A method and apparatus for cross-domain route generation based on the coordination of computing, storage, and transmission resources.
By modeling the routing problem as a sequence learning task and constructing a deep neural network to generate cross-domain routes, the problems of resource coordination and scheduling and intra-domain privacy in cross-domain route generation are solved, and efficient path calculation and resource utilization are achieved.
Patent Information
- Application Number
- CN202311388465.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-24
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-10-24
AI Technical Summary
Existing technologies cannot comprehensively consider the coordinated scheduling of computing, storage, and transmission resources in cross-domain route generation, resulting in resource shortages and high latency, and making it difficult to balance intra-domain privacy and computational complexity.
The routing problem is modeled as a sequence learning task. A deep neural network is constructed, and the input and output sequences are mapped using sparse historical routing results. Cross-domain routes are generated through deep learning, which is applicable to multi-domain scenarios with different levels of openness and takes into account the privacy of network domains.
It greatly reduces path calculation complexity and latency, is suitable for multi-domain scenarios, takes into account the privacy of network domains, and improves resource utilization.
Smart Images

Figure CN117376239B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence and communication technology, and in particular to a method and apparatus for cross-domain route generation based on the coordination of computing, storage and transmission resources. Background Technology
[0002] Driven by emerging businesses, computing and storage resources originally belonging to data centers will inevitably move closer to the network edge. In networks primarily based on edge cloud data centers, current point-to-point communication services in telecommunications networks are gradually evolving into distributed computing tasks, such as federated learning and intelligent video surveillance, which require the collaborative completion of multiple edge nodes and even multiple network domains of different operators. In the process of deploying distributed tasks and generating corresponding routes, comprehensively considering available computing, storage, and transmission resources and achieving coordinated scheduling is an essential step. Furthermore, due to the privacy of information within different operator network domains and their relatively independent management, establishing a route across multiple operator domains is extremely time-consuming, making it difficult to meet the latency-sensitive requirements for service establishment.
[0003] Regarding the resource coordination problem, existing route generation methods based on single-dimensional resource optimization (such as transmission resources) cannot comprehensively consider the availability of various resources. This will lead to various resource shortages caused by improper routing strategies, which in turn introduce waiting delays, resulting in a high total task completion time and making it difficult to effectively improve resource utilization.
[0004] There are two main solutions for cross-domain routing: centralized routing and distributed routing. Centralized routing collects intra-domain information from each network domain through a central controller in the control plane. Therefore, after initiating a cross-domain service path calculation request, it can directly obtain the routing result using the shortest path algorithm based on the network information it possesses. The advantage of this architecture is that the routing calculation can be completed solely by the central controller, resulting in low complexity; however, its disadvantage is that the privacy of information from each network domain is not guaranteed. In the distributed routing scheme, each optical network domain has a controller at the same hierarchical level. Controllers of network domains with related domain sequences need to cooperate through certain signaling mechanisms to jointly complete the cross-domain route calculation. Its advantage is that it can effectively protect the privacy of each network domain; however, its disadvantage is that the domain sequence needs to be determined in advance under opaque network conditions, and complex signaling procedures are required between controllers of network domains with related domain sequences to complete the route calculation, resulting in high complexity. Summary of the Invention
[0005] This application aims to at least partially address one of the technical problems in the related art.
[0006] Therefore, the first objective of this application is to propose a cross-domain route generation method based on the coordination of computing, storage, and transmission resources. This method solves the problems of traditional route calculation schemes lacking resource coordination and scheduling, and the inability to balance intra-domain privacy and computational complexity when calculating cross-domain routes. It can directly generate routing results based on input requests without the need for coordination among multiple network domain control planes, which greatly saves the complexity and latency of the path calculation stage. At the same time, it can flexibly adjust the parameters of the neural network based on the available network state information, making it applicable to multi-domain scenarios with different degrees of openness. Thus, it balances the privacy of network domains while maintaining resource coordination and scheduling.
[0007] The second objective of this application is to propose a cross-domain route generation device based on the coordination of computing, storage, and transmission resources.
[0008] To achieve the above objectives, the first aspect of this application proposes a cross-domain route generation method based on the coordination of computing, storage, and transmission resources, comprising: modeling the routing problem in the network, transforming it into a sequence learning task, and constructing a deep neural network based on the sequence learning task; when a dynamic service request arrives, using the dynamic service request and the current network state as input to the deep neural network, outputting a predicted route, and setting an evaluation index to evaluate the predicted route, obtaining an evaluation result, wherein the evaluation result indicates whether the predicted route can be established in the network; if the predicted route can be established in the network, then the dynamic service is established.
[0009] The cross-domain route generation method based on computation, storage, and transmission resource coordination in this application model route requests, network nodes, and link resources as input sequences and route results as output sequences. By using the sequence-to-sequence method in deep learning, a portion of sparse historical route results is used to find the mapping relationship between the input and output sequences. Furthermore, a trained neural network that has mastered this relationship is used to process cross-domain route requests from more services in the network.
[0010] Optionally, in one embodiment of this application, the routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, network link set, available capacity set of network links, available storage resource set of network nodes, and available computing resource set of network nodes. The network link set includes inter-domain link set and intra-domain link set. The available capacity set of network links includes available capacity set of inter-domain links and available capacity set of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the network node set.
[0011] Optionally, in one embodiment of this application, the routing problem in the network is modeled and transformed into a sequence learning task, including:
[0012] The routing results are segmented to obtain segmented data, which includes source node-destination node pairs and intermediate node sets. When a suitable path cannot be found based on the routing request, the intermediate node set is represented by auxiliary nodes.
[0013] The learning task is constructed by taking the source node-destination node pair, the available capacity set of inter-domain links, the available capacity set of intra-domain links, the available storage resource set of each node, and the available computing resource set of each node as the input sequence, and the intermediate node set as the output sequence.
[0014] Optionally, in one embodiment of this application, constructing a deep neural network based on a sequence-based learning task includes:
[0015] A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network.
[0016] Dynamic service requests include the current source node-destination node pair. The current network state includes the available capacity set of current inter-domain links, the available capacity set of current intra-domain links, the available storage resource set of current each node, and the available computing resource set of current each node. Using the dynamic service requests and the current network state as input to a deep neural network, the output predicted route includes:
[0017] The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form the first vector;
[0018] The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector.
[0019] The first vector, the second vector, the third vector, the fourth vector, and the fifth vector are concatenated to form a concatenated vector, which represents the service request under the current network condition.
[0020] The concatenated vector is repeated m times to form the output route sequence as the predicted route, where m represents the maximum possible number of hops in the route result.
[0021] Optionally, in one embodiment of this application, after constructing the deep neural network, the following steps are further included:
[0022] The deep neural network is trained using a supervised learning model and the Adam optimizer.
[0023] During training, missing input information is replaced with 0, and a masking mechanism is used to cover up the missing information.
[0024] Optionally, in one embodiment of this application, the evaluation metric is feasible routing, and the evaluation of the predicted routing results using feasible routing includes:
[0025] For each path segment in the predicted routing results [r i ,r i+1 ], determine r i and r i+1 Are they in different domains?
[0026] If r i and r i+1 If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation.
[0027] If r i and r i+1 Within the same domain, the calculation starts from r. i to r i+1 If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops.
[0028] After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
[0029] To achieve the above objectives, a second aspect of the present invention proposes a cross-domain route generation device based on the coordination of computing, storage, and transmission resources, comprising a model building module, a prediction module, and an establishment module, wherein...
[0030] The model building module is used to model the routing problem in the network, transform it into a sequence learning task, and build a deep neural network based on the sequence learning task.
[0031] The prediction module is used to take dynamic service requests and the current network status as input to a deep neural network when dynamic service requests arrive, output a predicted route, and set evaluation metrics to evaluate the predicted route and obtain an evaluation result. The evaluation result indicates whether the predicted route can be established in the network.
[0032] A module is established to dynamically establish services when it is determined that a predicted route can be established in the network.
[0033] Optionally, in one embodiment of this application, the routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, network link set, available capacity set of network links, available storage resource set of network nodes, and available computing resource set of network nodes. The network link set includes inter-domain link set and intra-domain link set. The available capacity set of network links includes available capacity set of inter-domain links and available capacity set of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the network node set.
[0034] Optionally, in one embodiment of this application, constructing a deep neural network based on a sequence-based learning task includes:
[0035] A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network.
[0036] Dynamic service requests include the current source node-destination node pair. The current network state includes the available capacity set of current inter-domain links, the available capacity set of current intra-domain links, the available storage resource set of current each node, and the available computing resource set of current each node. Using the dynamic service requests and the current network state as input to a deep neural network, the output predicted route includes:
[0037] The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form the first vector;
[0038] The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector.
[0039] The first vector, the second vector, the third vector, the fourth vector, and the fifth vector are concatenated to form a concatenated vector, which represents the service request under the current network condition.
[0040] The concatenated vector is repeated m times to form the output route sequence as the predicted route, where m represents the maximum possible number of hops in the route result.
[0041] Optionally, in one embodiment of this application, the evaluation metric is feasible routing, and the evaluation of the predicted routing results using feasible routing includes:
[0042] For each path segment in the predicted routing results [r i ,r i+1 ], determine r i and r i+1Are they in different domains?
[0043] If r i and r i+1 If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation.
[0044] If r i and r i+1 Within the same domain, the calculation starts from r. i to r i+1 If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops.
[0045] After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
[0046] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0047] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0048] Figure 1 This is a flowchart illustrating a cross-domain route generation method based on the coordination of computing, storage, and transmission resources, provided in Embodiment 1 of this application.
[0049] Figure 2 This is a schematic diagram of a deep neural network structure according to an embodiment of this application;
[0050] Figure 3 This is a schematic diagram of the closed-loop process of the cross-domain route generation method based on the coordination of computing, storage, and transmission resources according to an embodiment of this application.
[0051] Figure 4 This is a schematic diagram illustrating the operation of the route generation method from the perspective of service arrival / departure events in an embodiment of this application.
[0052] Figure 5 This is a schematic diagram of a cross-domain route generation device based on the collaboration of computing, storage, and transmission resources, provided in an embodiment of this application. Detailed Implementation
[0053] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0054] Addressing the shortcomings of existing traditional routing computation schemes, such as the lack of resource coordination and scheduling, and the inability to balance intra-domain privacy and computational complexity in cross-domain routing computation, this application employs a deep sequence learning approach frequently used in natural language processing. It attempts to model the routing problem as a natural language processing problem, transforming the process from routing request to routing result into a dialogue between input and output statements. Simultaneously, machine learning methods are used to learn routing paths between network nodes from sparse historical network path trajectories, and to complete combinatorial optimization problems on the graph (routing and resource allocation, etc.). Based on this, this application models routing requests, network nodes, and link resources as input sequences, and routing results as output sequences. Through sequence-to-sequence methods in deep learning, it utilizes a subset of sparse historical routing results to find the mapping relationship between the input and output sequences. Finally, it leverages a pre-trained neural network that has mastered this relationship to handle cross-domain routing requests from more services within the network.
[0055] Compared with the prior art, the advantages of this application are:
[0056] (1) A well-trained neural network can directly generate routing results based on input requests without the need for coordination of multiple network domain control planes, which greatly saves the complexity and delay of the path calculation stage.
[0057] (2) By flexibly adjusting the parameters of the neural network based on the available network state information, this invention can be applied to multi-domain scenarios with different degrees of openness, thereby taking into account the privacy of the network domain on the basis of resource collaborative scheduling.
[0058] The following describes, with reference to the accompanying drawings, a method and apparatus for cross-domain route generation based on computational, storage, and transmission resource collaboration, according to embodiments of this application.
[0059] Figure 1 This is a flowchart illustrating a cross-domain route generation method based on the collaboration of computing, storage, and transmission resources, as provided in Embodiment 1 of this application.
[0060] like Figure 1 As shown, the cross-domain route generation method based on the coordination of computing, storage, and transmission resources includes the following steps:
[0061] Step 101: Model the routing problem in the network, transform it into a sequence learning task, and construct a deep neural network based on the sequence learning task;
[0062] Step 102: When a dynamic service request arrives, the dynamic service request and the current network state are used as inputs to the deep neural network to output a predicted route. An evaluation index is set to evaluate the predicted route and obtain an evaluation result. The evaluation result indicates whether the predicted route can be established in the network.
[0063] Step 103: If the predicted route can be established in the network, then the dynamic service will be established.
[0064] The cross-domain route generation method based on computation, storage, and transmission resource coordination in this application model route requests, network nodes, and link resources as input sequences and route results as output sequences. By using the sequence-to-sequence method in deep learning, a portion of sparse historical route results is used to find the mapping relationship between the input and output sequences. Furthermore, a trained neural network that has mastered this relationship is used to process cross-domain route requests from more services in the network.
[0065] Optionally, in one embodiment of this application, the routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, network link set, available capacity set of network links, available storage resource set of network nodes, and available computing resource set of network nodes. The network link set includes inter-domain link set and intra-domain link set. The available capacity set of network links includes available capacity set of inter-domain links and available capacity set of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the network node set.
[0066] Optionally, in one embodiment of this application, the routing problem in the network is modeled and transformed into a sequence learning task, including:
[0067] The routing results are segmented to obtain segmented data, which includes source node-destination node pairs and intermediate node sets. When a suitable path cannot be found based on the routing request, the intermediate node set is represented by auxiliary nodes.
[0068] The learning task is constructed by taking the source node-destination node pair, the available capacity set of inter-domain links, the available capacity set of intra-domain links, the available storage resource set of each node, and the available computing resource set of each node as the input sequence, and the intermediate node set as the output sequence.
[0069] Optionally, in one embodiment of this application, constructing a deep neural network based on a sequence-based learning task includes:
[0070] A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network.
[0071] Dynamic service requests include the current source node-destination node pair. The current network state includes the available capacity set of current inter-domain links, the available capacity set of current intra-domain links, the available storage resource set of current each node, and the available computing resource set of current each node. Using the dynamic service requests and the current network state as input to a deep neural network, the output predicted route includes:
[0072] The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form the first vector;
[0073] The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector.
[0074] The first vector, the second vector, the third vector, the fourth vector, and the fifth vector are concatenated to form a concatenated vector, which represents the service request under the current network condition.
[0075] The concatenated vector is repeated m times to form the output route sequence as the predicted route, where m represents the maximum possible number of hops in the route result.
[0076] Optionally, in one embodiment of this application, after constructing the deep neural network, the following steps are further included:
[0077] The deep neural network is trained using a supervised learning model and the Adam optimizer.
[0078] During training, missing input information is replaced with 0, and a masking mechanism is used to cover up the missing information.
[0079] Optionally, in one embodiment of this application, the evaluation metric is feasible routing, and the evaluation of the predicted routing results using feasible routing includes:
[0080] For each path segment in the predicted routing results [r i ,r i+1 ], determine r i and r i+1 Are they in different domains?
[0081] If r i and r i+1 If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation.
[0082] If r iand r i+1 Within the same domain, the calculation starts from r. i to r i+1 If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops.
[0083] After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
[0084] The following describes in detail the cross-domain route generation method based on the collaboration of computing, storage, and transmission resources according to embodiments of this application.
[0085] (a) First, the routing problem in the network is modeled and transformed into a sequence-to-sequence learning task that is more suitable for deep learning.
[0086] Consider a network topology G = (N, E), where N is the set of network nodes and E is the set of network links. Let E1 be the set of inter-domain links and E2 be the set of intra-domain links; therefore, E1 ∪ E2 = E. The set of available capacities for inter-domain links is L1, where L1 = {capacity(l i )},l i ∈E1; The set of available capacities for links within the domain is L2, L2={capacity(l i )},l i ∈E2; The set of available storage resources for each node is S, S={storage(r i )}, r i ∈N; the set of available computing resources for each node is C, C = {cpu(r i )}, r i ∈N. The routing result can be represented as a subset R of the set of network nodes. The routing result R is divided into two parts: source node-destination node pairs R1 and intermediate node set R2. R1 = [s, d], s, d ∈ N, where s and d represent the source node and destination node of the routing request, respectively. R1 ∪ R2 = R, Subsequently, R1, L1, L2, S, C are used as inputs to the neural network, and R2 is used as the output of the neural network, thus transforming a routing computation problem into a sequence-to-sequence learning problem.
[0087] (ii) Then, construct as follows Figure 2 The deep neural network structure shown is used to implement cross-domain routing functionality.
[0088] The input to this neural network consists of two parts: the service request sequence and the current network computing, storage, and transmission resource status. The output is the routing result. For the input information of the neural network, R1 and L1, L2, S, C are integer sequences. For a Transformer-based neural network, each integer element needs to be encoded into a binary vector for processing. Therefore, one-hot encoding is used. Generally, if an integer variable 'a' is to be one-hot encoded, and the range of 'a' is known to be A, and the variable to be encoded is a = i, then the encoded result is a vector of length |A|, where only the i-th element has a value of 1, and the remaining elements have values of 0. Through the above operation, the service request sequence R1 is transformed into a binary matrix of size 2 × |N|.
[0089]
[0090] The encoding process for the routing result sequence R2 is similar. Specifically, if a routing request cannot find a suitable path in the network, i.e., the service request is blocked, then an auxiliary node x will be used to represent the routing result [s,x,d]. Due to the addition of the auxiliary node, The size should be increased to 2×|N+1|. Furthermore, The data are sequentially fed into an LSTM (Bi-LSTM) and a pooling layer to extract key information and form a vector S. The sequences of inter-domain link capacity, storage resources, computational resources, and intra-domain link capacity are respectively passed through a fully-connected layer (FC) with ReLU activation function to form vectors V1, V2, V3, and V4.
[0091] After encoding preprocessing, S and V1, V2, V3, V4 are concatenated into I to represent the service request in the current network state. To obtain a vector form R2, I is repeated m times to form K. m This is used to generate an output route sequence of sufficient length, where m represents the maximum possible number of hops in the output route result.
[0092]
[0093] K m The data is fed into a Transformer for processing. The output is then passed through a pooling layer, followed by a fully connected layer with ReLU activation to adjust the size of the output matrix. Finally, another fully connected layer with softmax activation produces a probability output matrix P of size |R²|×|N|. Overfitting is avoided in the neural network using a dropout mechanism with a dropout ratio of 0.2.
[0094] The output of the neural network is a probability output matrix P of size |R²|×|N|, where each element represents a probability value between 0 and 1. Since the initial input data to the neural network undergoes one-hot encoding, a corresponding decoding operation is required at the output layer. For each vector of size |N| in the output matrix R², the index of the element with the highest probability value is selected as the decoding result, converting the vector into an integer value (similar to the reverse process of one-hot encoding). In this way, an output sequence of size |R²| is obtained, which is the desired routing result.
[0095] The neural network is trained using supervised learning. In supervised learning, for a specific business request input (R1, L1, L2, S, C), the training objective path for the problem is R2. The Adam optimizer is used to minimize the cross-entropy function θ of the problem during training.
[0096]
[0097] Here, p + p0 is an element in the output matrix P, and p0 is an element of the matrix formed after one-hot encoding of the training target path R2.
[0098] During training, to maintain network domain privacy, only publicly available information from each network domain and historical business information and routing trajectories between nodes collected by the network management system can be utilized. Therefore, missing or unavailable network status information is inevitable. To address this, unavailable input information is replaced with 0, and the Transformer's masking mechanism is used to mask the missing or unavailable information, thus ignoring this information during training. Based on this strategy, this application can be applied to multi-domain scenarios with varying degrees of openness.
[0099] (III) After training the neural network, evaluation metrics need to be set to assess the prediction results and determine the prediction performance. It is important to note that in supervised learning, the method of determining the results does not affect the training process. Here, two evaluation criteria for prediction results are introduced: consistent routing and feasible routing.
[0100] (1) Consistent routing: The route given by the neural network is completely consistent with the target route in the test dataset.
[0101] (2) Feasible route: The route given by the neural network is inconsistent with the target route in the test dataset, but the route can be used in the network.
[0102] The method for checking consistent routes is relatively simple; a direct comparison is sufficient. The following describes the method for checking feasible route results.
[0103] The input to the routing result checking method is: the predicted path R0 = [r] given by the neural network, r ∈ N; the network topology G = (N, E); the bandwidth b, computing resources d, and storage resources e required by the service request. The output is: R0, whose value corresponds to whether it can be accessed by the network (1) or not (0), and the service completion time t.
[0104] For each path [r] on R0 i ,r i+1 ],if:
[0105] (1)r i and r i+1 If the path and node resources meet the business requirements, modify the relevant resource status and then continue to check the next path segment; otherwise, output [0, inf] and the checking algorithm ends.
[0106] (2)r i and r i+1 In the same domain: calculation from r i to r i+1 If the calculated intra-domain path can be established, modify the relevant resource status and then continue to check the next path segment; otherwise, output [0, inf] and the checking algorithm ends.
[0107] (3) If all nodes and paths meet the business requirements, calculate the business completion time t and output [1,t], then check that the algorithm has ended.
[0108] (iv) Finally, based on the trained neural network, real-time cross-domain routing calculation is achieved according to dynamic business needs and network resource status.
[0109] Figure 3 and Figure 4 This paper details a cross-domain route generation method based on the coordination of computing, storage, and transmission resources. When a dynamic service request arrives, a predicted route is directly provided based on the publicly available capacity information of inter-domain links in the current network. This predicted route is then tested. If the test results indicate that the predicted route can be successfully established in the network, the service is established, and the time required for service completion is given. Simultaneously, the capacity of various network resources is adjusted accordingly. Otherwise, the route is discarded. When the next service arrives, the above process continues. If a service path leaves (is dismantled) after operating in the network for a period of time, the resource status of the inter-domain links in the network is directly changed.
[0110] To implement the above embodiments, this application also proposes a cross-domain route generation device based on the coordination of computing, storage, and transmission resources.
[0111] Figure 5 This is a schematic diagram of a cross-domain route generation device based on the collaboration of computing, storage, and transmission resources, provided in an embodiment of this application.
[0112] like Figure 5 As shown, the cross-domain route generation device based on the coordination of computing, storage, and transmission resources includes a model building module, a prediction module, and an establishment module, wherein...
[0113] The model building module is used to model the routing problem in the network, transform it into a sequence learning task, and build a deep neural network based on the sequence learning task.
[0114] The prediction module is used to take dynamic service requests and the current network status as input to a deep neural network when dynamic service requests arrive, output a predicted route, and set evaluation metrics to evaluate the predicted route and obtain an evaluation result. The evaluation result indicates whether the predicted route can be established in the network.
[0115] A module is established to dynamically establish services when it is determined that a predicted route can be established in the network.
[0116] Optionally, in one embodiment of this application, the routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, network link set, available capacity set of network links, available storage resource set of network nodes, and available computing resource set of network nodes. The network link set includes inter-domain link set and intra-domain link set. The available capacity set of network links includes available capacity set of inter-domain links and available capacity set of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the network node set.
[0117] Optionally, in one embodiment of this application, constructing a deep neural network based on a sequence-based learning task includes:
[0118] A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network.
[0119] Dynamic service requests include the current source node-destination node pair. The current network state includes the available capacity set of current inter-domain links, the available capacity set of current intra-domain links, the available storage resource set of current each node, and the available computing resource set of current each node. Using the dynamic service requests and the current network state as input to a deep neural network, the output predicted route includes:
[0120] The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form the first vector;
[0121] The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector.
[0122] The first vector, the second vector, the third vector, the fourth vector, and the fifth vector are concatenated to form a concatenated vector, which represents the service request under the current network condition.
[0123] The concatenated vector is repeated m times to form the output route sequence as the predicted route, where m represents the maximum possible number of hops in the route result.
[0124] Optionally, in one embodiment of this application, the evaluation metric is feasible routing, and the evaluation of the predicted routing results using feasible routing includes:
[0125] For each path segment in the predicted routing results [r i ,r i+1 ], determine r i and r i+1 Are they in different domains?
[0126] If r i and r i+1 If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation.
[0127] If r i and r i+1 Within the same domain, the calculation starts from r. i to r i+1 If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops.
[0128] After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
[0129] It should be noted that the foregoing explanation of the cross-domain route generation method based on the collaboration of computing, storage, and transmission resources also applies to the cross-domain route generation device based on the collaboration of computing, storage, and transmission resources in this embodiment, and will not be repeated here.
[0130] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0131] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0132] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.
[0133] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.
[0134] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0135] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0136] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0137] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
Claims
1. A cross-domain routing generation method based on computing storage transmission resource coordination, characterized in that, Includes the following steps: The routing problem in the network is modeled and transformed into a sequence-to-sequence learning task. A deep neural network is constructed based on the sequence-to-sequence learning task, wherein the source node-target node pair and the network state are used as the input sequence and the intermediate node set is used as the output sequence to construct the sequence-to-sequence learning task. Training the deep neural network includes: The deep neural network is trained using a supervised learning model and the Adam optimizer. During training, missing input information is replaced with 0, and a masking mechanism is used to cover up the missing information. When a dynamic service request arrives, the dynamic service request and the current network state are used as inputs to the deep neural network, a predicted route is output, and an evaluation index is set to evaluate the predicted route to obtain an evaluation result, wherein the evaluation result indicates whether the predicted route can be established in the network. If the predicted route can establish a path in the network, then the dynamic service will be established.
2. The method as described in claim 1, characterized in that, The routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, set of network links, set of available capacity of network links, set of available storage resources of network nodes, and set of available computing resources of network nodes. The set of network links includes set of inter-domain links and set of intra-domain links. The set of available capacity of network links includes set of available capacity of inter-domain links and set of available capacity of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the set of network nodes.
3. The method as described in claim 2, characterized in that, The process of modeling the routing problem in the network and transforming it into a sequence-to-sequence learning task includes: The routing results are segmented to obtain segmented data, wherein the segmented data includes source node-destination node pairs and intermediate node sets. When a suitable path cannot be found based on the routing request, the intermediate node set is represented by auxiliary nodes. Using the source node-destination node pair, the available capacity set of inter-domain links, the available capacity set of intra-domain links, the available storage resource set of each node, and the available computing resource set of each node as input sequences, and the intermediate node set as output sequences, a sequence-to-sequence learning task is constructed.
4. The method as described in claim 3, characterized in that, The construction of a deep neural network based on the sequence-to-sequence learning task includes: A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network. The dynamic service request includes the current source node-destination node pair, and the current network state includes the available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node. The step of using the dynamic service request and the current network state as input to the deep neural network and outputting a predicted route includes: The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form a first vector; The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector. The first vector is concatenated with the second, third, fourth, and fifth vectors to form a concatenated vector, wherein the concatenated vector represents a service request under the current network condition; The concatenated vector is repeated m times to form an output routing sequence, which is then used as the predicted route. This indicates the maximum possible number of hops for the routing result.
5. The method as described in claim 1, characterized in that, The evaluation metric is feasible routing. The predicted routing results are evaluated using feasible routing, including: For each path segment in the predicted routing results ,judge and Are they in different domains? like and If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation. like and Within the same domain, computation from arrive If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops. After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
6. A cross-domain route generation device based on the coordination of computing, storage, and transmission resources, characterized in that, It includes a model building module, a prediction module, and a model creation module, among which, The model building module is used to model the routing problem in the network, transform it into a sequence-to-sequence learning task, and build a deep neural network based on the sequence-to-sequence learning task. The sequence-to-sequence learning task is constructed by taking the source node-target node pair and the network state as the input sequence and the intermediate node set as the output sequence. Training the deep neural network includes: The deep neural network is trained using a supervised learning model and the Adam optimizer. During training, missing input information is replaced with 0, and a masking mechanism is used to cover up the missing information. The prediction module is used to take the dynamic service request and the current network state as input to the deep neural network when the dynamic service request arrives, output a predicted route, and set an evaluation index to evaluate the predicted route and obtain an evaluation result, wherein the evaluation result indicates whether the predicted route can be established in the network. The establishment module is used to establish dynamic services when it is determined that the predicted route can be established in the network.
7. The apparatus as claimed in claim 6, characterized in that, The routing problem is a problem of solving routing results based on routing requests and network states. The network states include: network topology, set of network links, set of available capacity of network links, set of available storage resources of network nodes, and set of available computing resources of network nodes. The set of network links includes set of inter-domain links and set of intra-domain links. The set of available capacity of network links includes set of available capacity of inter-domain links and set of available capacity of intra-domain links. The routing result is the path found in the network based on the routing request, and the path is a subset of the set of network nodes.
8. The apparatus as claimed in claim 7, characterized in that, The construction of a deep neural network based on the sequence-to-sequence learning task includes: A deep neural network is constructed by using the input sequence as the input of the neural network and the output sequence as the output of the neural network. The dynamic service request includes the current source node-destination node pair, and the current network state includes the available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node. The step of using the dynamic service request and the current network state as input to the deep neural network and outputting a predicted route includes: The current source node-destination node pair is processed through a bidirectional long short-term memory network and a pooling layer to extract key information and form a first vector; The available capacity set of the current inter-domain links, the available capacity set of the current intra-domain links, the available storage resource set of each current node, and the available computing resource set of each current node are respectively passed through a fully connected network with ReLU as the activation function to form the second vector, the third vector, the fourth vector, and the fifth vector. The first vector is concatenated with the second, third, fourth, and fifth vectors to form a concatenated vector, wherein the concatenated vector represents a service request under the current network condition; The concatenated vector is repeated m times to form an output routing sequence, which is then used as the predicted route. This indicates the maximum possible number of hops for the routing result.
9. The apparatus as claimed in claim 6, characterized in that, The evaluation metric is feasible routing. The predicted routing results are evaluated using feasible routing, including: For each path segment in the predicted routing results ,judge and Are they in different domains? like and If the path and node resources are in different domains, determine whether they meet the business requirements. If they do, determine whether the network can be accessed. If they do not, determine whether the network can be accessed and stop the evaluation. like and Within the same domain, computation from arrive If the calculated intra-domain path can be established, it is determined that it can be accessed by the network; otherwise, it is determined that it cannot be accessed by the network, and the evaluation stops. After the evaluation is completed, if each path segment is determined to be accessible to the network, then the evaluation result is determined to be that the predicted route can be established in the network.
Citation Information
Patent Citations
Space-air-ground integrated network resource allocation method based on deep reinforcement learning
CN114599043A
