A Data Splitting Method Based on Link Congestion Awareness
By implementing conservation-based traffic splitting intervention and orthogonal response relationship correction in multi-path data splitting, the problem of distinguishing between congestion response caused by path migration and changes in network background is solved, improving the reliability and consistency of the data splitting process and ensuring the stability of link status.
Patent Information
- Application Number
- CN202611527586.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing multipath data offloading methods struggle to distinguish between congestion responses caused by path migrations and changes in network context, leading to inconsistencies between offloading control and link status. Furthermore, the response relationships for small path migrations are not applicable to large-scale adjustments, potentially causing link status to exceed the response range.
By acquiring the diversion status of candidate paths, determining the congestion observation location and uniform sampling period, implementing conservatism diversion intervention, utilizing orthogonal response relationships and identifiable response boundaries for data migration, and combining actual time-domain congestion changes for correction, the reliability and consistency of the diversion process are ensured.
It improves the reliability of judging the relationship between path migration direction and link congestion changes, reduces the possibility of diversion adjustments exceeding the verified response range, and enables feedback correction based on actual link status, ensuring the stability and accuracy of the data diversion process.
Smart Images

Figure CN123123268A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data communication network traffic scheduling and congestion control technology, and more specifically, to a data diversion method based on link congestion awareness. Background Technology
[0002] This invention relates to multi-path data scheduling and link congestion control technology in data communication networks. In a network where the same data to be transmitted can reach the destination node through multiple candidate paths, the data transmission status in the network can be adjusted by changing the proportion of data carried by different candidate paths.
[0003] In existing multi-path data offloading processes, the data ratio of candidate paths can usually be adjusted based on path load, queue occupancy, or other congestion states. However, candidate paths may share common links, be different physical network segments, or be tunnel network segments. A single change in path ratio will result in corresponding state changes at different congestion observation locations and at different subsequent times. Normal network operation itself also involves changes in service arrival and measurement quantification. If the actual congestion response caused by path migration cannot be distinguished from these background changes, it is difficult to determine which offloading directions can form identifiable congestion mitigation effects, and which path changes fall within the low-response range near the current operating state. Furthermore, the response relationship obtained from small path migrations may not be applicable to larger offloading adjustments. If the migration amount is directly expanded based on local identification results, the actual link state may exceed the range of the obtained response data. Even if a formal offloading is completed based on existing response relationships, changes in network operating states may still cause deviations between the expected temporal congestion changes and the actual changes. Without checking for these deviations and re-identifying the actual adjustment directions, it is difficult to maintain consistency between offloading control and the current link state when continuing to use the original response relationships.
[0004] Therefore, it is necessary to propose a data diversion method based on link congestion awareness, which controls the multi-path data diversion process through real business conservation diversion intervention, congestion response identifiability judgment, local diversion constraints, and actual response correction. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a data diversion method based on link congestion awareness.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A data offloading method based on link congestion awareness specifically includes the following steps: Obtain the candidate path diversion status of the same data to be transmitted, determine the congestion observation location, unified sampling period and time base; for the congestion observation location where the actual queue occupancy and available queue capacity can be obtained, determine the safe queue boundary by the difference between the queue capacity and the maximum single-cycle arrival data volume during the period without manual path adjustment, and determine the safe congestion status by the ratio of the safe queue boundary to the available queue capacity. Under the condition that the total allocation ratio of candidate paths remains unchanged, a conservation diversion intervention is implemented on the data migration direction between linearly independent paths, and the orthogonal response relationship is determined based on the congestion status changes of each congestion observation location at multiple sampling times. The identifiable response boundary is determined based on the background changes and measured quantitative changes during the unmanned path adjustment period. The congestion causal null space and the identifiable response direction are determined based on the orthogonal response relationship and the identifiable response boundary. When the congestion state exceeds the safe congestion state, a local congestion adjustment amount is determined in the identifiable response direction, so that the absolute value of the cumulative adjustment amount of each path data migration direction since the baseline diversion state when the current orthogonal response relationship is established does not exceed the intervention ratio that has been actually implemented and obtained response data in that direction, and data diversion is performed. The expected time-domain congestion change is determined based on the current diversion adjustment amount and the orthogonal response relationship. The response prediction deviation is obtained by comparing it with the actual time-domain congestion change after diversion. If the response prediction deviation is not higher than the identifiable response boundary, the current response relationship is continued to be used. If it is higher, the actual non-zero data migration direction between paths in this round is re-identified.
[0007] Furthermore, the conservation traffic diversion intervention is as follows: using a migration ratio that can be stably identified by the number of bytes sent from the source and target paths before and after migration and has not reached the safety queue boundary, a forward real service migration is implemented from the same baseline traffic diversion state and a response is collected to restore the baseline traffic diversion state; after the adjacent sampling changes at each congestion observation location enter the actual change range that occurred during the period without manual path adjustment, a reverse real service migration is implemented at the same ratio; the migration ratio is the ratio of the number of migrated service bytes to the total number of service bytes of all candidate paths in the same period.
[0008] Furthermore, the process of determining the orthogonal response relationship is as follows: the data migration directions between the linearly independent paths are used to form a migration matrix, which is then decomposed using a thin QR decomposition to obtain a column orthogonal matrix; the time when the data unit sent according to the new diversion relationship leaves the diversion device is used as the time reference, and the congestion state changes at multiple sampling times at each congestion observation location are obtained according to a unified sampling period. The changes in the diversion state relative to the same reference are normalized according to the migration ratio to form a unit migration response column vector; each column vector forms a response matrix, which is then converted into an orthogonal response matrix according to the coordinate relationship of the thin QR decomposition to characterize the orthogonal response relationship.
[0009] Furthermore, the process for determining the identifiable response boundary is as follows: during the period without manual path adjustment, select an adjacent background window of equal length to the collection period of the conservation diversion intervention response, and normalize the background congestion change according to the corresponding migration ratio to form a background response matrix; form a quantization error matrix based on the minimum measurement change of the congestion state collection device; transform the two matrices according to the same orthogonal coordinate relationship as the conservation diversion intervention, and determine the identifiable response boundary based on the spectral norm of the transformed weighted matrix.
[0010] Furthermore, a weighted singular value decomposition is performed on the orthogonal response matrix, and the directions with singular values no higher than the identifiable response boundary are mapped to the candidate path splitting space to form the congestion causal null space; the directions with singular values higher than the identifiable response boundary are determined as the identifiable response directions.
[0011] Furthermore, the unit data migration direction of the candidate path pair is projected onto the orthogonal split coordinates, and the unit split response gain is determined according to the orthogonal response matrix. For candidate path pairs whose gain is not higher than the identifiable response boundary, bidirectional real business migration is implemented according to the migration ratio determined in the conservation split intervention. Only when the actual unit migration response in both directions is not higher than the identifiable response boundary is the path pair determined to have a low response path relationship.
[0012] Furthermore, for candidate path pairs with the aforementioned low-response path relationship, an incremental migration ratio sequence is formed according to the actual executable data scheduling granularity of the diversion device. Each ratio implements bidirectional migration from the same baseline diversion state. When the unit migration response in either direction exceeds the identifiable response boundary for the first time, the migration ratio that most recently met the condition and the first time it did not meet the condition are used as the lower boundary and upper boundary, respectively. Verification continues between the two until there are no new executable migration ratios. The causal neutral migration upper limit is determined by the migration ratio that finally simultaneously meets the bidirectional low-response condition and the safety queue boundary.
[0013] Furthermore, when the causal neutral migration upper limit has been determined for any two candidate paths in the path set, the change directions of the paths within the group are orthogonalized and the corresponding weighted response matrix is determined; when the maximum singular value of the weighted response matrix is not higher than the identifiable response boundary, the path set is determined as a causal equivalent path group.
[0014] Furthermore, the observation location and sampling time where the baseline state before formal diversion is higher than the safe congestion state are determined as the target congestion location. The local congestion adjustment amount is calculated in the identifiable response direction to minimize the maximum residual excess of the target congestion location's expected state relative to the safe congestion state. At the same time, the adjusted candidate path allocation ratio is constrained, and the absolute value of the cumulative adjustment amount since the baseline diversion state when the current orthogonal response relationship is established after switching to the linear independent data migration direction does not exceed the intervention ratio that has been actually implemented and obtained response data in the corresponding direction.
[0015] Furthermore, after data diversion is completed, the actual time-domain congestion change is obtained according to the same sampling period and time benchmark as the conservation diversion intervention, and the expected time-domain congestion change is determined based on the current diversion adjustment amount and the orthogonal response relationship; the response prediction deviation corresponding to the unit actual diversion adjustment is determined based on the fitting deviation between the two and the current diversion adjustment amount; when the response prediction deviation is higher than the identifiable response boundary, the conservation diversion intervention is re-implemented only for the data migration direction between the linearly independent paths that are actually non-zero in the current round, and the corresponding response column and orthogonal response relationship are updated.
[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention implements a conservation-based diversion intervention on the data migration direction between linearly independent paths while maintaining a constant total allocation ratio of candidate paths. It collects state changes at different congestion observation locations and multiple sampling times, and combines these with background changes and quantified changes during periods of no manual path adjustment to determine identifiable response boundaries, thereby distinguishing between the congestion causal null space and the identifiable response direction. This approach helps improve the reliability of determining the response relationship between path migration direction and link congestion changes, providing a practically validated adjustment basis for subsequent data diversion. 2. When the link congestion exceeds the safe congestion state, this invention determines the local congestion adjustment amount only within the identifiable response direction, and limits the absolute value of the cumulative adjustment amount between each path's data migration direction from the baseline diversion state when the current orthogonal response relationship is established to no more than the intervention ratio that has been actually implemented and whose response data has been obtained in the corresponding direction. After diversion is completed, the deviation between the expected time-domain congestion change and the actual time-domain congestion change is used to determine whether the response relationship needs to be re-identified. This approach helps reduce the situation where diversion adjustments exceed the verified response range, enabling the data diversion process to be corrected based on actual link state changes. Attached Figure Description
[0017] Figure 1 A flowchart of a data offloading method based on link congestion awareness; Figure 2 This is a flowchart of step S4 of the present invention: determination of causal neutral migration boundary and confirmation of path group; Figure 3 This is a flowchart of step S5 of the present invention: solving for the shunt control quantity and correcting the response. Detailed Implementation
[0018] Example 1, refer to Figure 1 This embodiment presents a data splitting method based on link congestion awareness, applicable to networks where the same data to be transmitted can reach the destination node via multiple candidate paths. The network device performing data splitting can be a sending end, access switching equipment, edge gateway, or other network devices with multi-path scheduling capabilities. Candidate paths can share a common link or traverse different physical network segments or tunnel network segments. Specifically, it includes the following steps: Step S1: Characterization of candidate path status and determination of observation conditions.
[0019] First, obtain the actual amount of business data currently handled by the candidate path, and then uniformly determine the congestion observation location, sampling timing, and queue safety boundary for subsequent response identification. The same data collection criteria will be used for each subsequent path migration.
[0020] Step S11: Determine the candidate path splitting status: Suppose that for the same destination node or the same service destination address, there are The candidate paths are denoted as follows: .
[0021] The actual number of bytes sent for each candidate path is read within the continuous statistical window. When using hash bucket scheduling, the number of service bytes is accumulated for each hash bucket already allocated to each candidate path; when using flow scheduling or flowlet scheduling, the number of service bytes carried by each candidate path within the statistical window is accumulated.
[0022] The current traffic splitting status is represented as follows: in, Indicates time The corresponding candidate path routing status; Indicate candidate path The proportion of data it undertakes, its value is the amount of data within the current statistical window. The ratio of the number of service bytes sent to the total number of service bytes sent across all candidate paths; This represents the number of candidate paths.
[0023] The statistics window continues until all currently active scheduling units show changes in byte count. Round-robin scheduling can use a full scheduling round; when scheduling is based on streams, flowlets, or hash buckets, the statistics continue until all currently participating paths show changes in sent byte count.
[0024] Step S12, Link congestion status characterization: When the hop-by-hop status of candidate paths can be obtained, the output port of the candidate path concatenation is used as the initial observation position. For multiple ports that are passed consecutively without any path forks or convergences in between, the ports with smaller available queue capacity and the ports near the path convergence position can be retained.
[0025] When a candidate path passes through a tunnel network segment where the internal hop-by-hop state cannot be directly obtained, observation positions are set at the output queue corresponding to the tunnel entrance, the first observable output queue after the tunnel exit, and the destination-side receiving entrance. This is achieved through the above processing. One congestion observation point.
[0026] For the For each congestion observation point, read the actual number of bytes occupied by the output queue and the queue capacity available for the current service type, and obtain the corresponding congestion status according to the following formula: in, Indicates the first Each congestion observation point at time [time] The congestion state; Indicates the corresponding output queue at time [time]. The actual number of bytes occupied; This indicates the available byte capacity of the output queue allocated to the current business category; The values correspond to the congestion observation locations.
[0027] When only ECN tag information is available, the number of data bytes with ECN tags and the corresponding total number of data bytes can be counted in each sampling period, and the ratio of the two can be used as the congestion status at that observation location. Within the same control period, the same congestion status representation method is maintained for the same observation location. Congestion observation locations where only the ECN tag ratio is available, but not the actual queue occupancy and available queue capacity, are only used for time-domain congestion response identification and are not used independently for determining safe queue boundaries and safe congestion states; the ECN tag ratio and queue occupancy ratio are not directly compared numerically.
[0028] Step S13: Determining the observation time sequence and safety boundaries: Each congestion observation location is first sampled according to the telemetry period during which the equipment can continuously and stably output the queue status. When the original telemetry periods of different observation devices are inconsistent, the telemetry period with the longest duration is used as the unified sampling period, and the data from the remaining observation locations are summarized according to the same time boundary.
[0029] Perform a minimum schedulable data migration between candidate paths and record the moment when the first data unit sent according to the new splitting relationship leaves the splitting device. This moment serves as the unified time reference for this splitting event. Each observation location records the moment when the corresponding migrated data or the corresponding queue change first appears, and these moments determine the actual time range required for the splitting action to propagate to each observation location.
[0030] All sampling times in the subsequent time-domain congestion response are represented relative to the aforementioned unified time reference. The time when the response first appears at each observation location is only used to determine the response observation interval and is not used as their respective time zero point.
[0031] During normal operation without manual data migration, congestion status changes between adjacent sampling periods are continuously recorded. After a test migration is completed, sampling continues. When the adjacent sampling changes at all observation locations re-enter the range of changes that actually occurred during normal operation, and this continues for at least the time required for the aforementioned diversion action to propagate to the farthest observation location, the corresponding moment is taken as the end point of the response observation interval. This determines the response observation interval uniformly used in the current control process.
[0032] The safety queue boundary is determined based on the actual queue capacity and the amount of data arriving in a recent single sampling period. For the first... At each observation location, the number of data bytes entering the corresponding queue for multiple consecutive sampling cycles during the period without manual path adjustment is counted, and the largest single-cycle arrival byte count is obtained. .Will deduct The corresponding queue position is used as the safety queue boundary. If the device has been configured with an ECN trigger position or a lossless network flow control trigger position, and the corresponding position is lower than the above-mentioned safety queue boundary, the lower queue position is used.
[0033] When the maximum single-cycle arrival data volume is not less than the corresponding available queue capacity, it is determined that the current congestion observation position does not have the safe probing conditions to actively increase the data ratio, and zero or negative values are not used as the available safe queue boundary; before obtaining normal operation statistics results that satisfy the safe queue boundary being greater than zero, the data ratio is not increased to the target path passing through the observation position.
[0034] Step S2: Conservative diversion intervention and time-domain response identification.
[0035] Controllable inputs are formed by pairwise path migration using actual business data. First, a unified orthogonal coordinate system is established in the path allocation space for each migration direction, and then the link congestion changes at multiple locations and sampling times are recorded using the same time base.
[0036] Step S21: Determine the conserved migration coordinates: Since the sum of the data allocation ratios for all candidate paths remains constant, in Select from candidate paths Data migration directions between linearly independent paths.
[0037] For candidate paths To candidate paths The migration data, along with the corresponding migration direction and coordinate relationships, is represented as follows: in, and They represent the first The element and the first A unit vector in which one element is 1 and the rest are zero; Indicates the first Data migration directions between paths; ; This represents a matrix consisting of all independent migration directions; This indicates the proportion of data migration corresponding to each migration direction; This indicates the change in the proportion of traffic splitting along the candidate path.
[0038] For matrix Thin QR decomposition is performed to obtain column orthogonal matrices. and invertible upper triangular matrix . Each column together forms a diversion adjustment space that satisfies the condition that the total ratio of candidate path allocation remains unchanged; Indicates and An identity matrix of the same order as the number of columns.
[0039] Actual determination When this happens, a candidate path currently carrying the service can be selected as a reference path, and then migration directions between this reference path and the other candidate paths can be established. After QR decomposition, subsequent low response space judgments are all performed within... The process is performed within a defined orthogonal coordinate system; therefore, the selection of the reference path is not used as the basis for subsequent response space partitioning.
[0040] Step S22: Determining and implementing the conservation diversion intervention amount: First, determine the smallest scheduling unit that the distribution device can independently change the path. For hash bucket scheduling, a reassignable hash bucket corresponds to one smallest scheduling unit; for flowlet scheduling, a complete flowlet is the smallest scheduling unit; for data block scheduling, a data block that cannot be further divided is the smallest scheduling unit.
[0041] The actual path migration begins from a minimum scheduling unit, and the byte counts of the source and target paths before and after the migration are compared. If the change in data volume caused by this migration cannot be stably identified from the difference in counts between adjacent statistical windows, another minimum scheduling unit is added. For congestion observation positions determined in step S13 as not currently meeting the conditions for actively increasing the data proportion for safe probing, no conservation diversion intervention to increase the data proportion passing through that observation position is performed. Before each addition of migrated data, each congestion observation position involved in the target path is checked; if the remaining byte space from the current queue to the safe queue boundary of the corresponding observation position is less than the data byte volume of the scheduling unit to be added, the addition is stopped. The amount of data migrated in this instance must not exceed the amount of data that the source path can actually provide within the equal-length response observation interval.
[0042] The minimum actual migration ratio that meets the above counting identifiable conditions and has not reached the safety boundary is denoted as . Its value is the ratio of the actual number of bytes of business data migrated in this transaction to the total number of bytes of business data from all candidate paths within the same time period.
[0043] Regarding the direction of migration First, the baseline traffic splitting state before the start of this round of testing is recorded. Then, a forward migration from the source path to the target path is implemented starting from this baseline state. After the forward response acquisition is completed, the system returns to the baseline traffic splitting state and continues to acquire the state of each congestion observation location. When the adjacent sampling changes of all observation locations re-enter the normal operating change range obtained in step S13, a reverse migration with the same data ratio is implemented starting from the same baseline traffic splitting state.
[0044] The response observation interval length, sampling period, and time reference determination method remain consistent for both forward and reverse migration. This paired data migration process is a conserved diversion intervention.
[0045] Step S23: Determining the relationship between time-domain congestion response and orthogonal response: Within the response observation intervals of both forward and reverse migration, data are obtained according to a uniform sampling period. Each sampling time is recorded sequentially as follows: .each Both represent the time offset relative to the execution time of the corresponding conservation diversion intervention.
[0046] In the Each sampling time will include all The congestion observation locations are arranged in a fixed order to form a state vector. Forward transition corresponds to... Reverse migration corresponds to .
[0047] The response relationship for each independent migration direction is represented as follows: in, Indicates the first The response column vector corresponding to each independent migration direction at a unit migration ratio; This represents the actual migration proportion used in this migration direction; This indicates that the state difference vectors are arranged vertically according to the order of sampling time; Based on the original migration base The response matrix; To transform to a column orthogonal matrix The response matrix corresponding to the coordinates, referred to as the orthogonal response matrix in this embodiment, refers to the time-domain congestion response matrix represented under the orthogonal columns of the split input coordinates. The orthogonality corresponds to the split input coordinates and does not require that the response columns of the response matrix be mutually orthogonal. Indicates the coordinates of orthogonal splitting; This indicates the expected congestion changes at multiple locations and sampling times resulting from traffic diversion changes along a given path.
[0048] Both forward and reverse migrations start from the same baseline traffic splitting state, allowing for comparison of the two link responses around the same network operating state. The aforementioned changes resulting from a single path migration across multiple observation locations and multiple subsequent sampling times are called the time-domain congestion response.
[0049] Step S3: Low-response control space identification and path relationship determination.
[0050] The natural fluctuations under normal operating conditions are treated with the same scale as the conservation diversion intervention. Based on this, the response boundary that can be stably distinguished from background changes is determined, and the low response direction and corresponding candidate path relationship are identified in the orthogonal diversion coordinate.
[0051] Step S31: Determine the identifiable response boundary: Regarding the first For each independent migration direction, two adjacent background windows of equal length to the response observation interval are selected during normal operation without manual path adjustments. Both background windows use the same sampling times as the formal identification and are based on the actual sampling method used in that direction. Normalize.
[0052] The background response and identifiable response boundaries are determined by selecting several groups of background windows consecutively, according to the following formula: in, Indicates the background window group number; and These represent the two corresponding unattended path adjustment windows in the [number]th [year]. The congestion state vector at each sampling time; Indicates the first The background response matrix formed by the group background window under the original migration basis; This represents the background response matrix after transformation to orthogonal split coordinates; This represents a matrix that applies consistent weights to all congestion observation locations and all sampling times; for An identity matrix of order 1; Represents the spectral norm of a matrix; This represents the unit shunt response boundary that can be stably distinguished from natural background variations and measurement quantization errors under current network conditions.
[0053] This is used to represent the minimum metering change of the congestion state acquisition device itself. When using queue byte counting, for the 1st... The minimum resolvable byte change of the queue counter is obtained at each observation position, and this byte change is divided by... This yields the minimum resolvable congestion state change at the observed location. All Form vectors according to the order of observation positions. .
[0054] Regarding the first One migration direction, according to Each sampling time is repeated and divided by . The first quantization error matrix is formed. Columns. All columns are arranged to form... Then according to Transform to orthogonal split coordinates. When using ECN counting, divide the amount of data corresponding to one ECN count change by the total amount of data in the same sampling period to obtain the corresponding result. .
[0055] Therefore, background changes, measurement quantification changes, and formal conservation diversion interventions all undergo the same migration ratio normalization and orthogonal coordinate transformation.
[0056] Step S32, Determining the congestion causality null space: Weighted singular value decomposition is performed on the orthogonal response matrix, and the response gain is not greater than... The direction is mapped back to the actual candidate path splitting space: in, , and The matrix obtained from singular value decomposition; express The first in One singular value; This represents the corresponding right singular vector; This indicates the direction of change of the corresponding direction in the actual candidate path splitting ratio space; This represents the congestion causal zero space.
[0057] satisfy The direction corresponds to a shunting response that cannot be stably distinguished from changes in the natural background near the current network operating state. The congestion causality null space referred to in this embodiment is an engineering near-zero response space formed using a identifiable response boundary as the criterion. It indicates that the time-domain congestion response corresponding to the shunting direction cannot be stably distinguished from background changes and measurement quantization changes within the current operating state, congestion observation location, and response observation interval. It does not require the corresponding mathematical response to be strictly equal to zero. (The last sentence appears to be incomplete and possibly refers to a requirement for a specific network configuration.) The direction is retained as a diversion direction that can form a discernible congestion change.
[0058] There is no satisfying When the direction is singular, the congestion causal null space of the current control cycle is empty, and subsequent data reallocation based on low-response path relationships is not performed. There is no condition that satisfies... When the direction is singular, the current candidate path splitting degree of freedom is not used to handle link congestion.
[0059] Step S33: Determining the low response relationship and path group of candidate paths: For candidate paths and candidate paths The proportion of unit data from Migrate to The corresponding actual path change direction is projected onto the orthogonal split coordinate system, and the overall response generated by a unit path change is calculated: in, Indicate candidate path To candidate paths The representation of the unit data migration direction in orthogonal split coordinates; Represents the response vector The overall response at all observation locations and at all sampling times; This indicates the gain of the corresponding unit shunt response for that path.
[0060] when Not higher than At that time, directly target according to step S22. and Implement bidirectional real-world business migration with the minimum identifiable proportion. Each direction starts from the same baseline traffic splitting state, and the actual unit migration response in both directions does not exceed [a certain threshold]. When this happens, the path pair is retained as a low-response path relationship.
[0061] When forming a group of paths to be verified, first select a candidate path that continuously carries the service, and then add other candidate paths one by one. The path to be added and every existing path in the group must meet the above-mentioned true bidirectional migration conditions.
[0062] When there are more than two paths within a group, one path is selected as the reference path within the group. A path change vector is generated between this reference path and other paths within the group, and this vector is mapped to orthogonal split coordinates. These coordinate vectors are then subjected to Gram-Schmidt orthogonalization to form orthogonal directions within the group. The maximum singular value of the weighted response matrix corresponding to these orthogonal directions is then calculated. Only when this maximum singular value is not higher than... Only then does the path set enter the next step of actual migration range verification. This ensures that any small-amplitude redistribution of conserved data within the path group is subject to low-response conditions.
[0063] Step S4: Determine the causal neutral migration boundary and confirm the path group.
[0064] like Figure 2 As shown, the low-response relationship that holds true under small data migration conditions may become invalid as the migration amount increases. For each candidate path pair, verification is performed step by step with discrete executable data scheduling, and each verification restores the same baseline split state.
[0065] Step S41: Determining the migration verification sequence and security scope: For candidate paths and Each migration direction is checked to see the amount of data that can be provided within the current response observation interval.
[0066] from relocation At that time, read The system records the current queue occupancy, safe queue boundary, and the actual number of arriving and departing bytes in the most recent unmanaged path adjustment window for each observation location. For each observation location, it calculates the remaining byte space between the safe queue boundary and the current queue occupancy. If the number of arriving bytes in the most recent window is greater than the number of departing bytes, the difference is subtracted from the remaining byte space.
[0067] The minimum remaining byte space after processing all observation positions along the target path is taken and compared with the actual number of data bytes carried by the source path within the equal-length window. The smaller value is taken as the upper bound of the allowed data volume for testing in that direction. The same method is used for the reverse direction. Bidirectional relationship verification uses the smaller of the upper bounds of the data volume in both directions.
[0068] Starting from the smallest schedulable unit, an incremental migration ratio sequence is formed according to the actual data scheduling granularity that the device can execute. Each migration ratio in the sequence must not exceed the diversion ratio corresponding to the upper bound of the aforementioned two-way verification data volume.
[0069] Each time a new one is verified Previously, the process was restored to the baseline diversion state recorded when the path pair began verification, and waited for adjacent sampling changes at all congestion observation locations to re-enter the normal operating change range obtained in step S13. After restoration, a new data migration was implemented from this baseline diversion state.
[0070] Step S42: Determining the upper limit of causal neutral migration: For each actual executable proportion in the migration sequence, data migration is performed in both directions, using the baseline state before the start of this test as the response reference. The upper limit of bidirectional migration between candidate paths is determined by the following formula: in, Indicates the proportion of data migration actually performed in a single operation; Indicated by proportion Data from Migrate to Then, the time-domain congestion change vector obtained relative to the baseline state before the test; This indicates that the same proportion is used from Migrate to The obtained temporal congestion change vector; and These represent the response gain per unit path change in both directions; Indicate candidate path With candidate paths The upper limit of causal neutral migration between them.
[0071] If a certain migration ratio first causes the unit response gain in any direction to exceed The lower boundary is the proportion of executable migrations that most recently met the conditions, and the upper boundary is the proportion of migrations that first failed to meet the conditions. The data volume that the splitter can actually implement and is closest to the midpoint between the two boundaries is then used for further verification. Based on the verification results, the boundaries are narrowed until there are no new executable minimum scheduling units between the upper and lower boundaries.
[0072] During the verification process, when any target path observation position reaches the safety queue boundary determined in step S13, the migration amount is stopped from being expanded further, and the actual migration ratio that was previously satisfied with both the low response condition and the safety queue condition is adopted.
[0073] Any two paths in the path group to be verified were valid. After the low response space check within the group in step S33 is passed, the path set is determined as a causal equivalent path group.
[0074] When determining causal equivalence path groups, the path allocation status within the group is simultaneously recorded as a reference status within the group. For each specific path pair, a directed data migration record is maintained, with an initial value of zero. Towards Increase the number of records during data migration and decrease the number of records during reverse migration. Its absolute value must not exceed [a certain limit]. Therefore, reverse migration can restore the migration range that has already been occupied, and intra-group path adjustments are always limited to the vicinity of the offloading location that has been verified in practice.
[0075] Step S5: Solving for the shunt control quantity and correcting the response.
[0076] like Figure 3 As shown, current link congestion is handled based on the identified orthogonal response relationships. Formal traffic offloading adopts a small-scale iterative approach, with each adjustment limited to the actually identified local range. After the adjustment is completed, a decision is made based on the actual time-domain response to determine whether to continue using the current response relationship.
[0077] Step S51: Determine the congestion constraint state: Divide the safe queue boundary for each congestion observation location where the actual queue occupancy and available queue capacity can be obtained by the corresponding... This yields the safe congestion state at the observed location. Following the index order of the congestion observation locations where actual queue occupancy and available queue capacity can be obtained, and their corresponding response sampling times within the complete time-domain congestion response vector, the corresponding safe congestion states are repeatedly arranged to form a safe state vector. The safe congestion state vector is only compared with the expected congestion states with the same index in the full time-domain congestion response vector; observation locations with only ECN marker proportions are not assigned a safe congestion state, and their corresponding response data are still retained in the full time-domain congestion response vector.
[0078] Before the formal diversion, a window of equal length to the response observation interval without manual adjustment is obtained, and a reference state vector is formed according to the same sampling time. .
[0079] The indices of vector elements whose baseline state is higher than the safe state are recorded in the set. The indices of the remaining elements are recorded in the set. The elements of the set correspond to specific congestion observation locations and specific response sampling times. Congestion observation locations with only ECN marker proportions are not included in the above-mentioned safe state constraints and target congestion location sets, but their sampled data are still retained in the time-domain congestion response vector, participating in the calculation of orthogonal response relationships and identifiable response boundaries. When there are no safe constraint observation locations where the actual queue occupancy and available queue capacity can be obtained for the target path to increase the data proportion, no active conservation diversion intervention is implemented in the migration direction pointing to that target path.
[0080] Singular values greater than The corresponding right singular vectors are sequentially arranged into a matrix. singular values not higher than The corresponding right singular vectors form a matrix It does not exist. In the corresponding direction, data congestion regulation based on candidate paths is not performed.
[0081] Step S52, Solving for local congestion adjustment: exist The actual flow rate is obtained from the corresponding identifiable response direction. Let... Each of them All of these represent the proportion of interventions that have been actually implemented and for which response data has been obtained in the corresponding original migration direction in step S22.
[0082] This round of local adjustments is solved using the following constraints: in, This represents the cumulative migration proportion implemented along each original conserved migration direction in step S22 since the baseline diversion state when the current orthogonal response relationship was established; when establishing a new orthogonal response relationship, the cumulative migration proportion is initialized to zero. After obtaining the adjustment amount in each round of formal diversion, the adjustment amount of this round is converted to the original conserved migration direction and added to the corresponding cumulative migration proportion. The result of the addition is used to determine whether it exceeds the range of intervention proportions that have been actually verified. This represents the control coefficient corresponding to the identifiable response direction; This indicates the expected congestion status at each observation location and at each response sampling time after the execution of this round of partial diversion; This indicates that the target congestion location may still be higher than the maximum remaining amount in a safe state; , These represent the expected state and the safe state at the corresponding index position, respectively. This indicates the allocation ratio of candidate paths after this round of traffic splitting; This indicates the migration ratio after the formal adjustment in this round is converted back to the original conservation migration directions actually executed in step S22.
[0083] The final constraint ensures that the cumulative migration ratio of each original migration direction from the baseline split state when the current orthogonal response relationship is established does not exceed the range of the link response that has been actually identified, thus avoiding the gradual exceedance of the verified local response range in multiple rounds of split adjustments.
[0084] First find the smallest There exist multiple sets of control variables that achieve the same minimum. When, fix the minimum Then select from these feasible control quantities The smaller group results in a smaller actual path adjustment in this round.
[0085] This constraint problem can be solved using the activity set method or the interior point method. When using the activity set method, the path proportion boundary, local migration boundary, and safe congestion boundary are treated as linear inequality constraints. Starting from the zero adjustment point corresponding to the current split state, these constraints are released sequentially to reduce... Feasible migration directions; when further movement in any available direction cannot continue to reduce [the risk / risk]... If continuing to move would violate at least one boundary constraint, the current solution is terminated.
[0086] If this round is the smallest If the value is still greater than zero, but has fallen below the maximum safe state excess of the target observation position before adjustment, then the obtained local flow is executed, and after the response verification is completed, the next round of solution is performed with a new working state.
[0087] If the current allowed local migration range cannot produce a reduction in congestion at any target location exceeding the level currently identifiable by measurement, then the unverified path migration amount will not be expanded further. The remaining congestion can be handled by existing transmission rate control or service admission measures.
[0088] Step S53: Data reallocation within the low response range: Once the target congestion location has returned to a safe state, if there are other service scheduling needs, the expected diversion change will be determined by the existing service scheduling rules. The sum of the elements of this change remains zero.
[0089] First calculate its orthogonal migration coordinates Then retain the coordinates in The projection in the space is obtained The corresponding actual candidate path adjustment amount is .
[0090] Based on this adjustment, determine the paths that need to reduce the proportion of data and the paths that need to increase the proportion of data, and perform actual path pairing only within the same causal equivalent path group.
[0091] For each pair of paths within the group, read the directional migration amount recorded in step S42. If the current directional migration record is... Then from Towards When continuing the migration, the number of new records must not exceed The corresponding decrease occurs when migrating in the opposite direction. The actual migration amount for each migration is limited by both the amount of data to be reduced in the outgoing path and the amount of data to be added in the incoming path.
[0092] When the proposed intra-group migration is opposite to the current directional migration record volume, and this migration will cause the record volume to cross zero, the migration is split into two actions: first, the first action is performed to restore the directional migration record volume of the path pair to zero, and the corresponding actual time-domain congestion response is obtained according to step S54; after confirming that the actual unit migration response is not higher than the identifiable response boundary, the second action is performed in the opposite direction from the zero-record state of the path pair. No single migration action may exceed the migration ratio that has been actually verified for the path pair.
[0093] Update the corresponding path after adjusting the path pair. Then, the next path pair is processed. If there is no receiving path that satisfies the causal equivalence path group relationship and still has an allowable migration range, the corresponding remaining scheduling requirements are stopped within the current control cycle.
[0094] Step S54, Formal shunt response correction: After completing the formal traffic diversion in step S52, the actual time-domain congestion change vector is obtained using the formal traffic diversion execution time as a unified time reference, following the same response observation interval and sampling period as the conservation traffic diversion intervention. This vector represents the actual change in congestion state after formal traffic diversion relative to the baseline state in step S51.
[0095] The fitting bias of the current response relationship is expressed as: in, This indicates the deviation in response prediction corresponding to the actual diversion adjustment per unit; This indicates the actual time-domain congestion changes resulting from this round of formal traffic diversion; This represents the expected time-domain congestion changes resulting from the current orthogonal response relationship; The method for calculating the comprehensive response quantity in step S33 is used.
[0096] Only The deviation is calculated when it is greater than zero. This check is unnecessary if no actual path adjustment has occurred.
[0097] when Not higher than At this time, the formal traffic splitting in this round remains consistent with the current response relationship. If the target congestion has not yet returned to a safe state, the baseline state is re-acquired based on the candidate path allocation state adjusted in this round, and the next local traffic splitting solution is executed. The re-acquired baseline state is used for the response change calculation in this round and does not change the baseline traffic splitting state when the current orthogonal response relationship was established; before re-establishing the orthogonal response relationship, the migration proportion of each original migration direction continues to be accumulated according to the original baseline traffic splitting state.
[0098] when Higher than When the actual non-zero original migration direction is read, the conservation diversion intervention is re-executed only for the candidate paths involved in the migration direction, and the corresponding response column, orthogonal response relationship, and causal neutral migration upper limit of the relevant path are updated. Candidate paths that did not participate in the actual migration in this round and whose background congestion change has not exceeded the original identifiable response boundary continue to retain the existing identification results.
[0099] The intra-group path adjustment in step S53 uses the same response check method. The number of directional migration records for a certain path has not yet reached the required threshold. The actual unit migration response has exceeded If the condition is not met, immediately stop the path pair from continuing to migrate in the current direction, and re-execute the boundary verification in step S42 with the path allocation position that most recently met the conditions and the path allocation position that does not meet the conditions this time.
[0100] After the response matrix completes its initial identification, it is used as a continuous control cycle until the aforementioned local re-identification conditions occur or the candidate link capacity changes. Upon entering a new control cycle, the candidate paths involved in the changes re-acquire their response relationships and corresponding migration ranges.
[0101] Example 2 illustrates the technical chain of the data offloading method based on link congestion awareness of the present invention using a four-path example, as follows: Assume there are candidate paths , , and The current traffic splitting status is: in: choose As a reference path for establishing the initial conserved transfer base, we obtain: And composed of: right Implement QR decomposition: Subsequently, forward and reverse conservation diversion interventions were implemented for the three original migration directions to obtain the original response matrices. Then, after conversion, we get: Assume that the identifiable response boundary is obtained based on the absence of manual adjustment of the background window and the actual metering resolution of the equipment. .
[0102] After weighted singular value decomposition, there exists a low-response direction, which, when mapped to the actual candidate path splitting space, mainly corresponds to... and Data exchange between them, and with and Both of the main adjustment directions can produce more than The identifiable response.
[0103] Further calculate candidate paths To candidate paths Orthogonal coordinates: If the following conditions are met: Then directly target and Implement a small-scale, two-way migration of real business data. The unit migration response time in both directions should not exceed [a certain value]. At that time, proceed to the migration range verification.
[0104] The actual migration ratio is increased based on the minimum scheduling unit of the diversion equipment: Assuming from to All corresponding bidirectional actual migrations satisfy the low response condition, while using hour, At a certain congestion observation location, more than The unit migration response.
[0105] Further in and After verification based on the actual executable scheduling granularity of the equipment, the following results were obtained: Thus confirming and It has an effective low-response migration range in the current operating state.
[0106] Subsequently, a congestion observation position for a certain target is established. The queue state exceeds the safe queue boundary. The current response matrix indicates... and Data exchange between them mainly occurs in the congestion causality null space, while with The relevant migration direction can enable The congestion status decreased at multiple response sampling times.
[0107] The local flow adjustment amount is obtained according to step S52. Let the obtained orthogonal control coordinates be... The corresponding candidate path splitting changes are as follows: And the adjustment amounts in each original migration direction are still within the range that the conservation diversion intervention has actually identified.
[0108] After this round of traffic splitting, the new candidate path allocation status is recorded as follows: Subsequently, actual link changes were collected according to the unified response observation interval. If the unit response prediction deviation generated by the formal traffic diversion is no higher than... Continue using the current response relationship; if the target observation position has not yet returned to a safe state, then... The new current split state will be used to start the next round of local solution.
[0109] Once the target congestion has been reduced to a safe level, if existing service scheduling rules require further reduction... The data to be taken on and increased The data to be received will then be projected onto the congestion causality null space according to the corresponding expectations.
[0110] Assuming this round of implementation is permitted to Intragroup migration rate The number of directional records before migration was If the migration still satisfies the following conditions: After completing the group adjustments, the time-domain congestion response was collected again. The actual unit migration response was still no higher than [previous data]. At that time, retain the current time. and The low response relationship; the actual unit migration response is higher than When the current path assignment position is compared with the most recent path assignment position that satisfies the condition, the upper limit of causal neutral migration is re-verified.
[0111] Through the detailed description of the above embodiments, the data diversion method based on link congestion awareness of the present invention is based on the actual diversion status of candidate paths, link congestion observation results, and real service migration responses. First, it uniformly determines the congestion observation location, sampling period, time base, and safe congestion state. Then, it obtains time-domain congestion changes at multiple locations and times through conservation diversion intervention between linear independent paths and establishes orthogonal response relationships. After forming identifiable response boundaries by combining background changes and measurement quantification changes during normal operation, it distinguishes between low-response directions and identifiable response directions in the diversion space, and further limits the low-response path relationships and their verifiable migration range through bidirectional actual migration. When the link state exceeds the safe congestion state, local congestion adjustment is calculated only within the identified response directions and actual verification range, and data diversion is performed. After diversion, the current response relationship is verified using actual time-domain congestion changes. If necessary, the migration directions actually involved in the adjustment in this round are re-identified, thus forming a data diversion process that connects response identification, local diversion, and actual feedback correction.
[0112] The preset parameters in the above formulas shall be set by those skilled in the art according to the actual situation.
[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.
[0114] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0115] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0116] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0117] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0118] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A data offloading method based on link congestion awareness, characterized in that, The method flow is as follows: Obtain the candidate path diversion status of the same data to be transmitted, determine the congestion observation location, unified sampling period and time base; for the congestion observation location where the actual queue occupancy and available queue capacity can be obtained, determine the safe queue boundary by the difference between the queue capacity and the maximum single-cycle arrival data volume during the period without manual path adjustment, and determine the safe congestion status by the ratio of the safe queue boundary to the available queue capacity. Under the condition that the total allocation ratio of candidate paths remains unchanged, a conservation diversion intervention is implemented on the data migration direction between linearly independent paths, and the orthogonal response relationship is determined based on the congestion status changes of each congestion observation location at multiple sampling times. The identifiable response boundary is determined based on the background changes and measured quantitative changes during the unmanned path adjustment period. The congestion causal null space and the identifiable response direction are determined based on the orthogonal response relationship and the identifiable response boundary. When the congestion state exceeds the safe congestion state, a local congestion adjustment amount is determined in the identifiable response direction, so that the absolute value of the cumulative adjustment amount of each path data migration direction since the baseline diversion state when the current orthogonal response relationship is established does not exceed the intervention ratio that has been actually implemented and obtained response data in that direction, and data diversion is performed. The expected time-domain congestion change is determined based on the current diversion adjustment amount and the orthogonal response relationship. The response prediction deviation is obtained by comparing it with the actual time-domain congestion change after diversion. If the response prediction deviation is not higher than the identifiable response boundary, the current response relationship is continued to be used. If it is higher, the actual non-zero data migration direction between paths in this round is re-identified.
2. The data offloading method based on link congestion awareness according to claim 1, characterized in that, The conservation traffic splitting intervention is as follows: using the migration ratio that can be stably identified by the number of bytes sent on the source and target paths before and after migration and has not reached the safety queue boundary, a positive real service migration is implemented from the same baseline traffic splitting state and the response is collected to restore to the baseline traffic splitting state; Once the adjacent sampling changes at each congestion observation location enter the range of changes that actually occurred during the period of no manual path adjustment, reverse real service migration is implemented in the same proportion; the migration proportion is the ratio of the number of migrated service bytes to the total number of service bytes of all candidate paths in the same period.
3. The data offloading method based on link congestion awareness according to claim 2, characterized in that, The process of determining the orthogonal response relationship is as follows: the data migration directions between the linear independent paths are used to form a migration matrix, and the column orthogonal matrix is obtained by thin QR decomposition; the time when the data unit sent according to the new diversion relationship leaves the diversion device is used as the time reference, and the congestion state changes of forward and reverse migration at multiple sampling times at each congestion observation position are obtained according to a unified sampling period; the changes of diversion state relative to the same reference are normalized according to the migration ratio to form a unit migration response column vector. The column vectors form a response matrix, which is then converted into an orthogonal response matrix according to the coordinate relationship of the thin QR decomposition, thereby representing the orthogonal response relationship.
4. The data offloading method based on link congestion awareness according to claim 3, characterized in that, The process of determining the identifiable response boundary is as follows: during the period without manual path adjustment, select an adjacent background window with the same length as the collection period of the conservation diversion intervention response, and normalize the background congestion change according to the corresponding migration ratio to form a background response matrix. A quantization error matrix is formed based on the minimum measurement change of the congestion state acquisition device; the two matrices are transformed according to the same orthogonal coordinate relationship as the conservation diversion intervention, and the identifiable response boundary is determined based on the spectral norm of the transformed weighted matrix.
5. A data offloading method based on link congestion awareness according to claim 4, characterized in that, Weighted singular value decomposition is performed on the orthogonal response matrix, and directions with singular values no higher than the identifiable response boundary are mapped to the candidate path splitting space to form the congestion causal null space; directions with singular values higher than the identifiable response boundary are determined as the identifiable response directions.
6. The data offloading method based on link congestion awareness according to claim 5, characterized in that, Project the unit data migration direction of the candidate path pair onto the orthogonal split coordinates, and determine the unit split response gain according to the orthogonal response matrix; for candidate path pairs whose gain is not higher than the identifiable response boundary, implement bidirectional real business migration according to the migration ratio determined in the conservation split intervention, and determine that the path pair has a low response path relationship only when the actual unit migration response in both directions is not higher than the identifiable response boundary.
7. A data offloading method based on link congestion awareness according to claim 6, characterized in that, For candidate path pairs with the aforementioned low-response path relationship, an incremental migration ratio sequence is formed according to the actual data scheduling granularity that the diversion device can execute, and each ratio implements bidirectional migration from the same baseline diversion state. When a unit migration response in either direction exceeds the identifiable response boundary for the first time, the migration proportions that most recently met the conditions and those that first failed to meet the conditions are used as the lower and upper boundaries, respectively. Verification continues between the two until no new executable migration proportions exist. The upper limit of causal neutral migration is determined by the migration proportion that last simultaneously meets both the bidirectional low response condition and the safe queue boundary.
8. A data offloading method based on link congestion awareness according to claim 7, characterized in that, When the upper limit of the causal neutral migration has been determined for any two candidate paths in the path set, the change direction of the paths within the group is orthogonalized and the corresponding weighted response matrix is determined; when the maximum singular value of the weighted response matrix is not higher than the identifiable response boundary, the path set is determined as a causal equivalent path group.
9. A data offloading method based on link congestion awareness according to claim 5, characterized in that, The observation location and sampling time where the baseline state before formal diversion is higher than the safe congestion state are determined as the target congestion location. The local congestion adjustment amount is calculated in the identifiable response direction to minimize the maximum residual excess of the expected state of the target congestion location relative to the safe congestion state. At the same time, the adjusted candidate path allocation ratio is constrained, and the absolute value of the cumulative adjustment amount since the baseline diversion state when the current orthogonal response relationship is established does not exceed the intervention ratio that has been actually implemented and obtained response data in the corresponding direction after the data migration direction between the linear independent paths is switched.
10. A data offloading method based on link congestion awareness according to claim 9, characterized in that, After data diversion is completed, the actual time-domain congestion change is obtained according to the same sampling period and time benchmark as the conservation diversion intervention, and the expected time-domain congestion change is determined based on the current diversion adjustment amount and orthogonal response relationship. The response prediction deviation corresponding to the unit actual diversion adjustment is determined based on the fitting deviation between the two and the current diversion adjustment amount. When the response prediction deviation is higher than the identifiable response boundary, the conservation diversion intervention is re-implemented only for the data migration direction between the linear independent paths that are actually non-zero in the current round, and the corresponding response column and orthogonal response relationship are updated.
