Conflict graph driven incremental joint optimization method, electronic device, medium and product

Abstract

This application relates to the field of computer network technology and discloses a conflict graph-driven incremental joint optimization method, electronic device, medium, and product. The method includes acquiring communication semantic information and a physical network topology graph from a semantic source and / or a training control plane; establishing a mapping relationship between logical communication edges in the communication tree and physical network links on the physical network topology to determine the physical path occupancy set corresponding to the active communication tree; constructing a communication tree conflict graph based on the physical path occupancy set of each active communication tree; filtering out high-conflict communication trees from the communication tree conflict graph; identifying a high-conflict subset containing hotspot logical edges from the high-conflict communication trees after triggering a rebalancing mechanism; and performing incremental rebalancing on this high-conflict subset. This invention only performs local incremental rebalancing on the high-conflict subset, reducing online operational risks and effectively alleviating tail latency problems during multi-task or multi-communication flow concurrency.

Description

Technical Field

[0001] This application relates to the field of computer network technology, and in particular to conflict graph-driven incremental joint optimization methods, electronic devices, media, and products. Background Technology

[0002] As the scale of large-scale model training continues to expand, various parallel modes such as data parallelism, tensor parallelism, pipeline parallelism, and expert parallelism coexist. The ensemble communication traffic during training has become a core bottleneck limiting training throughput and scalability. In real-world clusters, multiple training tasks run in parallel, and each training task typically contains multiple communication groups. Each communication group logically corresponds to one or more ensemble communication trees, and these logical communication trees share the same underlying physical network.

[0003] Existing technologies typically optimize the following two aspects respectively:

[0004] 1. Optimize the collection communication algorithm on the training framework or communication library side, for example, select a better mode for Ring, Tree, and HierarchicalAllReduce.

[0005] 2. Optimize physical paths on the network controller side, for example, by performing ECMP, WCMP, shortest path allocation, or load balancing.

[0006] Both of the above optimizations generally suffer from the problem of separation between the logical and physical layers. The training framework is unaware of external conflicts on shared links, and the network controller is also unaware of which communication group, communication tree, or set of communication primitives a given flow belongs to. During multi-task concurrency, multiple communication trees may occupy the same batch of Spine-Leaf links or links between the same group, causing local congestion to be amplified into global synchronous blocking.

[0007] To mitigate network conflicts between communication trees, existing flow control and rebalancing schemes typically employ global rerouting or static hash allocation strategies. However, when faced with distributed AI training scenarios involving active communication trees, large message volumes, and complex spatiotemporal overlap, existing rebalancing mechanisms reveal the following serious drawbacks:

[0008] First, existing rebalancing schemes lack effective means to quantify network occupancy conflicts between communication trees in the spatiotemporal dimensions.

[0009] Secondly, existing rebalancing mechanisms typically tend to perform global, large-scale path recalculation and adjustments. In a distributed training cluster with multiple concurrent tasks, global rerouting can have far-reaching consequences. This indiscriminate adjustment forces many previously functioning, conflict-free communication trees to undergo path mapping changes. Frequent and disordered switching of network paths can easily trigger large-scale asynchronous control and data plane operations, online network oscillations, and brief periods of network congestion. This can lead to temporary pauses in the distributed training framework due to connection reconnection and increased latency, and in severe cases, even cause task performance degradation or training deadlock / crash.

[0010] Therefore, how to provide conflict graph-driven incremental joint optimization methods, electronic devices, media, and products to address the obvious deficiencies in existing technologies, such as the lack of effective means to quantify network occupancy conflicts between communication trees in the spatiotemporal dimension and the lack of indiscriminate adjustment, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0011] One objective of this application is to provide a conflict graph-driven incremental joint optimization method, electronic device, medium, and product, at least to address the shortcomings of existing technologies in terms of the lack of effective means to quantify network occupancy conflicts between communication trees in the spatiotemporal dimension and the lack of indiscriminate adjustment.

[0012] To achieve the above objectives, some embodiments of this application provide the following aspects:

[0013] In a first aspect, some embodiments of this application provide a conflict graph-driven incremental joint optimization method, characterized by: acquiring communication semantic information and physical network topology graphs from training-side semantic sources and / or training control planes; establishing a mapping relationship between logical communication edges in the communication tree and physical network links on the physical network topology to determine the physical path occupancy set corresponding to the active communication tree; and constructing a communication tree conflict graph based on the physical path occupancy set of each active communication tree.

[0014] High-conflict communication trees are selected from the communication tree conflict graph. After the rebalancing mechanism is triggered, a high-conflict subset containing hotspot logical edges is identified from the high-conflict communication trees, and incremental rebalancing is performed on the high-conflict subset.

[0015] Secondly, some embodiments of this application also provide an electronic device, the electronic device comprising: one or more processors; and a memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the conflict graph-driven incremental joint optimization method as described above.

[0016] Thirdly, some embodiments of this application also provide a computer-readable medium having computer program instructions stored thereon, which can be executed by a processor to implement the conflict graph-driven incremental joint optimization method as described above.

[0017] Fourthly, some embodiments of this application also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the conflict graph-driven incremental joint optimization method as described above.

[0018] Compared with related technologies, the solution provided in this application firstly constructs a communication tree conflict graph by extracting training-side semantics and physical network topology, realizing the explicit and accurate quantification of network occupancy conflicts between multiple communication trees; secondly, when generating candidate physical paths for hotspot logical edges and performing comprehensive scoring, a predicted inter-tree interference penalty term is introduced, employing a layered progressive filtering chain to accurately extract high-conflict communication tree nodes and their internal hotspot logical edges as the minimum optimization unit, performing local incremental rebalancing only on this high-conflict subset; by employing a layered... The progressive filtering chain precisely extracts high-conflict communication tree nodes and their internal hotspot logical edges as the smallest optimization unit, performing local incremental rebalancing only on this high-conflict subset to reduce the computational overhead and state search space of rebalancing. Secondly, a rigorous path version switching closed-loop control mechanism is used to construct a safe and smooth data plane switching closed loop, reducing online operation risks. Finally, incremental optimization of the high-conflict subset effectively alleviates the tail latency problem when multiple tasks or multiple communication streams are running concurrently, significantly reducing synchronization waiting time and congestion propagation in multi-task concurrent training. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0020] Figure 1 An exemplary flowchart of the conflict graph-driven incremental joint optimization method provided in the first embodiment;

[0021] Figure 2 An exemplary flowchart of the conflict graph-driven incremental joint optimization method provided in the second embodiment;

[0022] Figure 3 An exemplary structural diagram of the communication tree conflict graph provided in the second embodiment;

[0023] Figure 4 An exemplary structural diagram of an electronic device provided for a third embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In this embodiment of the disclosure, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information comply with relevant laws and regulations and do not violate public order and good morals.

[0026] First Embodiment

[0027] The first embodiment relates to a conflict graph-driven incremental joint optimization method. For example... Figure 1 As shown, the method may include the following steps:

[0028] S11, Obtain communication semantic information and physical network topology diagrams from the semantic source of the training book and / or the training control plane;

[0029] S12, On the physical network topology, establish the mapping relationship between logical communication edges in the communication tree and physical network links to determine the physical path occupancy set corresponding to the active communication tree;

[0030] S13, Construct a communication tree conflict graph based on the physical path occupancy set of each active communication tree;

[0031] S14, select high-conflict communication trees from the communication tree conflict graph, and after triggering the rebalancing mechanism, identify a high-conflict subset containing hotspot logical edges from the high-conflict communication trees, and perform incremental rebalancing on the high-conflict subset.

[0032] The core of this embodiment lies in obtaining communication semantic information and physical network topology from the semantic source of the training book and / or the training control plane; on the physical network topology, establishing a mapping relationship between logical communication edges in the communication tree and physical network links to determine the physical path occupancy set corresponding to the active communication tree; based on the physical path occupancy set of each active communication tree, constructing a communication tree conflict graph; filtering out high-conflict communication trees from the communication tree conflict graph, and after triggering the rebalancing mechanism, identifying a high-conflict subset containing hotspot logical edges from the high-conflict communication trees, and performing incremental rebalancing on the high-conflict subset.

[0033] It is not difficult to see that, compared with related technologies, this embodiment firstly, by extracting the semantics and physical network topology of the training side, innovatively constructs a communication tree conflict graph to achieve explicit and accurate quantification of network occupancy conflicts between multiple communication trees; secondly, by introducing a predicted inter-tree interference penalty term when generating candidate physical paths for hotspot logical edges and performing comprehensive scoring, a layer-by-layer progressive filtering chain is adopted to accurately extract high-conflict communication tree nodes and their internal hotspot logical edges as the minimum optimization unit, and only performs local incremental rebalancing solutions on this high-conflict subset; by adopting a layer-by-layer progressive filtering chain... The advanced filtering chain precisely extracts high-conflict communication tree nodes and their internal hotspot logical edges as the smallest optimization unit, performing local incremental rebalancing only on this high-conflict subset to reduce the computational overhead and state search space of rebalancing. Secondly, a rigorous path version switching closed-loop control mechanism is used to construct a safe and smooth data plane switching closed loop, reducing online operation risks. Finally, incremental optimization of the high-conflict subset effectively alleviates the tail latency problem when multiple tasks or multiple communication streams are running concurrently, significantly reducing synchronization waiting time and congestion propagation in multi-task concurrent training.

[0034] Second Embodiment

[0035] The second embodiment relates to another conflict graph-driven incremental joint optimization method. The second embodiment is an improvement on the first embodiment, with the core improvement being that it provides a more thorough and in-depth explanation of the relevant technical details compared to the first embodiment. Please refer to [link / reference]. Figure 2 This is shown as another exemplary flowchart of the conflict graph-driven incremental joint optimization method. Figure 2 As shown, the conflict graph-driven incremental joint optimization method specifically includes the following steps:

[0036] S21, acquire communication semantic information and physical network topology from the training semantic source and / or training control plane. In this embodiment, the communication semantic information is a set of structured multidimensional metadata characterizing the communication features of the AI ​​ensemble, including communication group information, communication primitive type, participating rank set, communication tree structure or equivalent logical transmission edge structure, message size, scheduling round, stage boundary, priority, and tree change events. The training-side semantic source includes the training framework, communication library, communication runtime, or equivalent components thereof. In actual deployment, by embedding hook functions or event listeners in the training-side semantic source, the control plane and data plane behaviors generated during the operation of the distributed AI training task are captured in real time.

[0037] The physical network topology map is derived from the cross-combination and fusion of four data sources: underlying physical connections, network controller policies, switch local configurations, and upper-layer AI training task location labels.

[0038] Specifically, the underlying physical connections are the device connection relationships output by the cluster asset management system or topology database. Network controller policies are the link states, fault domains, routing domains, and path policies maintained by the network controller. Switch local configurations include switch configurations, routing tables, ECMP / WCMP hash configurations, and tunnel policies. The upper-layer AI training task location labels are the mapping relationships between GPUs, nodes, racks, Pods, and task labels maintained by the training control plane.

[0039] Preferably, the physical network topology is defined as:

[0040]

[0041] in, Represents the physical network topology diagram, where N is the set of physical nodes, including at least a graphics processing unit (GPU), a network interface card (NIC), a top / leaf switch, a spine switch, a cross-cluster unit gateway, or an equivalent network device.

[0042] L is a set of directed links, and each link corresponds to at least one link ID and direction.

[0043] Attr N Node attributes include at least device type, rack, pod, fault domain, and management label.

[0044] Attr L Link attributes include at least capacity, network layer, baseline delay, routing domain, equal-cost multipath group, and health.

[0045] S22, on the physical network topology, establish the mapping relationship between logical communication edges in the communication tree and physical network links to determine the physical path occupancy set corresponding to the active communication tree.

[0046] To achieve refined digital modeling of cross-layer traffic, a standardized core data foundation is provided for constructing conflict graphs and ensuring incremental rebalancing of security, such as... Figure 2 As shown, step S22 specifically includes the following steps:

[0047] S221, for each logical parent-child edge in each communication tree, establish the sending endpoint and receiving endpoint of each logical parent-child edge.

[0048] Specifically, the communication tree in the physical network topology is represented at the logical layer as a parent-child structure consisting of a set of logical nodes or processes. The algorithm traverses each currently active communication tree and extracts its logical topology. For any directed logical parent-child edge in the tree... First, by training the task location labels and the distribution mapping of the cluster scheduler, the specific computing hardware identifier bound to the logical parent-child edge at the application layer is obtained. This clarifies the logical parent-child edge. The sending end corresponds to the source physical device, and the receiving end corresponds to the destination physical device.

[0049] S222, based on the sending and receiving endpoints of each logical parent-child edge, as well as the network configuration and strategy, a combined solution is performed to obtain one or more physical links corresponding to each logical parent-child edge.

[0050] Specifically, by combining the cluster topology database, existing equivalent multipath / weighted multipath hash rules, static tunnel tables, segmented routing policies, or explicit paths issued by the software-defined network controller, one or more actual physical links corresponding to the logical parent-child edge can be obtained. For example, the path is: sender → local topology → source network card → cabinet top / leaf switch → spine switch → cross-cluster unit gateway → destination leaf / cabinet top switch → destination network card → receiver.

[0051] S223 distributes the communication load to each physical link from a spatial dimension and distributes the time window to each physical link from a temporal dimension.

[0052] Specifically, distributing the communication load to each physical link from a spatial perspective means obtaining the message size transmitted in the set of communication tasks and combining it with the flow distribution weights in the network routing strategy (such as the hash dynamic distribution of Equal Cost Multipath (ECMP) or the bandwidth ratio of Weighted Multipath (WCMP)). If multiple physical paths are determined, the total number of bytes expected to be occupied is proportionally divided and distributed to the corresponding physical links according to the weights.

[0053] Distributing the time window across each physical link from a temporal perspective involves combining the scheduling rounds captured by the semantic sources on the distributed training side with the predicted baseline of stage boundary timeouts in historical iterations to calculate the transmission and cutoff times of traffic on that logical edge. This allows for the derivation of the expected start and end time periods of that traffic on the underlying physical link, thus establishing the time window for the expected start and end times of occupancy on that directed physical link.

[0054] S224, generate a physical link occupancy set for each communication tree. The physical link occupancy set includes the communication tree identifier, logical edge identifier, physical path identifier, physical link identifier, estimated occupancy bytes, training phase / batch, estimated occupancy start time, estimated occupancy end time, path mapping version number, and traffic binding key. The communication tree identifier and logical edge identifier are used for accurate tracing of upper-layer services. The physical path identifier and physical link identifier are used to locate the underlying hardware physical trajectory.

[0055] The estimated byte size is used to measure space resource consumption. The training phase / batch, estimated start time, and estimated end time of the estimated byte size lock the time dimension window. The path mapping version number records the iterative version of the current mapping scheme for subsequent hot-swapping between versions. The traffic binding key is used to strongly bind the network controller to hardware devices to perform precise incremental routing switching for local hotspot logical edges.

[0056] S23. Construct a communication tree conflict graph based on the physical path occupancy set of each active communication tree.

[0057] To construct a communication tree conflict graph that can dynamically and intuitively reflect the actual interference intensity among communication tasks across the entire network, a key graph theory decision model and quantitative data foundation are provided for subsequent low-overhead, accurate local rebalancing. Figure 2 As shown, step S23 specifically includes the following steps:

[0058] S231: For any two communication trees, project the corresponding physical link occupancy sets according to the directed link dimension, sum the expected occupancy bytes on the same physical link, and merge the expected occupancy time windows.

[0059] Specifically, any two communication trees T i And Tj, will the communication tree T i The physical link occupancy set O corresponding to Tj i and O j Projecting along the directed link dimension yields the link occupancy record. and .

[0060]

[0061] Where l represents a directed physical link, and Communication Tree T i The number of bytes occupied by the aggregation of Tj on the directed link l. and Represents the communication tree T i The estimated time window occupied by Tj on directed link l.

[0062] ,

[0063] ,

[0064] ,

[0065] ,

[0066] ,

[0067] This indicates the number of bytes in the message or traffic represented by record r; when record r is a single message record, it indicates the byte length of that message; when record r is a traffic statistics record, it indicates the cumulative number of bytes counted by that record. It does not refer to the size of the data structure of record r itself.

[0068] dirlink(r) represents extracting the underlying directed link key from the occupancy record r.<link_id,direction> verlink(r) represents the extraction of the versioned directed link key from the occupied record r when dual-version parallelism exists.<link_id,direction, route_version> .

[0069] In the spatial dimension, the total estimated load of the communication tree on the physical link is obtained by linearly summing the number of bytes it is expected to occupy. In the temporal dimension, if multiple estimated time windows of the communication tree on the link overlap or are consecutive, they are merged into one or more disjoint, consecutive estimated time windows.

[0070] S232, when the estimated time windows of two communication trees on the same directed physical link overlap, the directed physical link is included in the shared link set of the two communication trees. The shared link set includes link identifier, link direction, link capacity, real-time load within the detection window, estimated byte count of each of the two communication trees on the directed physical link, time overlap ratio, and link level identifier.

[0071] Specifically, the communication tree T i and T j The estimated time window for merging on the same directed physical link. If the estimated time windows of the two overlap, it is determined that there is a risk of time collision in the communication flow on that physical link, and the directed physical link is classified as two communication trees T. i and T j Share the conflicting link and incorporate the directed physical link into the two communication trees T. i and T j Shared link set ， .

[0072] S233, from the shared link set The real-time load and link capacity of each link are obtained and combined with the normalized link occupancy share and time overlap ratio of the two communication trees on the corresponding directed physical links to determine the conflict edge weight between the two communication trees.

[0073] Specifically, in the shared link set Calculate Ti and T for all affected physical links within the network. j The severity of the conflict between them, i.e., the conflicting borders .

[0074]

[0075] in, This is the shared directed physical link between Ti and Tj.

[0076] Indicates the link normalized load.

[0077]

[0078] The preferred method is the average throughput rate or equivalent port utilization within the current detection window. For the nominal or available bandwidth rate of the link, faulty or unavailable links are preferably removed directly from the candidate path set.

[0079] This shows the normalized link occupancy share of the two communication trees on the link.

[0080]

[0081]

[0082] in, , This represents the total number of bytes occupied by the corresponding communication tree within the current detection window.

[0083] This is the link weight coefficient, preferably between 0 and 1, used to distinguish the importance of links in the spine layer, leaf layer, cross-cluster unit, or cross-rack.

[0084] The ratio of time overlap between two communication trees on the link is preferably defined as:

[0085]

[0086] in, This represents the total duration of the interval set. If the denominator is 0, then... Record it as 0.

[0087] S234, using the communication tree as nodes, and using edges in the shared physical link set that have at least one directed physical link and whose corresponding time windows overlap as conflict edges, and using the conflict intensity as the conflict edge weight, constructs a system as follows: Figure 3The communication tree conflict diagram is shown.

[0088] S24, select high-conflict communication trees from the communication tree conflict graph, and after triggering the rebalancing state, identify a high-conflict subset containing hotspot logical edges from the high-conflict communication trees, and perform incremental rebalancing on the high-conflict subset.

[0089] To reduce the overhead of global recalculation, such as Figure 2 As shown, step S24 specifically includes the following steps:

[0090] S241, based on the communication tree conflict graph, select high-conflict communication trees, and identify hotspot logical edges that cause conflicts within the high-conflict communication trees.

[0091] Specifically, based on the constructed communication tree conflict graph, the conflict edge weights of the conflict edges connected to each node are extracted. The weighted conflict degree of each communication tree is calculated. Then, the entire network's communication trees are sorted in descending order of their weighted conflict degrees, or the endpoint nodes connected by high-weight conflict edges are extracted. This allows us to identify the communication trees that have contributed the most to the network performance degradation and urgently require optimization. As the target to be optimized, its associated set of highly conflicting neighbors is identified. .

[0092] The union of all shared physical links between this communication tree and its high-collision neighbor set forms the hotspot observation link set. Hotspot observation link collection for:

[0093]

[0094] Traverse every logical parent-child edge within the communication tree. Determine the logical parent-child edge. The physical path currently used Collection of hotspot observation links Intersection links.

[0095] Based on the normalized load of the intersection links and the normalized occupancy share of the logical edge on the links, the logical parent-child edge is calculated. Hot topic heuristic contribution. Logical parent-child edge. The hot topic heuristic contribution is defined as:

[0096]

[0097] in, For link normalization load, This represents the normalized link occupancy share of the total link occupancy bytes of the communication tree within the current detection window.

[0098] , .

[0099] Sort the edges in descending order of their hotspot heuristic contribution, and select the top N logical parent-child edges or those with a cumulative contribution exceeding a preset ratio as the hotspot logical edges that cause conflicts.

[0100] S242, for the hotspot logical edges, candidate physical paths are generated on the physical network topology graph, and a comprehensive path score is determined based on the basic network score of the candidate physical paths and the predicted inter-tree interference penalty term. In this embodiment, the inter-tree interference penalty term can prevent path search from pursuing only the local optimum of a single communication tree and migrating it to another path that is also occupied by other critical communication trees; at the same time, it not only considers the currently adjacent conflicting trees, but also predicts the conflicting objects that may be added after the candidate path is switched, so that the path selection can explicitly avoid the critical links that high-priority tasks are using, reduce the situation where the benefit of this tree increases but the global synchronization tail delay deteriorates, and suppress path oscillations caused by frequent switching.

[0101] Specifically, for the identified hotspot logical edges, a routing algorithm is run on the latest physical network topology graph based on their source / destination endpoints. This algorithm may employ methods such as K-shortest path, constrained shortest path, explicit tunnel pooling, pre-computed segmented routing path pooling, or equivalent pathfinding to avoid links currently restricted by hardware failures or static policies, generating candidate physical paths on the physical network topology graph. The physical characteristics of the candidate physical paths are evaluated, and combined with the priority of potentially interfered communication trees and the predicted intensity of new conflicts, an inter-tree interference penalty is determined. Based on the basic network score and its weights, and the inter-tree interference penalty and its weights, a comprehensive path score is determined. This scoring mechanism effectively suppresses global conflicts caused by local path optimization, ensuring that new paths have global collaborative benefits. In this embodiment, the physical characteristics of the candidate physical paths include the physical hop count, the dynamic remaining bandwidth of each physical link, the interface queue cache status, and the background traffic load. The basic score is directly proportional to the remaining bandwidth and inversely proportional to the hop count and dynamic load.

[0102] In this embodiment, the basic network score includes a latency benefit term Q(p), an available bandwidth benefit term C(p), a congestion steady-state benefit term R(p), and a path switching cost term P(p).

[0103] The basic network score is further defined as:

[0104]

[0105] Where delay(p) represents the predicted delay of the candidate path, which can be estimated based on link propagation delay, queuing delay, and number of switching hops; available_bw(p) represents the remaining bandwidth of the candidate path bottleneck; and rho(l) represents the link normalized load, which is the predicted normalized load after superimposing candidate traffic in the candidate path scoring. If the current load metric is used, it is necessary to combine the remaining bandwidth threshold and capacity constraints for filtering; switch_cost(p) represents the control plane overhead, cache mismatch overhead, transient out-of-order risk, or version switching waiting cost caused by path switching; delay ref bw ref cost ref This serves as the reference upper bound or baseline value used for normalization within the current detection window. Preferably, the same set of reference values ​​is used for the candidate path scoring in the same round. The reference values ​​can be the maximum value in the current candidate path set, the historical baseline window quantile, the SLA upper bound, or the configuration upper bound. Thus, Q(p), C(p), R(p), and P(p) are all normalized to within 0 to 1, facilitating a unified comparison with the inter-tree interference penalty term.

[0106] When using the predicted load metric, the set of candidate switching units that have been accepted but not yet submitted in this round can be denoted as A, and preferably defined as:

[0107]

[0108] in, This represents the expected rate contribution of the old path along the logical parent-child edge on link l. This represents the expected rate contribution on link l after the logical parent-child edge switches to candidate path p; if the old path or candidate path does not contain link l, the corresponding item is 0. When scoring a single candidate, the candidate switching unit set A is empty; when accepting multiple candidates in the same detection window according to their revenue order, the candidate switching unit set, remaining bandwidth, and predicted load caliber of the candidate switching unit are updated after each candidate switching unit is accepted. Then, evaluate the mutual exclusion relationships and capacity constraints of subsequent candidates.

[0109] The inter-tree interference penalty term is further defined as:

[0110]

[0111] in, The reference upper bound used to normalize the interference penalty term to the interval between 0 and 1 is preferred to be consistent in the same round of candidate path scoring. This represents the total, unnormalized, predictive inter-tree interference penalty. It quantifies the total potential spatiotemporal conflict interference intensity caused to all other communication trees in the entire network when a hotspot logical edge of the communication tree to be optimized switches to a new candidate path p. Specifically defined as:

[0112]

[0113] in, This represents the normalized load of the link, indicating the current level of congestion on the physical link. This represents the link weight coefficient, used to distinguish the importance of backbone links from edge links; This indicates the newly added tree-level normalized occupancy share, which refers to the net increase in traffic on link l after this logical edge switches to the new link l, relative to the old path. Represents the neighbor communication tree T j Normalized link occupancy share on link l: Predicted time overlap ratio: This item predicts the degree of overlap between the traffic transmission / transmission time window on the new path and the transmission time window of the neighboring tree on the link on the time axis. Indicates the normalization priority of the neighbor tree. This indicates the communication trees Ti and T when the candidate path p to be evaluated is adopted. j The set of shared directed links formed with candidate path p. Among them, Specifically defined as:

[0114]

[0115] in, , This represents the estimated number of bytes occupied on link l after the same logical parent-child edge switches to candidate path p. This represents the estimated number of bytes occupied on link l after the same logical parent-child edge switches to candidate path p. This represents the total number of hop bytes in the old path of the logical parent-child edge. This represents the total number of hop bytes in the link after the logical parent-child edge adopts candidate path p. This represents the total number of bytes occupied by the predicted tree-level links after candidate path p becomes effective.

[0116] Therefore, for the candidate path p of optimizing the logical parent-child edge, the comprehensive path score is defined as:

[0117]

[0118] Where Q(p) is the delay benefit term, C(p) is the available bandwidth benefit term, R(p) is the congestion steady-state benefit term, P(p) is the path switching cost term, and I(p) is the inter-tree interference penalty term. , , , , All are non-negative weights, and preferably satisfy the following conditions: .

[0119] S243, when the rebalancing mechanism is triggered, the path mapping of the communication tree except for the high-conflict subset remains unchanged, and path switching is performed only on the hotspot logical edges that meet the switchability qualification and the comprehensive path score meets the benefit conditions, thereby realizing the incremental rebalancing closed loop.

[0120] In this embodiment, rebalancing is triggered when any of the following conditions are met:

[0121] The global conflict index is greater than the global conflict trigger threshold;

[0122] There exists a conflict edge weight between any two communication trees that is greater than the local conflict trigger threshold;

[0123] The actual communication latency of a communication tree exceeds a preset multiple of its predicted baseline;

[0124] The creation of new communication groups, the destruction of existing communication groups, link failures, or path version changes cause changes in the topology of the communication tree conflict graph. Among these changes is the global conflict index.

[0125]

[0126] In this embodiment, GCI is used to characterize the average conflict intensity between communication tree pairs in the current entire training network; for communication tree pairs without conflicting edges, let =0, therefore the above summation can be equivalently understood as summing all communication tree pairs and setting non-shared link tree pairs to zero. Since Since it has been normalized to the 0-1 range, the GCI also falls within the 0-1 range. It is a global health indicator that determines whether a transition from continuous monitoring to proactive rebalancing is needed. Local conflict edge weights... The edge weights are directly taken from the conflict graph of the communication tree and used to identify specific conflict hotspots.

[0127] In a preferred embodiment, step S243 specifically includes the following steps:

[0128] S2431, For each hotspot logical edge that meets the switching eligibility and comprehensive path score benefit conditions, calculate its switching benefit judgment value, and combine the corresponding communication tree, hotspot logical edge and candidate physical path to construct a candidate switching unit. .

[0129] To avoid inconsistencies in the comparison criteria between the old and new paths, it is preferable to regard the current old path as the baseline path p. old The same detection window, the same upper reference bound, and the same weights are used to calculate the Score(p).old Before submitting the switchover, further define the switchover benefit determination value:

[0130]

[0131] in, and To enable configurable non-negative weights, This represents the expected decrease in the conflict level of the target communication tree after the candidate path takes effect. Preferably, the weighted conflict level of the target communication tree is defined first. ,but This represents the weighted conflict degree under the current old path mapping. This indicates that the logical parent-child edge has been switched to a candidate path. The predicted weighted conflict degree after it takes effect This indicates the upper bound of the conflict reduction normalization reference. Indicates except For newly added risk items outside of the initial risk level, a normalized combination of candidate path traversal of fault domains, version switch waiting, or high-priority task protection constraints can be used. For simplified implementation, it can also be made... and ,at this time Degenerate to the same caliber .

[0132] S2432, the candidate handover units are sorted in descending order according to the handover benefit judgment value, and resource mutual exclusion filtering is performed based on the link capacity constraints of each candidate physical path in the underlying network and the physical link resource occupancy between each candidate handover unit to eliminate candidate paths that have link conflicts within the same time window, forming a set of non-mutually exclusive candidate handover units. Non-mutually exclusive candidate switching unit set Defined as:

[0133]

[0134] in, Let H represent the switching benefit judgment value of the candidate path relative to the logical parent-child edge of the communication tree, and let H represent the set of communication trees that contains at least one high-benefit candidate switching unit after scoring, defined as:

[0135]

[0136] in, This represents the pre-selected communication tree set for candidate path generation and scoring. This represents the set of candidate switching units within the communication tree. Each candidate switching unit is bound to the communication tree, logical parent-child edges, and candidate paths, and is defined as follows:

[0137]

[0138] in, This represents the set of hotspot logical edges within the communication tree that lead to the recalculation of candidate paths, and is defined as follows:

[0139] ={ | For T i Hotspot logic edges selected based on contribution

[0140] The set of candidate paths representing hotspot logical edges is specifically defined as follows:

[0141]

[0142] S2433, count the number of communication trees involved in the set of non-mutually exclusive candidate switching units. When the number exceeds the preset threshold for the maximum number of switching trees in a single window, intercept the candidate switching units that are ranked first and send them down. Allow the candidate switching units that are ranked first to switch paths, so as to limit the total number of communication trees that actually switch paths in the current detection window. Units that do not meet the interception queuing conditions are postponed to the next detection window for processing, thereby strictly limiting the total number of communication trees that actually switch paths in the current window and reducing the concurrent pressure on the control plane.

[0143] The path switching of the candidate switching unit can be implemented using the following scheme:

[0144] The conflict graph controller first retains the old path version (route_version_old) for the target logical edge and then selects the new path version (route_version_new).

[0145] The network controller first pre-installs the forwarding table, explicit tunnel, SRv6 segment list, MPLS label path, ECMP weight scheme or its equivalent path rules corresponding to the new path version route_version_new on the data plane, without immediately deleting the old path version.

[0146] When the training framework enters a set communication boundary, training step boundary, or preset switching window that allows switching, the conflict graph controller issues a version switching instruction to the network controller, causing subsequent target logical edge traffic to be forwarded according to the new path version route_version_new.

[0147] Within the verification window, continuously collect the load of the affected links, the actual latency of the affected communication trees, and the changes in the weights of local conflict edges.

[0148] If the verification results meet the benefit conditions, submit the new path version and release the old version resources; if the benefit conditions are not met, high-priority business deterioration occurs, or the link is abnormal, roll back to the old path version route_version_old.

[0149] After rollback, the conflict graph controller restores the affected communication tree with the path occupancy record corresponding to the old path version route_version_old, recalculates the affected conflict edge weights and prediction baseline, marks the new path version route_version_new as a failure candidate and records the reason for failure; the cooling status can distinguish between successful switch cooling and failure suppression cooling, avoiding misrecording ineffective candidates as successful switches.

[0150] S2434, the network controller pre-installs new path forwarding rules on each physical switching device involved in the intercepted candidate switching unit, and retains the old path forwarding rules currently in use for the hotspot logical edge. At this time, the new path forwarding rules are inactive, and the corresponding path mapping version number is incremented.

[0151] S2435, continuously collect the running status of the AI ​​training framework, and when it is detected that the training framework has entered the task boundary or switching window that allows switching, issue a version switching instruction to the physical switching device to activate the new path forwarding rule, so that the subsequent aggregated communication traffic can be seamlessly switched to the physical path corresponding to the new path forwarding rule for forwarding.

[0152] S2436, within the preset verification window, the network controller and topology database continuously collect the real-time load and latency of the affected physical links and dynamically update the communication tree conflict graph; when the overall network conflict level is detected to decrease or the preset benefit conditions are met, the new path forwarding rule is submitted as the officially effective version, and the corresponding path mapping version number is fixed and updated in the communication tree conflict graph, and the network resources occupied by the old path forwarding rule are released asynchronously.

[0153] S2437 If the benefit condition is not met within the verification window, or if a higher-priority AI training task is detected to have degraded performance, a rollback command is immediately issued to switch the aggregated communication traffic transiently back to the old path forwarding rule, and the incremental state of the communication tree conflict graph is restored based on the historical link occupancy record corresponding to the old path forwarding rule, thus completing the rebalancing control closed loop.

[0154] S2438, after completing the path switch or determining not to switch, update the physical path occupancy set, conflict edge weights, cooling status, and predicted baseline of the affected communication tree, and generate a persistent conflict event record. The persistent conflict event record refers to persisting the detection and processing results in the form of logs, database records, or message queue messages. It includes fields such as timestamp, communication tree ID, hotspot logical edge ID, shared link ID, edge weights before and after the switch, trigger reason, selected new path version, latency change, benefit value, and whether rollback is required. These fields are used for subsequent traceability, parameter tuning, and preservation of evidence for patent implementation.

[0155] The above conflict graph-driven incremental joint optimization method is applied to a multi-parallel group training scenario. The following is a detailed description of the application:

[0156] In a training cluster containing 512 GPUs, there are 16 data parallel groups simultaneously.

[0157] There are 8 tensor parallel groups and 4 pipeline parallel groups.

[0158] The conflict graph controller obtains all communication group mappings from the training framework and constructs 28 active communication trees. At a certain moment, it detects:

[0159] Tree T7 in the tensor parallel group and tree T in the data parallel group 13 Shared Spine link.

[0160] Bian Quan w 7,13 It has remained above the threshold for the past 5 periods.

[0161] The actual full reduction delay of the corresponding communication tree exceeded the baseline by 24% and 31%, respectively.

[0162] Based on this, the conflict graph controller selects the three trees with the highest conflict degree for incremental recalculation, and further identifies T7 and T... 13 The hottest logic edges that contribute the most; then candidate switching units are formed for these hottest logic edges. ,calculate and And obtain a new path mapping by reducing the overlap ratio of shared links. If the threshold is reached and the switching window is open, the new path version is pre-installed and the switch is performed. After the update, the weights of locally conflicting edges decrease, and the tail waiting time of the training iteration converges significantly.

[0163] To illustrate the feasibility of the above formula, a non-restrictive small-scale calculation example is given below. Suppose there are three communication trees T1, T2, and T3 within the current detection window, and the total link occupancy (occ) used for normalization is given. i All are 1000-byte hops, and the load on the three shared links is... =0.70、 =0.80、 =0.45, and All are set to 1. If T1's share in L1 and L2 is 0.5 and 0.5 respectively, T2's share in L2 and L3 is 0.6 and 0.4 respectively, and T3's share in L1 and L3 is 0.7 and 0.3 respectively, and let... (L2)=0.75 (L1)=0.60 (L3)(=0.50, then:

[0164]

[0165] like This triggers the incremental rebalancing mechanism. For a certain hotspot logic edge, set the old path The scoring items are , , , , Candidate new paths The scoring items are , , , , Weighting , , , , ,but:

[0166]

[0167] like If so, then the candidate path meets the switching condition. After switching, if If the weights are reduced to 0.12, and the remaining edge weights remain unchanged, then... It dropped to approximately 0.1325, a decrease of about 31.2%, thus reproducibly demonstrating the causal chain between conflict graph metrics, candidate path scores, and rebalancing gains.

[0168] The above conflict graph-driven incremental joint optimization method is applied to a scenario of dynamic changes in communication groups. The following is a detailed description of the application:

[0169] When a new expert parallel group is launched, the training framework sends a communication group creation event to the conflict graph controller. The conflict graph controller only adds nodes and edges related to the new tree and calculates the weights of the affected edges; if the GCI does not exceed the threshold, it only records and continuously monitors, without triggering rebalancing; if it exceeds the threshold, it only recalculates the relevant subset. If the candidate path score is not significantly better than the current path, it keeps the original mapping unchanged, only refreshes the prediction baseline and event records, and does not add a successful switch cooldown timer.

[0170] The conflict graph-driven incremental joint optimization method described above is applied to a rollback scenario due to insufficient returns. The following is a detailed description of the application:

[0171] When a high-priority communication tree T 21 When a conflict occurs with the normal priority communication tree T9, the conflict graph controller selects a candidate path for the hotspot logical edge of T9 and performs a path version switch at the next collective boundary. If T is found in the verification window... 21 If the actual latency rises above the threshold or the local conflict edge weights in T9 decrease by less than a preset proportion, the network controller rolls back to route_version_old. The conflict graph controller restores the physical path occupancy set and affected conflict edge weights corresponding to the old path version, and writes this failure record to the event log for future improvement. Extend the failure suppression cooling window or reduce the priority of the same candidate path.

[0172] It is not difficult to see that, compared with related technologies, this application firstly, by extracting the semantics of the training side and the physical network topology, innovatively constructs a communication tree conflict graph to achieve explicit and accurate quantification of network occupancy conflicts between multiple communication trees; secondly, by introducing a predicted inter-tree interference penalty term and using a layer-by-layer progressive filtering chain when generating candidate physical paths for hotspot logical edges, it accurately extracts high-conflict communication tree nodes and their internal hotspot logical edges as the minimum optimization unit, and only performs local incremental rebalancing on this high-conflict subset; by adopting a layer-by-layer progressive... The filtering chain precisely extracts high-conflict communication tree nodes and their internal hotspot logical edges as the smallest optimization unit, and performs local incremental rebalancing only on this high-conflict subset, reducing the computational overhead and state search space of rebalancing. Secondly, a rigorous path version switching closed-loop control mechanism is used to construct a safe and smooth data plane switching closed loop, reducing the risk of online operation. Finally, through incremental optimization of the high-conflict subset, the tail latency problem in multi-task or multi-communication stream concurrency is effectively alleviated, and the synchronization waiting time and congestion spread in multi-task concurrent training are significantly reduced.

[0173] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0174] Third Embodiment

[0175] This application provides an electronic device. The electronic device can be various forms of digital computer, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, mainframe computers, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0176] The electronic device includes: one or more processors; and a memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the methods provided in any one or more of the above embodiments.

[0177] Figure 4 An exemplary structural diagram of the electronic device is disclosed. The electronic device includes one or more processors 1101, a memory 1102, an input device 1103, and an output device 1104. The various components are interconnected via a bus or other means (the diagram shows an example of bus connection). The processor 1101 can be used to execute instructions stored in the memory 1102 to control the overall operation of the electronic device. The memory 1102 may include a program storage area and a data storage area, wherein the program storage area stores the operating system and applications required for at least one function; the data storage area stores data created according to the use of the electronic device, etc. The memory 1102 may include high-speed random access memory and may also include non-transitory memory, such as disk storage devices, flash memory devices, or other non-transitory solid-state storage devices. In some embodiments, the memory 1102 may also include storage resources located remotely to the processor and accessible via a network.

[0178] Input device 1103 can be used to receive input numerical or character information or user operation signals, such as a touch screen, keypad, mouse, trackpad, touchpad, indicator, one or more mouse buttons, trackball, joystick, etc. Output device 1104 may include display devices (such as liquid crystal displays, light-emitting diode displays, plasma displays, and optional touch screens), auxiliary lighting devices (such as LEDs), and haptic feedback devices (such as vibration motors), etc.

[0179] To facilitate user interaction, the electronic device may be configured to include a display device (such as an LCD or CRT monitor) and input devices such as a keyboard and pointing devices (e.g., a mouse or touchpad). Feedback can be any form of sensory feedback (e.g., visual feedback, auditory feedback); input may also be received via voice, touch, or other means.

[0180] This application also relates to a computer-readable medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the conflict graph-driven incremental joint optimization method provided in any one or more of the above embodiments. The computer-readable medium may be a memory included in an electronic device, or it may be a standalone storage medium not assembled into the device.

[0181] It should be noted that the computer-readable medium described in this application may be a computer-readable signal medium, a computer-readable storage medium, or a combination of both. Examples include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. Specific examples of storage media may include, but are not limited to, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory, optical fibers, portable CD-ROMs, optical storage devices, magnetic storage devices, etc., or any suitable combination thereof.

[0182] Computer-readable media may store one or more programs that can be used by or in conjunction with an instruction execution system. The media may be permanent or non-permanent, removable or non-removable, and may store information by any method or technology, including computer-readable instructions, data structures, program modules, or other data.

[0183] The computer program code used to implement the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​(such as Java, Smalltalk, and C++) and conventional procedural programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer, partially on a remote computer, or entirely on a remote computer or server. The remote computer can be connected to the user's computer via any network (including a local area network or a wide area network) or can be connected to an external computer.

[0184] In the above embodiments, the functions can be implemented in whole or in part by software, hardware, firmware, or any combination thereof, for example, by using application-specific integrated circuits, general-purpose computers, or other similar hardware devices. In some embodiments, the software program of this application can be executed by a processor to implement the steps or functions; it can also be implemented by hardware, for example, as a circuit that works in conjunction with the processor to execute the steps or functions.

[0185] This application also provides a computer program product, including one or more computer programs / instructions, which, when executed by a processor, generate all or part of the processes or functions described in this application. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium or transmitted from one storage medium to another via wired (e.g., DSL) or wireless (e.g., wireless, microwave) means. The computer-readable storage medium may be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive).

[0186] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0187] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0188] 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 made 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, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A conflict graph-driven incremental joint optimization method, characterized in that, Obtain communication semantic information and physical network topology from semantic sources on the training side and / or the training control plane; In the physical network topology, a mapping relationship is established between logical communication edges in the communication tree and physical network links to determine the physical path occupancy set corresponding to the active communication tree; Based on the physical path occupancy set of each active communication tree, a communication tree conflict graph is constructed; High-conflict communication trees are selected from the communication tree conflict graph. After the rebalancing mechanism is triggered, a high-conflict subset containing hotspot logical edges is identified from the high-conflict communication trees, and incremental rebalancing is performed on the high-conflict subset.

2. The conflict graph-driven incremental joint optimization method according to claim 1, characterized in that, The communication semantic information includes communication group information, communication primitive type, participating rank set, communication tree structure or equivalent logical transmission edge structure, message size, scheduling round, stage boundary, priority and tree change event, wherein the semantic source on the training side includes training framework, communication library, communication runtime or equivalent components thereof; The physical network topology map is derived from the cross-combination and fusion of four data sources: underlying physical connections, network controller policies, switch local configurations, and upper-layer AI training task location labels.

3. The conflict graph-driven incremental joint optimization method according to claim 1, characterized in that, In terms of the physical network topology, the steps to establish the mapping relationship between logical communication edges in the communication tree and physical network links, in order to determine the physical path occupancy set corresponding to the active communication tree, include: For each logical parent-child edge in each communication tree, determine the sending endpoint and receiving endpoint of each logical parent-child edge; The solution is obtained by combining the sending and receiving endpoints of each logical parent-child edge with the network configuration and routing strategy to obtain one or more physical links corresponding to each logical parent-child edge. The communication load is distributed to each physical link from a spatial dimension, and the time window is distributed to each physical link from a temporal dimension. A physical link occupancy set is generated for each communication tree; the physical link occupancy set includes the communication tree identifier, logical edge identifier, physical path identifier, physical link identifier, expected occupancy byte size, training phase / batch, expected occupancy start time, expected occupancy end time, path mapping version number, and traffic binding key.

4. The conflict graph-driven incremental joint optimization method according to claim 3, characterized in that, Based on the physical path occupancy set of each active communication tree, the steps to construct a communication tree conflict graph include: For any two communication trees, project their respective physical link occupancy sets according to the directed link dimension, sum the expected occupancy bytes on the same physical link, and merge the expected occupancy time windows. When the expected time windows of two communication trees overlap on the same directed physical link, the directed physical link is included in the shared link set of the two communication trees; the shared link set includes link identifier, link direction, link capacity, real-time load within the detection window, expected byte count of each of the two communication trees on the directed physical link, time overlap ratio, and link level identifier. The real-time load and link capacity of each link are obtained from the shared link set, and combined with the normalized link occupancy share and time overlap ratio of the two communication trees on the corresponding directed physical links to determine the conflict edge weight between the two communication trees. The communication tree conflict graph is constructed using each communication tree as a node, the edge that has an overlapping relationship with at least one directed physical link in the shared physical link set and the corresponding occupied time window as the conflict edge, and the conflict intensity as the conflict edge weight.

5. The conflict graph-driven incremental joint optimization method according to claim 4, characterized in that, The steps of filtering high-conflict communication trees from the communication tree conflict graph, identifying a high-conflict subset containing hotspot logical edges from the high-conflict communication trees after triggering a rebalancing state, and performing incremental rebalancing on the high-conflict subset include: Based on the communication tree conflict graph, high-conflict communication trees are selected, and hotspot logic edges that cause conflicts are identified within the high-conflict communication trees. For the hotspot logical edges, candidate physical paths are generated on the physical network topology graph, and a comprehensive path score is determined based on the basic network score of the candidate physical paths and the predicted inter-tree interference penalty term. When the rebalancing mechanism is triggered, the path mapping of the communication tree, except for the high-conflict subset, remains unchanged. Path switching is only performed on the hotspot logical edges that meet the switchability qualification and whose comprehensive path score meets the benefit conditions, thereby realizing an incremental rebalancing closed loop.

6. The conflict graph-driven incremental joint optimization method according to claim 5, characterized in that, The steps of filtering high-conflict communication trees based on the communication tree conflict graph and identifying hotspot logical edges causing conflicts within the high-conflict communication trees include: Based on the conflict edge weights in the conflict graph, each communication tree is sorted by its conflict degree, or the endpoints connected by high-weight conflict edges are extracted to filter out the communication trees to be optimized and determine their associated set of high-conflict neighbors. The shared links between the communication tree and its high-conflict neighbors are combined to form a hotspot observation link set. For any logical parent-child edge within the communication tree, calculate the intersection link between the physical path currently used by the logical parent-child edge and the hotspot observation link set, and calculate the hotspot heuristic contribution of the logical parent-child edge based on the normalized load of the intersection link and the normalized occupancy share of the edge on the intersection link. Based on the hotspot heuristic contribution, the logical parent-child edges in the communication tree are sorted in descending order, and the top N logical parent-child edges or those with a cumulative contribution exceeding a preset proportion are selected as the hotspot logical edges.

7. The conflict graph-driven incremental joint optimization method according to claim 4, characterized in that, The rebalancing mechanism is triggered by any of the following conditions: The global conflict index is greater than the global conflict trigger threshold; There exists a conflict edge weight between any two communication trees that is greater than the local conflict trigger threshold; The actual communication latency of a communication tree exceeds a preset multiple of its predicted baseline; The creation of new communication groups, the destruction of existing communication groups, link failures, or changes in path versions cause changes in the topology of the communication tree conflict graph; When the rebalancing mechanism is triggered, the path mapping of the communication tree, except for the high-collision subset, remains unchanged. The steps for performing path switching only on the hotspot logical edges that meet the switchability qualification and whose comprehensive path score meets the benefit conditions, to achieve incremental rebalancing closed loop, include: For each hotspot logical edge that meets the switching eligibility and the comprehensive path score benefit conditions, calculate its switching benefit judgment value, and combine the corresponding communication tree, hotspot logical edge and candidate physical path to construct a candidate switching unit; The candidate switching units are sorted in descending order according to the switching benefit judgment value. Based on the link capacity constraints of each candidate physical path in the underlying network and the physical link resource occupancy between each candidate switching unit, resource mutual exclusion filtering is performed to eliminate candidate paths that have link conflicts in the same time window, forming a set of non-mutually exclusive candidate switching units. The number of communication trees involved in the set of non-mutually exclusive candidate switching units is counted. When the number exceeds the preset threshold for the maximum number of switching trees in a single window, the candidate switching units that are ranked first are selected and sent down to limit the total number of communication trees that actually undergo path switching in the current detection window. The network controller pre-installs new path forwarding rules on each physical switching device involved in the intercepted candidate switching unit, and retains the old path forwarding rules currently in use for the hotspot logical edge. At this time, the new path forwarding rules are inactive, and the corresponding path mapping version number is incremented. The system continuously collects the running status of the AI ​​training framework. When it detects that the training framework has entered a task boundary or switching window that allows switching, it issues a version switching instruction to the physical switching device to activate the new path forwarding rule, so that subsequent aggregated communication traffic can be seamlessly switched to the physical path corresponding to the new path forwarding rule for forwarding. Within the preset verification window, the network controller and topology database continuously collect the real-time load and latency of the affected physical links and dynamically update the communication tree conflict graph. When the overall network conflict level decreases or the preset benefit conditions are met, the new path forwarding rule is submitted as the officially effective version, and the corresponding path mapping version number is fixed and updated in the communication tree conflict graph, and the network resources occupied by the old path forwarding rule are released asynchronously. If the benefit conditions are not met within the verification window, or if a higher-priority AI training task is detected to have degraded performance, a rollback command is immediately issued to switch the aggregated communication traffic transiently back to the old path forwarding rule, and the incremental state of the communication tree conflict graph is restored based on the historical link occupancy record corresponding to the old path forwarding rule to complete the rebalancing control closed loop. After completing the path switching or determining not to switch, update the physical path occupancy set, conflict edge weights, cooling status, and predicted baseline of the affected communication tree, and generate persistent conflict event records.

8. An electronic device, characterized in that, The electronic device includes: One or more processors; and A memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the conflict graph-driven incremental joint optimization method as described in any one of claims 1 to 7.

9. A computer-readable medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the conflict graph-driven incremental joint optimization method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the conflict graph-driven incremental joint optimization method according to any one of claims 1 to 7.

Images