CN-121984921-A - Processing method and device for setting flow scheduling flag based on SDN flow prediction

Abstract

The embodiment of the invention relates to a processing method and a device for setting a flow scheduling mark based on SDN flow prediction, wherein the method comprises the steps of confirming a switching layer topological graph of a hardware switching layer; the method comprises the steps of configuring a flow collection library for each switching node, continuously sampling the flow of each node and sampling data based on the flow collection library, constructing a flow prediction model, periodically constructing a latest data set based on all the flow collection libraries to train the model, predicting the flow of the switching node set in the future period by using the flow prediction model to obtain a node prediction set, and identifying whether the flow scheduling needs to be started or not based on the node prediction set to obtain a corresponding flow scheduling mark and scheduling node set. The invention can reduce the risk of network blocking and effectively assist in improving the dispatching precision, efficiency and timeliness of flow dispatching.

Inventors

• CHE CHENXIN
• Jia Shumeng

Assignees

• 北京兴云致雨科技有限公司

Dates

Publication Date

20260505

Application Date

20260318

Claims (10)

1. 1. A processing method for setting a flow scheduling flag based on SDN flow prediction, the method comprising: The method comprises the steps of confirming a physical host topological relation for carrying out network data exchange on service flow data of specified service in an SDN network hardware exchange layer to obtain an exchange layer topological diagram, wherein the SDN network hardware exchange layer comprises a plurality of physical hosts for network data exchange, the total number of the physical hosts of the SDN network hardware exchange layer is N V , each physical host is directly connected with one or more other physical hosts respectively, the total number of direct connection lines of the physical hosts in the SDN network hardware exchange layer is N E , the exchange layer topological diagram comprises an exchange node set V and a physical edge set E, the exchange node set V consists of N V exchange nodes V i , an index i is not more than 1 and not more than N V , the exchange nodes V i are in one-to-one correspondence with the physical hosts, each exchange node V i corresponds to a unique node identifier, and each exchange node V i corresponds to a maximum node flow corresponding to the specified service The node types of the switching node v i comprise edge switching nodes and internal switching nodes, wherein the edge switching nodes are used for carrying out network data exchange on each service end outside a switching layer and an internal network in the switching layer, the internal switching nodes are used for processing network data exchange of the internal network in the switching layer, the physical edge set E consists of N E physical edges E j , the index j is more than or equal to 1 and less than or equal to N E , the physical edges E j are in one-to-one correspondence with a direct connection line of a physical host, and each physical edge E j is used for connecting two switching nodes v i ; Configuring a flow collection library corresponding to the appointed service for each switching node V i of the switching node set V, continuously sampling the real-time data exchange flow of the appointed service on the physical host corresponding to each switching node V i according to a preset sampling time interval delta t and storing the sampled data into the corresponding flow collection library; The method comprises the steps of establishing a flow prediction model for predicting the service data flow of a switching node, establishing a latest first data set based on all flow acquisition libraries at regular intervals, training the flow prediction model based on the latest first data set, and predicting the service data flow of a future period and outputting a corresponding predicted flow sequence Y according to service flow characteristics X flow and basic characteristics X base input by the model; The method comprises the steps of obtaining a corresponding node prediction set G by predicting the future time period flow of a switching node set V through the flow prediction model at any time after each training of the model is finished, identifying whether a hardware switching layer of the SDN network needs to start flow scheduling or not based on the node prediction set G to obtain a corresponding flow scheduling mark and a scheduling node set, wherein the flow scheduling mark comprises starting scheduling and non-starting scheduling, the corresponding scheduling node set consists of one or more switching nodes V i in the switching node set V when the flow scheduling mark is starting scheduling, and the corresponding scheduling node set is empty when the flow scheduling mark is non-starting scheduling.
2. 2. The processing method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 1, wherein, The service types of the appointed service at least comprise a short message service, a data service, a video service and a streaming media service; The service flow characteristic X flow is formed by sequencing N X data flows X t in time sequence, sampling point indexes t are less than or equal to 1 and less than or equal to N X , the total number of sampling points N X is a preset positive odd number, the time interval of sampling points corresponding to each two adjacent data flows X t is fixed to be the sampling time interval delta t, the time length corresponding to the service flow characteristic X flow is preset time length L X ,L X =(N X -1)×△t,L X which is less than 24 hours, and the data flow X t is the service data flow generated by the appointed service at the t-th moment by the switching node; The basic features X base comprise a time feature X date , a holiday feature X holiday , a workday feature X work , a region feature X area , a season feature X season and a business type feature X type ; The time characteristic X date is formed by coding a time stamp, the time information of the time characteristic X date at least comprises year, month, day, hour, minute, second and millisecond, and the time characteristic X date is used for representing the time state corresponding to the first data flow X t of the service flow characteristic X flow ; The holiday characteristic X holiday is a specific holiday type code, the value range of the holiday type code is a preset holiday type code set, the holiday type code set consists of a non-holiday code and a plurality of holiday type codes, and the holiday characteristic X holiday is used for representing the holiday state of the corresponding day of the first data flow X t of the service flow characteristic X flow ; The working day characteristic X work is a binary state characteristic and comprises a yes state and a no state, wherein the working day characteristic X work is used for representing whether the corresponding day of the first data flow X t of the service flow characteristic X flow is a working day or not; The regional characteristic X area comprises longitude, latitude, country code, state/province code and city code, wherein the value ranges of the country code, the state/province code and the city code are respectively a preset country code set, a state/province code set and a city code set, and the regional characteristic X area is used for representing the physical host deployment position and the regional information of the switching node corresponding to the service flow characteristic X flow ; The seasonal characteristic X season is a specific seasonal type code, the value range of the seasonal type code is a preset seasonal code range [ spring code, summer code, autumn code and winter code ]; The service type characteristic x type is a service type code corresponding to the appointed service, the value range of the service type code is a preset service type code set, and each code of the service type code set corresponds to a service type; the predicted flow sequence Y is composed of N Y predicted flows The system is formed by sequencing according to time sequence, the time length corresponding to the predicted flow sequence Y is preset time length L Y ,L Y =L X /2=(N X -1) X delta t/2, sampling point index t ' ≤N Y which is less than or equal to 1, total number of predicted points N Y =1+L Y /△t=1+(N X -1)/2, and the last data flow of the service flow characteristic X flow Corresponding time T x-e to the first predicted flow of the predicted flow sequence Y The corresponding relation of the time T y-s is T y-s =T x-e +L X /2, and the predicted flow rate Traffic data traffic generated by said designated traffic at a future time t ' for the switching node; The first data set comprises a plurality of first data records, the first data records comprise the service flow characteristics X flow , the basic characteristics X base and a label traffic sequence Y tag , and the label traffic sequence Y tag comprises N Y label traffic Sequencing according to time sequence; The node prediction set G comprises N V predicted traffic sequences Y i , and the predicted traffic sequences Y i comprise N Y predicted traffic sequences 。
3. 3. The processing method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 2, wherein, The first model input end of the flow prediction model is used for receiving the service flow characteristic X flow , the second model input end is used for receiving the basic characteristic X base , and the model output end is used for outputting the predicted flow sequence Y; The flow prediction model comprises a first LSTM model, a first MLP model, a second MLP model, a first fusion layer, a third MLP model and a first linear layer; The input end of the first LSTM model is connected with the input end of the first model, the output end of the first LSTM model is connected with the input end of the second MLP model, the output end of the second MLP model is connected with the first input end of the first fusion layer, the input end of the first MLP model is connected with the input end of the second model, the output end of the first fusion layer is connected with the input end of the third MLP model, the output end of the third MLP model is connected with the input end of the first linear layer, and the output end of the first linear layer is connected with the output end of the model; The first LSTM model is used for carrying out sequence feature coding on the service flow feature X flow to obtain a corresponding feature sequence { h t }, and the last feature vector of the feature sequence The second MLP model is sent as a corresponding feature vector H flow , wherein the feature sequence { H t } consists of N X feature vectors H t , the feature vectors H t are in one-to-one correspondence with the data traffic x t , and the shape of each of the feature vectors H t and H flow is 1×D LSTM ,D LSTM , which is the feature dimension output by the first LSTM model; the first MLP model is used for carrying out feature coding on the basic feature X base to obtain a corresponding feature vector Z base , and sending the corresponding feature vector Z base to the first fusion layer, wherein the shape of the feature vector Z base is 1 xD 1 ,D 1 , and the feature vector is the feature dimension output by the first MLP model; The second MLP model is used for carrying out feature coding on the feature vector H flow to obtain a corresponding feature vector Z flow , and sending the corresponding feature vector Z flow to the first fusion layer, wherein the shape of the feature vector Z flow is 1 xD 2 ,D 2 , and the feature vector is the feature dimension output by the second MLP model; The first fusion layer is configured to perform feature fusion on the feature vector Z base and the feature vector Z flow in a feature stitching manner to obtain a corresponding feature vector Z fuse , and send the feature vector Z fuse to the third MLP model, where a shape of the feature vector Z fuse is 1× (D 1 +D 2 ); The third MLP model is used for carrying out feature coding on the feature vector Z fuse to obtain a corresponding feature vector Z out , and sending the corresponding feature vector Z out to the first linear layer, wherein the shape of the feature vector Z out is 1×D 3 ,D 3 , and the feature vector is the feature dimension output by the third MLP model; The first linear layer is implemented based on a full-connection layer, and is used for performing linear regression prediction according to the feature vector Z out to obtain the corresponding predicted flow sequence Y.
4. 4. The method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 2, wherein the periodically constructing the latest first data set based on all traffic collection libraries specifically includes: Performing one round of traversal on all the exchange nodes V i of the exchange node set V periodically, taking the currently traversed exchange node V i as a current node in the round of traversal, taking the flow collection library corresponding to the current node as a current collection library, extracting the node data flow with the time interval between the collection time stamp and the current time meeting the preset latest appointed time in the current collection library to form a corresponding current node sample sequence, wherein the latest appointed time is in month units, and the current node sample sequence is formed according to the preset sliding time L swip , The segment duration L piece is used for carrying out sliding slicing on the current node sample sequence to obtain a plurality of corresponding slice segment sequences, 0< L swip ≤L X ,L piece =2L X , taking each slice segment sequence as a current segment, composing the corresponding service flow characteristic X flow by the first N X node data flows of the current segment, composing the corresponding label flow sequence Y tag by the last N Y node data flows of the current segment, and based on the time state corresponding to the first node data flow of the current segment, Setting a group of corresponding time characteristics x date , holiday characteristics x holiday , holiday states, working day states and season states, The working day characteristic x work and the seasonal characteristic x season , and set the corresponding regional characteristic x area based on the physical host deployment position and regional information corresponding to the current node, and set the corresponding service type characteristic x type based on the service type of the specified service, and the time characteristic x date , the time characteristic x type corresponding to the current segment, The holiday characteristic x holiday , the workday characteristic x work , the region characteristic x area , the season characteristic x season , The business type characteristic X type forms the corresponding basic characteristic X base , the business flow characteristic X flow corresponding to the current segment, the basic characteristic X base and the label flow sequence Y tag form the corresponding first data record, and when the round of traversal is finished, all the first data records obtained by the round of traversal form the latest first data set.
5. 5. The method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 2, wherein the training the traffic prediction model based on the latest first data set specifically includes: Step 51, randomly dividing the first data set into two sub data sets based on a preset first dividing ratio, and recording the two sub data sets as a corresponding first training set and a first evaluation set; The first training set and the first evaluation set are composed of a plurality of first data records, wherein the total record number of the first training set is N tr , the total record number of the first evaluation set is N av , and the total record ratio N tr :N av of the first training set and the first evaluation set is matched with the first segmentation proportion; the tag traffic sequences Y tag of each of the first data records of the first training set are noted as corresponding Each of the tag traffic sequence Y tag Is marked as corresponding The index q is less than or equal to 1 and less than or equal to N tr , and the label flow sequence Y tag of each first data record of the first evaluation set is recorded as corresponding Each of the tag traffic sequence Y tag Is marked as corresponding Index q ' ≤N av is less than or equal to 1; Step 52, inputting the traffic flow characteristics X flow and the basic characteristics X base of each first data record of the first training set into the traffic prediction model for processing, and recording the predicted traffic sequence Y output by the current model processing as a corresponding traffic sequence And sequence the current predicted traffic Each of the predicted flows of (1) Is marked as corresponding And from each predicted traffic sequence And its corresponding tag traffic sequence Forming a corresponding first predictor-tag pair; Step 53, substituting the N tr first prediction-label pairs into a preset first model loss function L M1 to calculate a corresponding first loss value; wherein, the first model loss function L M1 is: ; step 54, identifying whether the first loss value meets a preset first loss value range, if yes, turning to step 55, if not, carrying out one-round modulation on model parameters of the flow prediction model based on a preset first model optimizer towards the direction of enabling the first model loss function L M1 to reach the minimum value, and returning to step 52 after the end of the one-round modulation; The first model optimizer comprises an Adam optimizer and an SGD optimizer; Step 55, inputting the traffic flow characteristics X flow and the basic characteristics X base of each of the first data records of the first evaluation set into the traffic prediction model for processing, and recording the predicted traffic sequence Y output by the current model processing as a corresponding traffic sequence And sequence the current predicted traffic Each of the predicted flows of (1) Is marked as corresponding And from each predicted traffic sequence And its corresponding tag traffic sequence Substituting N av second prediction-label pairs into a preset first model evaluation function F M1 to calculate a corresponding first evaluation value; Wherein the first model evaluation function F M1 is; ; Step 56, identifying whether the first evaluation value meets a preset first evaluation value range, if not, returning to step 51, and if so, stopping training and confirming that model training is finished.
6. 6. The method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 2, wherein the predicting the future period traffic of the switching node set V using the traffic prediction model to obtain a corresponding node prediction set G specifically includes: Taking the current time as the current ending time T end , setting corresponding current starting time T sta =T end -L X -Deltat based on the preset duration L X and the sampling time interval Deltat, and setting corresponding current latest time period [ T sta ,T end ]; And performing one round of traversal on all the switching nodes V i of the switching node set V, and in the round of traversal, marking the currently traversed switching node V i and the corresponding traffic collection library as corresponding current nodes, The method comprises the steps of obtaining a current flow library, extracting N X latest node data flows with the collection time stamp meeting the current latest time interval [ T sta ,T end ] in the current flow library to form a corresponding current flow sequence, taking the current flow sequence as a corresponding service flow characteristic X flow , and based on a time state corresponding to the collection time stamp of the first node data flow of the current flow sequence, Setting the corresponding time characteristics x date , the holiday characteristics x holiday , the working day characteristics x work , the working day conditions and the season conditions, The season feature x season , the region feature x area corresponding to the physical host deployment position and region information corresponding to the current node, the service type feature x type corresponding to the service type of the appointed service, the time feature x date corresponding to the current node, The holiday characteristic x holiday , the workday characteristic x work , the region characteristic x area , the season characteristic x season , The service type characteristic X type forms the corresponding basic characteristic X base , the service flow characteristic X flow and the basic characteristic X base corresponding to the current node are input into the flow prediction model to be processed, the predicted flow sequence Y output by the current model processing is used as the corresponding predicted flow sequence Y i , and when the round of traversal is finished, the corresponding node prediction set G is formed by all obtained predicted flow sequences Y i .
7. 7. The method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in claim 2, wherein the identifying whether the SDN network hardware switching layer needs to start traffic scheduling based on the node prediction set G obtains a corresponding traffic scheduling flag and scheduling node set specifically includes: Step 71, based on each of the predicted traffic in the node prediction set G The maximum node flow corresponding to the node flow Calculating corresponding node saturation ; Wherein the node saturation The calculation mode of (a) is as follows: ; Step 72, for all of the node saturation levels Identifying whether the traffic scheduling mark meets the preset saturation range, if so, setting the corresponding traffic scheduling mark as not starting scheduling and setting the corresponding scheduling node set as empty, and if not, saturating the nodes exceeding the saturation range Counting the total number of the corresponding traffic scheduling marks to obtain a first total number, identifying whether the first total number exceeds a preset total number threshold of nodes with the saturation exceeding the standard, if so, setting the corresponding traffic scheduling marks as starting scheduling, and judging whether all the nodes with the saturation exceeding the saturation range are saturated And the corresponding switching node v i forms the corresponding scheduling node set, if not, the corresponding flow scheduling mark is set to be not started for scheduling, and the corresponding scheduling node set is set to be empty.
8. 8. An apparatus for performing the processing method for setting a traffic scheduling flag based on SDN traffic prediction as set forth in any one of claims 1-7, wherein the apparatus comprises a traffic topology identification module, a node sampling module, a model construction and training module, and a traffic prediction and scheduling flag processing module; the service topology identification module is used for confirming a physical host topological relation for carrying out network data exchange on service flow data of a specified service in an SDN network hardware exchange layer to obtain an exchange layer topological diagram, the SDN network hardware exchange layer comprises a plurality of physical hosts for network data exchange, the total number of the physical hosts of the SDN network hardware exchange layer is N V , each physical host is directly connected with one or more other physical hosts, the total number of a direct connection line of the physical hosts in the SDN network hardware exchange layer is N E , the exchange layer topological diagram comprises an exchange node set V and a physical edge set E, the exchange node set V comprises N V exchange nodes V i , an index i is not more than N V , the exchange nodes V i are in one-to-one correspondence with the physical hosts, each exchange node V i is corresponding to one unique node identifier, each exchange node V i is corresponding to one maximum node flow corresponding to the specified service The node types of the switching node v i comprise edge switching nodes and internal switching nodes, wherein the edge switching nodes are used for carrying out network data exchange on each service end outside a switching layer and an internal network in the switching layer, the internal switching nodes are used for processing network data exchange of the internal network in the switching layer, the physical edge set E consists of N E physical edges E j , the index j is more than or equal to 1 and less than or equal to N E , the physical edges E j are in one-to-one correspondence with a direct connection line of a physical host, and each physical edge E j is used for connecting two switching nodes v i ; The node sampling module is used for configuring a flow collection library corresponding to the appointed service for each switching node V i of the switching node set V, continuously sampling the real-time data exchange flow of the appointed service on the physical host corresponding to each switching node V i according to a preset sampling time interval delta t and storing the sampled data into the corresponding flow collection library, wherein each flow collection library comprises a plurality of node data flows, and each node data flow corresponds to one collection time stamp; The model construction and training module is used for constructing a flow prediction model for predicting the service data flow of the switching node, constructing the latest first data set based on all the flow acquisition libraries at regular intervals, and training the flow prediction model based on the latest first data set, wherein the flow prediction model is used for predicting the service data flow of the future period according to the service flow characteristics X flow and the basic characteristics X base which are input by the model and outputting a corresponding predicted flow sequence Y; The flow prediction and scheduling flag processing module is used for predicting the flow of the switching node set V in the future period by using the flow prediction model at any moment after each training of the model is finished to obtain a corresponding node prediction set G, identifying whether the SDN network hardware switching layer needs to start flow scheduling or not based on the node prediction set G to obtain a corresponding flow scheduling flag and a scheduling node set, wherein the flow scheduling flag comprises a start scheduling and a non-start scheduling, when the flow scheduling flag is the start scheduling, the corresponding scheduling node set is composed of one or more switching nodes V i in the switching node set V, and when the flow scheduling flag is the non-start scheduling, the corresponding scheduling node set is empty.
9. 9. An electronic device comprising a memory, a processor, and a transceiver; the processor being operative to couple with the memory, read and execute instructions in the memory to implement the method of any one of claims 1-7; the transceiver is coupled to the processor and is controlled by the processor to transmit and receive messages.
10. 10. A computer readable storage medium storing computer instructions which, when executed by a computer, cause the computer to perform the method of any one of claims 1-7.

Description

Processing method and device for setting flow scheduling flag based on SDN flow prediction Technical Field The invention relates to the technical field of data processing, in particular to a processing method and a device for setting a flow scheduling flag based on SDN flow prediction. Background The software control layer of the software defined network (Software Defined Network, SDN) is separate from the hardware switching layer. The software control layer may distinguish the data exchange channels of any physical host (e.g., switch) of the hardware exchange layer by traffic type and set network parameters (e.g., traffic bandwidth) for each type of traffic data exchange channel. The decoupling architecture enables a network administrator to adjust node forwarding flow according to service requirements and optimize resource allocation. Currently, a flow scheduling policy deployed on an SDN is implemented based on an event activation manner, where a corresponding event activation flag (e.g., a flow scheduling flag) is configured for a flow scheduling operation flow, and when the flag is set to an active state (e.g., a start scheduling state), the corresponding flow scheduling operation flow is started/activated. However, how to set the traffic scheduling flag, the current conventional scheme is mostly implemented based on a passive response mechanism. The passive response mechanism is simply that each node of the exchange layer is monitored for flow, whether the monitored flow exceeds the limit is identified, and a flow scheduling flag is set to be in a starting scheduling state after the overrun is confirmed. The time delay of the passive response mechanism is found through application practice, so that the traffic scheduling strategy is difficult to cope with the transient congestion caused by the burst traffic, and large-scale packet loss and delay are often caused before the scheduling is effective. Disclosure of Invention The invention aims at overcoming the defects of the prior art and provides a processing method, a device, electronic equipment and a computer readable storage medium for setting a flow scheduling flag based on SDN flow prediction. The invention constructs a flow prediction model for predicting the data flow of the node in the future period according to the traffic flow sensing characteristic (traffic flow characteristic X flow) and the space-time background characteristic (basic characteristic X base) of the switching node, and optimally trains the model by periodically utilizing the latest sampling data of the SDN network. The traffic prediction model incorporates a series of spatiotemporal background features (time, holiday, workday, longitude, latitude, country, state/province, city, season, traffic type) related to periodic/regional network congestion in addition to using traffic flow awareness features related to a given traffic network topology in the prediction. And a sign setting mechanism for active prediction is provided based on a flow prediction model, wherein future flow of the switching nodes of the whole network is predicted through the flow prediction model, and whether scheduling is started or not and the scheduling nodes are positioned are judged in advance based on prediction information. The method and the system can effectively identify and predict potential network congestion events caused by periodic fluctuation, regional characteristics and burst modes (abnormal flow change trend) appearing in service flow perception through multidimensional characteristic fusion, can carry out advanced setting on a flow scheduling mark and can accurately position a flow scheduling object through a mark setting mechanism which is actively foreseen, can reduce network blocking risk, and can effectively assist in improving scheduling precision, scheduling efficiency and scheduling timeliness. To achieve the above object, a first aspect of the present invention provides a processing method for setting a traffic scheduling flag based on SDN traffic prediction, where the method includes: The method comprises the steps of confirming a physical host topological relation for carrying out network data exchange on service flow data of specified service in an SDN network hardware exchange layer to obtain an exchange layer topological diagram, wherein the SDN network hardware exchange layer comprises a plurality of physical hosts for network data exchange, the total number of the physical hosts of the SDN network hardware exchange layer is N V, each physical host is directly connected with one or more other physical hosts respectively, the total number of direct connection lines of the physical hosts in the SDN network hardware exchange layer is N E, the exchange layer topological diagram comprises an exchange node set V and a physical edge set E, the exchange node set V consists of N V exchange nodes V i, an index i is not more than 1 and not more than N V, the exchange nodes V i are in o