traffic unidirectional between each node, i.e., from i-node to the traffic of j node, to the net of n node
Network then shares n (n-1) a stream；The control information include distribute to the congestion error budget of node, serializing error budget and
The data loss rate of energy minimization network congestion configures.
Local control is assembled on each node, according to the control information received, to be each i.e. by the number of injection network
The setting that data loss rate is carried out according to packet, transfers to data cropping tool to be handled.The original data sequence for including in data packet is such as
20 nybble integers, original data sequence include several data cells.
Further, local message is first collected by global controller and carries out global Optimum Regulation, local control is again
Carry out data loss rate setting each time.
Further, by relaxing requirement of the system to data precision, damage by the data transmitted in network-on-chip into
Row is cut out, and reduces the data volume to circulate in a network, and restored to ensure data in data of the destination node to loss
Integrality.
Further, the volume forecasting process are as follows: the volume forecasting device of each node collects data when application operation
Information is the amount from the node to other nodes hair data, and the autoregression mould of time series forecasting is used for by off-line training building
Type；On the basis of the autoregression model of off-line training, by by before traffic prediction value and observed value be obtain it is true
The difference of value is modified next volume forecasting as feedback.
Further, the process that the data cropping tool cuts data are as follows: data cropping tool is given according to local control
Fixed data loss rate carries out data cutting with l, interval data cell；Every l data cell of reservation, data cropping tool are just lost
A data cell is abandoned, the cut data packet of generation is smaller than raw data packets.
Further, the recovery process for losing data are as follows: to the data sequence cut, data recoverer is according to sanction
Interval l used, is inserted into the data cell of recovery in the data sequence when cutting；Every l data cell, data recoverer is just
It is inserted into a recovery value, recovery value is being averaged for the sum of its data value of adjacent cells.
Further, the specific control process of the global controller are as follows:
1) mass loss model
By emulating the mechanism of data clipper, data recoverer with software, generated in operation by application
Different loss of data ratios, and the sample of different data loss rate and mass loss is obtained, by way of linear interpolation
Mass loss model is established to each application；Mass loss model is used to estimate that the data lost to will lead to using final output knot
The error of fruit.
2) error budget
According to the result of volume forecasting and user-defined quality requirement, the volume forecasting respectively flowed on having collected each node
After value, the data volume for allowing to be restored after approximate culling is calculated in turn using mass loss model, obtain permitting in whole network
Permitted the i.e. overall error budget of the total amount of data being tailored, the tailoring, that is, node error budget reallocated to different nodes is described
The error budget of node includes congestion error budget and serializing error.
3) judge the degree of Congestion of single link
For single link, if the flow flowed through has been more than the maximum bandwidth of the link, then it is assumed that gather around chain road
Plug, and the part for being more than is defined as link congestion degree；Degree of Congestion will be flowed and be defined as all links on the path of a stream and gathered around
The sum of plug degree.
4) congestion minimizes
It is one data loss rate of each stream calculation using Greedy strategy according to overall error budget and total flow predicted value,
To reach the minimum of congestion；The degree of Congestion flowed first to each link, respectively initializes, and then starts iteration；Each iteration
The highest stream of degree of Congestion is chosen, and selects highest link congestion degree on the flowed through path of the stream, sets the communication of the stream
Loss ratio is measured, if the traffic of the setting, which loses ratio, makes link no longer congestion, the data loss rate of the stream is set as
The ratio；If it is not, being set as the maximum data Loss Rate i.e. half of cut data packet load；If total error budget
Deficiency, then the data flow only loses the loss data volume that main controlled node is distributed to；One data loss rate of every setting, it is contemplated that loss
Data volume will be deducted from overall error budget, and each link, each stream update and enters next iteration.
Each link, the iteration respectively flowed continue to that overall error budget is used up, link no longer congestion or all streams are processed
Until crossing；After iteration, as soon as each stream just possesses the configuration for minimizing data loss rate for congestion, also subscribe every
A stream will lose how many data；For some node, all streams issued expect the sum of data volume lost, that is, are allocated to this
The congestion error budget of a node.
5) serializing time delay is reduced
On the basis of minimizing congestion data Loss Rate, main controlled node continues a possibility that losing data according to each node
Height in proportion reallocates remaining overall error budget after 4) middle distribution to each node, i.e. the serializing error of node
Budget；The serializing time delay of data packet is reduced by being cut to data packet, and cutting the data volume lost every time will be from section
It is deducted in the serializing error budget of point.
So far, global control is completed, and congestion error budget, serializing error budget and congestion are minimized loss of data
Rate distributes to each corresponding node, i.e., the allocation optimum of each node.
Further, due to global controller only calculate will to each node distribution how much, each node obtains
After the budget of distribution, it can go " to consume " (deduction) these error budgets when carrying out data and damaging transmission, local control needs
In the setting of further progress data loss rate.
Further, the control process of the local control: local control is gathered around according to what global controller was distributed to
Fill in error budget, serializing error budget and congestion minimize data loss rate, to it is each wait injection network data packet into
The setting of row data loss rate；Initial Loss Rate is set as zero, if the congestion error budget of node is not zero, will lose
Rate is set as the Loss Rate of energy minimization network congestion, and the data of this part discarding are deducted from congestion error budget；It deducts
After congestion error budget, if there remains serializing error budget, then Loss Rate is set as the maximum that data cropping tool provides
Loss Rate, that is, cut data packet load half, and the data that this increased part abandons are detained from serializing error budget
It removes.
Further, the estimated data volume abandoned in the data flow that the congestion error budget sets out for each node；Sequence
Columnization error budget is that each node abandons remaining error budget after congestion error budget.
The present invention have compared with prior art it is below the utility model has the advantages that
1, the data volume that the present invention is circulated in a network by reduction improves the congestion situation of network, reduces network delay,
To optimize the overall performance and energy consumption of network.
2, the present invention is meeting matter by cutting out on the basis of volume forecasting and quality requirement data and Optimum Regulation
Amount maximizes the income of performance in the case where requiring.
Detailed description of the invention
Fig. 1 is global traffic prognostic chart of the invention.
Fig. 2 is global information control figure of the invention.
Fig. 3 is error budget flow chart of the invention.
Fig. 4 is the structural block diagram of specific embodiment.
Fig. 5 is the flow chart that data cut and restore in the present invention.
Fig. 6 is the pseudocode that congestion minimizes algorithm in the present invention.
Specific embodiment
Present invention will now be described in further detail with reference to the embodiments and the accompanying drawings, but embodiments of the present invention are unlimited
In this.
Referring to fig. 4, C indicates that the processor core in network-on-chip node, M indicate processor Cache (caches
Device), R indicates that router, NI indicate network interface.Network-on-chip optimization method based on approximate calculation of the invention includes flow
Fallout predictor, data cropping tool, data recoverer, global controller and the several parts of local control.Wherein:
As shown in Figure 1, the volume forecasting device is run on each node L, in the beginning that each regulation is spaced according to section
The past operating condition of point, predicts the communication condition of the data flow issued in the interval by the node, and be sent to master
Control node G.
Each node of network-on-chip respectively configures a data cropping tool and a data recoverer, and data cropping tool is in number
It injects before network interface according to packet and it is cut according to the data loss rate of setting, data recoverer completes cropped data packet
The data wherein lost are restored after reception.
Global controller is assembled on selected main controlled node G, according to global flow information and user to computation
The requirement of outcome quality is that each node distributes an allocation optimum, as shown in Fig. 2, the control information generated is sent by network
To corresponding node.
Local control is assembled on each node, according to the control information received, to be each i.e. by the number of injection network
The setting that data loss rate is carried out according to packet, transfers to data cropping tool to be handled.
Volume forecasting specifically: information when collecting the operation of application is used for time series forecasting by off-line training building
Autoregression model,
yAR(t)=λ1yt-1+λ2yt-2+…+λpyt-p+μ+εt
Wherein yAR(t) be the t moment that autoregression model obtains predicted value, yt-1…yt-pIt is the observed value of network flow,
λ1…λpIt is independent variable coefficient, μ is constant term, εtIt is the model error of t moment, parameter can be fitted by least square method.?
On the basis of the autoregression model of off-line training, the feedback of the difference by obtaining traffic prediction value and observed value online, to prediction
As a result it is modified,
WhereinIt is final traffic prediction value, ΔtIt is the value of feedback of t moment,It is prediction error, α is study
Rate.
The specific cutting of data are as follows: referring to Fig. 5, to an original data sequence a to be processed1,…,al-1,al,
al+1..., according to given data loss rate, l carries out data cutting to data cropping tool at certain intervals.Every l data of reservation
Unit, data cropping tool just abandon a data cell, and the cut data packet of generation is smaller than raw data packets.
Lose the specific recovery of data are as follows: referring to Fig. 5, the data sequence a cut to one1,…,al-1,al+1...,
Data recoverer is inserted into the data cell of recovery according to interval l used when cutting in the data sequence.Every l data
Unit, data recoverer are just inserted into a recovery value, such as
a′l=(al-1+al+1)/2
That is the sum of its data value of adjacent cells is averaged.Due to l >=2, the data cropping tool in this example is provided most
Big Loss Rate was 50% (working as l=2).
Overall situation control specifically:
1) mass loss model
By the way that the mechanism of data clipper, restorer is realized and is embedded into influence caused by it in the source code of application
It is simulated, modified application is run with different loss of data ratios, is collected into shaped like data loss rate, mass loss
Sample carries out completion to unknown-value by linear interpolation, establishes a mass loss model.The model is certain for estimating to lose
The data of ratio will lead to using the error in final output.
2) error budget
Requirement according to global flow information and user to computation outcome quality, in being collected into whole network
After volume forecasting result, the data volume for allowing to be restored after approximate culling can be calculated in turn using mass loss model,
That is the overall error budget of whole network.Wherein q-1For the inverse function of mass loss model, θ is the quality that user gives
Loss limitation, q-1The ratio for the data that (θ) allows to lose, vijFor the flow that i-node is communicated to j node, n is network node
Quantity,The sum of flow i.e. in whole network.One will all be deducted by carrying out approximation to data in some node every time
The fixed error budget (congestion error budget, serializing error budget or both) for distributing to the node.
3) degree of Congestion
For single link, if the flow flowed through has been more than the maximum bandwidth of the link, then it is assumed that gather around chain road
Plug, and the part for being more than is defined as link congestion degree, the degree of Congestion of link k is
Wherein c is the maximum bandwidth of link, xijFor the data loss rate on stream i → j.rijkFor formula expression two into
Amount processed flows degree of Congestion when calculating flows through the flow of a link for shielding the stream for being not passed through the link in sum formula
Then it is defined as
The sum of the degree of Congestion of all links on the path of i.e. one stream.
4) congestion minimizes
Referring to Fig. 6, figure is detailed pseudo-code of the algorithm.Collecting the flow information (volume forecasting respectively flowed on each node
Value) after, calculate the error budget of whole network.It is each using Greedy strategy according to this overall error budget and traffic prediction value
One data loss rate of stream calculation, to reach the minimum of congestion；The degree of Congestion flowed first to each link, respectively is according to side above-mentioned
Method is initialized, and iteration is then started；Each iteration chooses the highest stream of degree of Congestion, and selects the flowed through path of the stream
Upper highest link congestion degree loses the data of the stream if the traffic for setting the stream, which loses ratio, makes link no longer congestion
Mistake rate is set as the ratio；If it is not, being set as maximum data Loss Rate；If overall error budgetary shortfall, only lose total
The amount that error budget allows.One data loss rate of every setting, the caused estimated data volume lost will be pre- from overall error
It deducts and updates in calculation, enter back into next iteration.
Each link, the iteration respectively flowed continue to that overall error budget is used up, link no longer congestion or all streams are processed
Until crossing.After iteration, as soon as each stream just has the data loss rate for being used to minimize a congestion configuration, also subscribe
Each stream will lose how many data.For some node, all streams issued expect the sum of data volume lost, that is, are allocated to
The congestion error budget of this node.
5) serialization delay is reduced
Long-time row in the unconspicuous situation of network congestion, when the main determining factor of network delay is no longer congestion
Team.Wherein, the serializing time delay of data packet can be reduced by being cut to data packet, and cut and just needed error budget
Support.
As shown in figure 3, main controlled node continues to lose number according to each node on the basis of minimizing congestion data Loss Rate
According to a possibility that give remaining overall error budget allocation after distribution in 4) to each node in proportion, as distributing to its
Serialize error budget；The serializing time delay of data packet can be reduced by being cut to data packet, cut lose every time
Data volume will from serializing error budget in deduct.
So far, global control is completed, and by the corresponding congestion error budget of each node, serializing error budget and congestion
It minimizes data loss rate configuration and is sent to the node.
The detailed process of Partial controll are as follows:
To each data packet for waiting injection network, local control carries out data according to the surplus of two kinds of error budgets and loses
The selection of mistake rate.Initial Loss Rate is set as zero, if Loss Rate is temporarily set as by congestion error budget there are also residue
The Loss Rate that congestion minimizes, and deduct from congestion error budget the data of this part discarding；If serializing error budget
There are also residues, then the data for Loss Rate being set as maximum Loss Rate again, and this increased part being abandoned are from serializing error
It is deducted in budget.
Above-mentioned is the preferable embodiment of the present invention, but embodiments of the present invention are not limited by the foregoing content,
His any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present invention, should be
The substitute mode of effect, is included within the scope of the present invention.
Claims (8)
1. a kind of network-on-chip optimization method based on approximate calculation, which is characterized in that this method passes through volume forecasting device, data
Cropping tool, data recoverer, global controller and local control are realized；Wherein, the volume forecasting device is on each node
Operation, in the beginning that each regulation is spaced according to the past operating condition of this node, prediction regulates and controls the data traffic in interval, often
Traffic prediction value is sent to global controller by a node；
Each node of network-on-chip configures a data cropping tool and a data recoverer, and data cropping tool is in data packet
It is cut according to the data loss rate of setting before injection network interface, to it after the cropped data packet of data recoverer reception
The data of middle loss are restored；
Global controller is assembled on main controlled node, and computation outcome quality is wanted according to global flow information and user
It asks and distributes an allocation optimum for each node, and corresponding node is sent by network by the control information of generation；It is described
Flow information traffic unidirectional between each node is right that is, from i-node to the traffic of j nodenThe network of a node, then
It is sharedn(n- 1) a stream；The control information includes congestion error budget, serializing error budget and minimum network congestion
Data loss rate configuration；
Local control is assembled on each node, according to the control information received, for each data packet for waiting injection network
Data loss rate setting is carried out, then data cropping tool is transferred to be handled；The original data sequence for including in data packet, original number
It include several data cells according to sequence.
2. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that first by the overall situation
Controller collects local message and carries out global Optimum Regulation, and the data loss rate that local control carries out each time again is set
It sets.
3. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that volume forecasting
The prediction process of device are as follows:
Data information when the volume forecasting device of each node collects application operation sends out data from the node to other nodes
Amount is used for the autoregression model of time series forecasting by off-line training building；On the basis of the autoregression model of off-line training
On, by using before traffic prediction value and observed value be obtain true value difference as feed back, it is pre- to next flow
Survey is modified.
4. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that the data
The process that cropping tool cuts data are as follows:
The data loss rate that cropping tool gives according to local control is with intervallA data cell carries out data cutting；Every reservationl
A data cell, data cropping tool just abandon a data cell, and the cut data packet of generation is smaller than raw data packets.
5. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that the loss
The recovery process of data are as follows: to the data sequence cut, data recoverer is according to interval used when cuttingl, in the data
The data cell of recovery is inserted into sequence；EverylA data cell, data recoverer are just inserted into a recovery value, and recovery value is
The sum of the data value of its left and right adjacent cells of the data cell of recovery is averaged.
6. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that described complete
The control process of office's controller are as follows:
It establishes mass loss model to emulate the mechanism of data clipper, data recoverer with software, by applying
Different loss of data ratios is generated in operation, and obtains the sample of different data loss rate and mass loss, by linear
Interpolation carries out completion to unknown-value, establishes the mass loss model of each application；Mass loss model is used for the number for estimating to lose
According to the error that will lead to using final output；
Complete net is being collected in requirement of the error budget according to global flow information and user to computation outcome quality
After the traffic prediction value respectively flowed on network, that is, each node, calculates permission in turn using mass loss model and restored after approximate culling
Data volume, obtain the i.e. overall error budget of the total amount of data for allowing to be tailored in whole network, cutting to different nodes of reallocating
Discretion, that is, node error budget, the error budget of the node include congestion error budget and serializing error；
Judge the degree of Congestion of single link for single link, if the flow flowed through has been more than the maximum bandwidth of the link,
Then think that there is congestion in chain road, and the part for being more than is defined as link congestion degree；Degree of Congestion will be flowed and be defined as a stream
The sum of the degree of Congestion of all links on path；
Congestion minimizes
It is one data loss rate of each stream calculation using Greedy strategy, to reach according to overall error budget and total flow predicted value
To the minimum of congestion；The degree of Congestion flowed first to each link, respectively initializes, and then starts iteration；Each iteration is chosen
The one highest stream of degree of Congestion, and highest link congestion degree, the traffic for setting the stream on the flowed through path of the stream is selected to lose
The data loss rate of the stream is set as the ratio if the traffic of the setting, which loses ratio, makes link no longer congestion by mistake ratio
Example；If it is not, being set as the maximum data Loss Rate i.e. half of cut data packet load；If total error budget is not
Foot, then the data flow only loses the loss data volume that main controlled node is distributed to；One data loss rate of every setting, it is contemplated that the number of loss
It will be deducted from overall error budget according to amount, each link, each stream update and enters next iteration；Each link, the iteration respectively flowed
Until continueing to that overall error budget is used up, link no longer congestion or all streams are processed；After iteration, each stream is just
As soon as possessing the data loss rate for being used to minimize a congestion configuration, also subscribe each stream will lose how many data；For some
Node, all streams issued expect the sum of data volume lost, that is, are allocated to the congestion error budget of this node；
Reduce serialization delay
On the basis of minimizing congestion data Loss Rate, main controlled node continues a possibility that losing data height according to each node
Give remaining overall error budget allocation after 4) middle distribution to each node, i.e. the serializing error budget of node in proportion；
The serializing time delay of data packet is reduced by being cut to data packet, and cutting the data volume lost every time will be from the sequence of node
It is deducted in columnization error budget；
So far, global control is completed, and congestion error budget, serializing error budget and congestion are minimized data loss rate point
The each corresponding node of dispensing, i.e., the allocation optimum of each node.
7. a kind of network-on-chip optimization method based on approximate calculation as described in claim 1, which is characterized in that the part
The control process of controller are as follows:
Congestion error budget, serializing error budget and the congestion that local control is distributed to according to global controller minimize number
The setting of data loss rate is carried out to each data packet for waiting injection network according to Loss Rate；Initial Loss Rate is set as zero,
If the congestion error budget of node is not zero, Loss Rate is set as to the Loss Rate of energy minimization network congestion, and from gathering around
The data of this part discarding are deducted in plug error budget；After deducting congestion error budget, if there remains serializing error budget,
Loss Rate is set as maximum Loss Rate that data cropping tool the provides i.e. half of cut data packet load again, and by it is increased this
The data that part abandons are deducted from serializing error budget.
8. a kind of network-on-chip optimization method based on approximate calculation as claimed in claim 7, which is characterized in that the congestion
The estimated data volume abandoned in the data flow that error budget sets out for each node；Error budget is serialized as the discarding of each node
Remaining error budget after congestion error budget.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811537135.6A CN109547263B (en) | 2018-12-15 | 2018-12-15 | Network-on-chip optimization method based on approximate calculation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811537135.6A CN109547263B (en) | 2018-12-15 | 2018-12-15 | Network-on-chip optimization method based on approximate calculation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109547263A true CN109547263A (en) | 2019-03-29 |
| CN109547263B CN109547263B (en) | 2021-08-20 |
Family
ID=65856475
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811537135.6A Active CN109547263B (en) | 2018-12-15 | 2018-12-15 | Network-on-chip optimization method based on approximate calculation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109547263B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113467938A (en) * | 2021-06-18 | 2021-10-01 | 山东云海国创云计算装备产业创新中心有限公司 | Bus resource allocation method, device and related equipment |
| CN113839878A (en) * | 2021-09-26 | 2021-12-24 | 南京宁麒智能计算芯片研究院有限公司 | Data-intensive application-oriented network-on-chip approximate communication system |
| CN115277563A (en) * | 2022-06-07 | 2022-11-01 | 南京大学 | On-chip network approximate control system based on offline reinforcement learning |
| CN113467938B (en) * | 2021-06-18 | 2024-05-17 | 山东云海国创云计算装备产业创新中心有限公司 | Bus resource allocation method and device and related equipment |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102065019A (en) * | 2011-01-21 | 2011-05-18 | 西安电子科技大学 | IP (Internet Protocol) core fast mapping method for network on chip based on region division |
| CN103814563A (en) * | 2011-09-21 | 2014-05-21 | 富士通株式会社 | Data communication apparatus, data transmission method, and computer system |
| CN104243242A (en) * | 2014-09-26 | 2014-12-24 | 厦门亿联网络技术股份有限公司 | Network dithering algorithm based network congestion reflection method |
| CN106453109A (en) * | 2016-10-28 | 2017-02-22 | 南通大学 | Network-on-chip communication method and network-on-chip router |
| US20170063634A1 (en) * | 2013-05-03 | 2017-03-02 | Netspeed Systems | Heterogeneous soc ip core placement in an interconnect to optimize latency and interconnect performance |
| CN107395503A (en) * | 2017-08-25 | 2017-11-24 | 东南大学 | A kind of network-on-chip method for routing based on linear programming |
| CN108156090A (en) * | 2018-03-15 | 2018-06-12 | 北京邮电大学 | Based on the optimal arrival rate method for routing of satellite DTN web impact factors |
| CN108965024A (en) * | 2018-08-01 | 2018-12-07 | 重庆邮电大学 | A kind of virtual network function dispatching method of the 5G network slice based on prediction |
-
2018
- 2018-12-15 CN CN201811537135.6A patent/CN109547263B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102065019A (en) * | 2011-01-21 | 2011-05-18 | 西安电子科技大学 | IP (Internet Protocol) core fast mapping method for network on chip based on region division |
| CN103814563A (en) * | 2011-09-21 | 2014-05-21 | 富士通株式会社 | Data communication apparatus, data transmission method, and computer system |
| US20170063634A1 (en) * | 2013-05-03 | 2017-03-02 | Netspeed Systems | Heterogeneous soc ip core placement in an interconnect to optimize latency and interconnect performance |
| CN104243242A (en) * | 2014-09-26 | 2014-12-24 | 厦门亿联网络技术股份有限公司 | Network dithering algorithm based network congestion reflection method |
| CN106453109A (en) * | 2016-10-28 | 2017-02-22 | 南通大学 | Network-on-chip communication method and network-on-chip router |
| CN107395503A (en) * | 2017-08-25 | 2017-11-24 | 东南大学 | A kind of network-on-chip method for routing based on linear programming |
| CN108156090A (en) * | 2018-03-15 | 2018-06-12 | 北京邮电大学 | Based on the optimal arrival rate method for routing of satellite DTN web impact factors |
| CN108965024A (en) * | 2018-08-01 | 2018-12-07 | 重庆邮电大学 | A kind of virtual network function dispatching method of the 5G network slice based on prediction |
Non-Patent Citations (1)
| Title |
|---|
| 陈树江等: "片上互连网络(NOCs)低功耗的研究现状及发展趋势", 《电脑知识与技术》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113467938A (en) * | 2021-06-18 | 2021-10-01 | 山东云海国创云计算装备产业创新中心有限公司 | Bus resource allocation method, device and related equipment |
| CN113467938B (en) * | 2021-06-18 | 2024-05-17 | 山东云海国创云计算装备产业创新中心有限公司 | Bus resource allocation method and device and related equipment |
| CN113839878A (en) * | 2021-09-26 | 2021-12-24 | 南京宁麒智能计算芯片研究院有限公司 | Data-intensive application-oriented network-on-chip approximate communication system |
| CN115277563A (en) * | 2022-06-07 | 2022-11-01 | 南京大学 | On-chip network approximate control system based on offline reinforcement learning |
| CN115277563B (en) * | 2022-06-07 | 2024-03-19 | 南京大学 | Network-on-chip approximation control system based on offline reinforcement learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109547263B (en) | 2021-08-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Hussein et al. | Efficient task offloading for IoT-based applications in fog computing using ant colony optimization | |
| CN108965024B (en) | Virtual network function scheduling method based on prediction for 5G network slice | |
| CN110365753B (en) | Low-delay load distribution method and device for Internet of things service based on edge calculation | |
| Zhang et al. | Decomposable intelligence on cloud-edge iot framework for live video analytics | |
| CN107172166A (en) | The cloud and mist computing system serviced towards industrial intelligentization | |
| CN109905470A (en) | A kind of expense optimization method for scheduling task based on Border Gateway system | |
| CN113810233B (en) | Distributed computation unloading method based on computation network cooperation in random network | |
| CN105718364A (en) | Dynamic assessment method for ability of computation resource in cloud computing platform | |
| CN112118312B (en) | Network burst load evacuation method facing edge server | |
| CN113238847B (en) | Distribution and scheduling method based on distributed network environment and capable of distributing tasks | |
| CN109547263A (en) | Network-on-chip optimization method based on approximate calculation | |
| CN110287034A (en) | The dynamic task allocation method of energy-delay balance in a kind of chargeable mobile edge calculations | |
| CN105376157A (en) | Method and device for routing path selection | |
| Zheng et al. | Learning based task offloading in digital twin empowered internet of vehicles | |
| Marbukh | Towards fog network utility maximization (FoNUM) for managing fog computing resources | |
| CN110460538A (en) | Data transmission method in a kind of Internet of Things of software definition | |
| CN110290539A (en) | Resource allocation device and its working method based on the application of the mobile augmented reality of user's mobile awareness and resource reservation | |
| CN111309472A (en) | Online virtual resource allocation method based on virtual machine pre-deployment | |
| CN105393518B (en) | Distributed cache control method and device | |
| CN113159539A (en) | Joint green energy scheduling and dynamic task allocation method in multilayer edge computing system | |
| CN110753101A (en) | Low-energy-consumption computing node selection and computing task allocation method in edge computing | |
| Liu et al. | Scalable traffic management for mobile cloud services in 5G networks | |
| Pan et al. | Satellite network load balancing strategy for SDN/NFV collaborative deployment | |
| Bertolli et al. | Analyzing memory requirements for pervasive grid applications | |
| Li et al. | Utility driven dynamic resource allocation using competitive markets in computational grid |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |