Data center and cloud computing switching method and switch

Abstract

The invention relates to the technical field of data switches, and discloses a data center and cloud computing switching method and a switch, and the method comprises the steps: analyzing east-west data traffic to obtain traffic feature information, and deploying a Spine layer controller, a Leaf layer controller and a ToR layer controller of a data center network; performing virtual output queue mapping based on a tenant identification field in the traffic feature information to obtain a bandwidth quota scheme; in the Spine layer controller, east-west flow prediction is carried out in combination with a bandwidth quota scheme, and a global path planning scheme is obtained; a scheduling instruction is created in the Leaf layer controller according to the global path planning scheme, the scheduling instruction is sent to the ToR layer controller for synchronization, and a cooperative exchange control result is obtained, the problems of single-point faults and expansibility of a centralized controller are solved, state synchronization of a distributed control system is ensured, and global convergence and optimality are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data exchange, in particular to a data center and cloud computing exchange method and switch. BACKGROUND

[0002] Traditional data center switches use fixed ECMP path selection and FIFO queue scheduling mechanism, which cannot perceive the business characteristics and quality of service requirements of different tenants, leading to uneven network resource allocation, serious interference between tenants, and increasingly prominent problems such as difficulty in guaranteeing service quality. In the mixed traffic scenario, when elephant flow and mouse flow are mixed in transmission, the traditional switch lacks effective traffic identification and classification mechanism, which is easy to cause network congestion and delay jitter. At the same time, the traditional static path planning method lacks the ability to predict future traffic trends, and cannot dynamically adjust the path selection strategy according to the network state, leading to uneven network load and low resource utilization. SUMMARY

[0003] The present application provides a data center and cloud computing exchange method and switch, which solves the single point failure and scalability problem of centralized controller, ensures the state synchronization of distributed control system, and guarantees global convergence and optimality.

[0004] In a first aspect, the present application provides a data center and cloud computing exchange method, which comprises: analyzing east-west data traffic to obtain traffic characteristic information, and deploying Spine layer controllers, Leaf layer controllers and ToR layer controllers of the data center network at the same time; mapping virtual output queues based on the tenant identification field in the traffic characteristic information to obtain a bandwidth quota scheme; in the Spine layer controller, predicting east-west traffic in combination with the bandwidth quota scheme to obtain a global path planning scheme; creating a scheduling instruction in the Leaf layer controller according to the global path planning scheme, and sending the scheduling instruction to the ToR layer controller for synchronization to obtain a collaborative exchange control result.

[0005] In combination with the first aspect, in a first implementation manner of the first aspect of the present application, the east-west data traffic is analyzed to obtain traffic characteristic information, and Spine layer controllers, Leaf layer controllers and ToR layer controllers of the data center network are deployed at the same time, which comprises: extracting the packet header field of the received east-west data traffic through the hardware parser of the ASIC chip to obtain the tenant identification field in the five-tuple information including the source IP address, the destination IP address, the source port number, the destination port number, the protocol type and the VLAN tag. performing CRC32 hash function calculation based on the five-tuple information and embedding metadata fields in the data packet header by INT technology to obtain identified flow data; performing double-threshold value determination on the flow duration timestamp and the cumulative byte count in the identified flow data to obtain a flow type classification result; performing level mapping on the flow type classification result and the tenant identification field and establishing active flow table records in the on-chip SRAM of the ASIC chip to obtain flow characteristic information; deploying Spine layer controllers, Leaf layer controllers and ToR layer controllers of a data center network according to the flow characteristic information.

[0006] In a second implementation manner of the first aspect, the deploying Spine layer controllers, Leaf layer controllers and ToR layer controllers of a data center network according to the flow characteristic information comprises: performing control responsibility division on the data center network based on the flow characteristic information to obtain Spine layer control responsibilities, Leaf layer control responsibilities and ToR layer control responsibilities; deploying a global coordinator instance in a Spine switch, deploying a domain controller instance in a Leaf switch and integrating a local decision maker instance in a ToR switch control plane according to the Spine layer control responsibilities, the Leaf layer control responsibilities and the ToR layer control responsibilities; connecting the global coordinator instance, the domain controller instance and the local decision maker instance and mapping control objects to different controllers to obtain Spine layer controllers, Leaf layer controllers and ToR layer controllers.

[0007] In a third implementation manner of the first aspect, the performing virtual output queue mapping based on the tenant identification field in the flow characteristic information to obtain a bandwidth quota scheme comprises: performing level classification on gold medal tenants, silver medal tenants, bronze medal tenants and common tenants according to the tenant identification field in the flow characteristic information and configuring mapping relationships of virtual output queues in the ASIC chip; performing queue allocation on data flows of each tenant level based on the mapping relationships and setting weighted scheduling parameters for each virtual output queue; performing bandwidth proportion calculation and sending opportunity allocation based on the weighted scheduling parameters to obtain a basic bandwidth allocation result; performing dynamic bandwidth borrowing according to the basic bandwidth allocation result and monitoring utilization rates of the virtual output queues, and when detecting that a utilization rate of a virtual output queue is lower than a utilization rate threshold value, performing bandwidth reallocation process to obtain a bandwidth quota scheme.

[0008] With reference to the first aspect, in a fourth implementation form of the first aspect of the present application, in the Spine layer controller, the east-west traffic prediction is combined with the bandwidth quota scheme to obtain a global path planning scheme, comprising: In the Spine layer controller, the bandwidth quota scheme is input into a graph neural network model, and a graph data structure is constructed, wherein the vertices of the graph data structure include Spine switches, Leaf switches, ToR switch nodes and their node feature vectors, and the edges include inter-switch physical links and their link feature vectors; In a three-layer graph neural network layer in the graph neural network model, feature aggregation and propagation calculation are performed on the graph data structure to obtain east-west traffic prediction data; Based on the east-west traffic prediction data, path planning is performed to obtain a global path planning scheme.

[0009] With reference to the first aspect, in a fifth implementation form of the first aspect of the present application, the path planning based on the east-west traffic prediction data to obtain a global path planning scheme, comprising: Based on the east-west traffic prediction data, a target function is set, which includes minimizing the total network delay, maximizing the load balancing degree, and maximizing the tenant SLA satisfaction rate; Link capacity constraints, flow conservation constraints and tenant SLA constraints are taken as constraint conditions, and a candidate ECMP path set of each node pair in the graph data structure is defined as a decision variable; According to the target function, the constraint conditions and the decision variables, an initial path planning result is solved; According to the initial path planning result, a delay shortest path is allocated to gold medal tenants, a load balanced path is allocated to silver medal tenants and bronze medal tenants under the SLA constraint, and a remaining path is allocated to ordinary tenants, to obtain a global path planning scheme.

[0010] With reference to the first aspect, in a sixth implementation form of the first aspect of the present application, in the Leaf layer controller, the global path planning scheme is created according to the global path planning scheme, and the scheduling instruction is sent to the ToR layer controller for synchronization to obtain a collaborative exchange control result, comprising: In the Leaf layer controller, the global path planning scheme is combined with the link utilization rate vector, the queue length vector and the tenant SLA margin vector of the current network environment to construct a state space of a Markov decision process; Based on the state space, the ECMP path set and its weight distribution are defined as an action space of a reinforcement learning algorithm, and a reward function calculation is performed on the execution result of each candidate path weight action in the action space to obtain a reward value; input the reward value into a deep Q network for path weight optimization to obtain a scheduling instruction; send the scheduling instruction to a ToR layer controller and synchronize through a Raft protocol to obtain a collaborative exchange control result.

[0011] In a seventh implementation manner of the first aspect, the inputting the reward value into the deep Q network for path weight optimization to obtain the scheduling instruction comprises: inputting the reward value into the deep Q network to calculate a Q value prediction result of each candidate path weight action under a current network environment; generating a Q value function weight parameter based on the Q value prediction result through an experience replay mechanism; selecting an optimal path weight of the current network environment from the candidate path weight actions according to an e-greedy strategy based on the Q value function weight parameter, and converting the optimal path weight into a corresponding scheduling instruction.

[0012] In an eighth implementation manner of the first aspect, the sending the scheduling instruction to the ToR layer controller and synchronizing through the Raft protocol to obtain the collaborative exchange control result comprises: encapsulating the scheduling instruction into a Raft protocol transmission message through the Raft protocol; confirming a first node and a second node based on the Raft protocol transmission message, sending the scheduling instruction from the first node to the second node, triggering re-election when the second node does not receive a heartbeat within a preset period, and obtaining control state synchronization information; when detecting that a deviation between a local version number and a global version number in the control state synchronization information exceeds a preset threshold, obtaining a latest control state from the first node and updating a local state, updating a hardware flow table cache and a path weight configuration in the ToR layer controller, and obtaining the collaborative exchange control result.

[0013] In a second aspect, the present application provides a data center and cloud computing switch, which is used to execute a data center and cloud computing exchange method.

[0014] The technical scheme provided by the application realizes multi-tenant east-west traffic hardware identification by deploying a P4 programmable data plane in a programmable ASIC chip built in a data center and a cloud computing switch, and significantly reduces identification delay compared with a traditional software processing mode. The three-layer decentralized control architecture effectively solves the single-point failure and scalability problems of the centralized controller, and realizes the coordination of the second-level global strategy and the microsecond-level packet forwarding through layered decision-making. The virtual output queue mapping based on tenant levels and the weighted DRR scheduling algorithm realize differentiated quality of service isolation and avoid interference between tenants. The graph neural network traffic prediction model provides active path planning capability by modeling network topology relationships, and the reinforcement learning algorithm realizes adaptive optimization of path weights. The hardware flow table cache and path rerouting mechanism of the ToR switch guarantee real-time response capability. The Raft protocol and version vector mechanism are used to ensure the state synchronization of the distributed control system, and guarantee the global convergence and optimality.

[0015] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the description, claims and drawings.

[0016] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 An embodiment schematic diagram of the data center and cloud computing exchange method in the embodiments of the present application. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions of the present application will be described in detail below in combination with the drawings. Obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0019] The terms "include" and "have" and any variations thereof mentioned in the embodiments of the present application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can also include other steps or units not listed, or can also include other steps or units inherent to the process, method, product or device.

[0020] For the convenience of understanding the present embodiment, first, a data center and cloud computing exchange method disclosed by the present embodiment is introduced in detail. As shown in the figure Figure 1 The data center and cloud computing exchange method includes the following steps: 101, analyzing east-west data traffic to obtain traffic characteristic information, while deploying Spine layer controllers, Leaf layer controllers and ToR layer controllers of the data center network; Specifically, a programmable ASIC chip is integrated in the data center and cloud computing switch, and a special hardware parser program is run in the ASIC chip, so that the hardware parser can analyze and process data packets at nanosecond level speed as soon as the data flows into the switch port. When the data packet enters, the hardware parser extracts the source IP address, destination IP address, source port number, destination port number and protocol type from the packet header to form five-tuple information, and parses the tenant identification information carried in the VLAN tag field to realize the preliminary binding of the data packet and the tenant level. Based on the five-tuple information, the CRC32 hash function is calculated to generate a unique flow identifier, and the key metadata fields including the flow identifier, cumulative byte count, flow duration timestamp, path hop count and path identifier are embedded into the data packet header through the INT technology to realize the real-time state coding of each flow. The flow duration timestamp and cumulative byte count extracted from each identified traffic data are subjected to a double threshold judgment strategy: only when the duration of the flow exceeds the preset millisecond level time threshold and the cumulative byte count exceeds the preset data volume threshold, the data flow will be classified as a "elephant flow" with high load, while other types of burst flow, long tail flow or short time flow are classified according to different combinations of conditions not met, in order to realize the classification of flow type. The flow type classification result is mapped with the tenant identification field, and the traffic is allocated to the corresponding priority category according to the preset tenant level system, and the identifier, type classification result, tenant level field of each flow and its corresponding statistical state are written into the active flow table in the on-chip SRAM of the ASIC chip, to establish a real-time updated traffic characteristic database. According to the structured traffic characteristic information summarized in the active flow table, the initial data basis of the data center network control system is constructed, and the Spine layer global coordination controller, Leaf layer intra-domain path selection controller and ToR layer local rapid scheduling controller are deployed accordingly.

[0021] 102, mapping the virtual output queue based on the tenant identification field in the traffic characteristic information to obtain a bandwidth quota scheme; Specifically, all tenant data flows are classified into four levels, namely gold medal tenant, silver medal tenant, bronze medal tenant and ordinary tenant, according to the tenant identifier field in the flow characteristic information, and 8 virtual output queues are configured inside the switch chip, and each two queues are bound to a tenant level, thereby establishing a mapping relationship between the tenant level and the virtual output queue. According to the mapping rule, each data packet flowing through the traffic generated by different tenant levels is allocated to the corresponding output queue, ensuring that high-priority tenant traffic will not compete with low-priority tenants for the same bandwidth resources, while setting appropriate weighted scheduling parameters for each virtual output queue, including scheduling weight and scheduling quantum value, for example, configuring higher scheduling weight and larger quantum value for gold medal tenants. Based on the weighted scheduling parameters, the weighted round-robin scheduling algorithm is executed in the scheduling module of the switch chip, and in each scheduling cycle, the sending opportunity is allocated to different tenant queues, so as to calculate the basic bandwidth allocation result according to the predetermined proportion, and to clearly obtain the actual bandwidth proportion that each tenant should obtain in the ideal case without resource conflict. According to the basic bandwidth allocation result, dynamic bandwidth borrowing is performed, and the current utilization rate of each virtual output queue is continuously monitored in real time during the running process, and its occupancy state is updated at a period of 1 millisecond; when it is detected that the utilization rate of a queue of a certain level is continuously lower than the set utilization rate threshold (such as 50%), the bandwidth reallocation process is triggered, allowing high-priority queues to temporarily borrow the bandwidth share of low-priority queues that is not used, but low-priority queues cannot borrow upwards. The reallocation result after bandwidth borrowing will update the effective scheduling weight of the current virtual output queue, and dynamically generate a bandwidth quota scheme that meets the actual running state.

[0022] 103. In the Spine layer controller, the east-west traffic prediction is combined with the bandwidth quota scheme to obtain a global path planning scheme; Specifically, the bandwidth quota scheme uploaded by the lower layer controller is received in the Spine layer controller, and the bandwidth quota scheme is input into the graph neural network model, while the graph data structure is constructed based on the current network state. The graph data structure is abstractly modeled based on the topology of the entire data center network, where the vertex set is composed of exchange nodes at each level, including Spine switches, Leaf switches and ToR switches, and each vertex is attached with a node feature vector, including the link utilization rate, average queue length, tenant level traffic distribution, switching capacity parameters and cache state of the current switch; the edge set represents the physical link connection between switches, and each edge is configured with a link feature vector containing link bandwidth capacity, delay index, packet loss rate, link load ratio and congestion level. The graph neural network model starts its internal neural network calculation process composed of three layers of graph convolution units, the first layer of graph convolution performs initial dimension lifting on the node features and completes local information extraction, the second layer of graph convolution performs feature aggregation and propagation across nodes to realize context fusion of state perception between different levels of exchange nodes, and the third layer of graph convolution further integrates and predicts the aggregation results of the overall graph structure to generate the traffic change trend of each link and each node in the future multiple time slices, forming structured east-west traffic prediction data. Based on the east-west traffic prediction data, path planning is performed, and a multi-objective optimization strategy is introduced in the path planning process, considering multiple objective functions such as minimizing link delay, balancing link load and guaranteeing tenant SLA, setting constraint conditions covering link capacity, tenant service level path priority and path hop limit, and using genetic algorithm as an optimization solver to iteratively generate optimal path combinations that meet the conditions to obtain a global path planning scheme.

[0023] 104、In the Leaf layer controller, create scheduling instructions according to the global path planning scheme, and send the scheduling instructions to the ToR layer controller for synchronization to obtain collaborative exchange control results.

[0024] Specifically, the global path planning scheme from the Spine layer controller is received and parsed in the Leaf layer controller, and the state space required for building the reinforcement learning model is constructed in combination with the current local network state data. The state space integrates three types of dynamic indicators: the link utilization rate vector maintained by the current Leaf exchange node, reflecting the bandwidth usage of each physical link; the queue length vector of the local switch, indicating the buffer occupancy level of different virtual output queues; and the tenant service level SLA margin vector, recording the difference between the current path delay, throughput, and service quality requirements of each tenant and the target SLA. The above three types of indicators are combined with the path combination and scheduling weights provided in the global path planning to form the state description of the Markov decision process. After the state space modeling is completed, the action space of the reinforcement learning algorithm is built based on the ECMP path set contained in the path planning scheme. Each action in the action space represents a path weight distribution strategy, such as equal weight distribution, bandwidth proportional weighting, delay weighting, congestion reverse weighting, etc. The Leaf controller selects the path combination scheme in the action space for execution simulation in each scheduling period, and evaluates the execution effect of each candidate action. The evaluation method relies on a pre-set multi-objective reward function, which integrates the minimization of packet queuing delay, the reduction of packet loss rate, the improvement of link load balancing, and the improvement of tenant SLA achievement rate, etc. to convert the four types of performance indicators into quantitative reward values. The reward values corresponding to different actions are input into the deep Q network model, the value function approximation of the path selection action is performed through the reinforcement learning strategy, the state-action value function is modeled using a deep neural network, and the network weights are iteratively optimized using the experience replay and target network mechanism. The optimal path weight configuration in the path combination corresponding to the action with the maximum Q value is selected as the scheduling result, forming a formatted scheduling instruction containing tenant identifier, path ID, path weight, execution timestamp, and update policy version number, etc. fields. The Leaf controller sends the scheduling instruction to the subordinate ToR controller through the established southbound control channel, and uses the Raft distributed consistency protocol to synchronize and confirm the instruction state, ensuring state synchronization and instruction consistency among multiple ToR controller instances. When the ToR controller receives the instruction, it writes the corresponding path weight configuration into the local hardware flow table, and at the same time feeds back the scheduling execution state to the Leaf controller, forming a top-down control link closed loop, realizing the fusion of path planning results and local state adaptive scheduling strategy in a multi-tenant environment, and forming a collaborative exchange control result through consistency synchronization.

[0025] In a specific embodiment, the process of performing step 101 can specifically include the following steps: The received east-west data traffic is subjected to packet header field extraction by a hardware parser of an ASIC chip to obtain five-tuple information including a source IP address, a destination IP address, a source port number, a destination port number, a protocol type, and a tenant identification field in a VLAN tag; CRC32 hash function calculation is performed based on the five-tuple information, and metadata fields are embedded in the packet header by INT technology to obtain identified traffic data; Double-threshold determination is performed on the flow duration timestamp and the cumulative byte count in the identified traffic data to obtain a flow type classification result; The flow type classification result and the tenant identification field are subjected to hierarchical mapping, and an active flow table record is established in an SRAM on the ASIC chip to obtain traffic characteristic information; A Spine layer controller, a Leaf layer controller, and a ToR layer controller of a data center network are deployed according to the traffic characteristic information.

[0026] Specifically, by the hardware parser module in the ASIC chip, the message header of the data packet is parsed in parallel when the data packet enters the ingress port of the switch, and the key information fields are extracted at line speed level processing capacity, including the source IP address and the destination IP address, which together constitute the end-to-end identification of network layer communication; the source port number and the destination port number are used to identify the specific process of the application layer communication session; the protocol type field is used to distinguish different transmission protocols such as TCP and UDP; and the tenant identification subsegment in the VLAN tag field is extracted in the link layer. The extracted fields together constitute the five-tuple and its additional tenant information. With the five-tuple information as input, the CRC32 hash function calculation module deployed in the ASIC logic unit is used to perform hash transformation on the five-tuple content to generate a flow identifier with strong uniqueness and extremely low collision probability. On this basis, the INT technology is started, and the metadata field used to describe the flow state is directly embedded into the INT extension section of the data packet header. The metadata field includes five types of content, such as flow identifier, flow cumulative byte count, flow duration timestamp, current path hop count, and path identifier, and propagates along the link to each level of the switching node with the data packet, ensuring that the system can continuously monitor the flow state without interrupting data transmission. As the data packet forwarding process progresses, the INT metadata field is continuously updated to record the dynamic evolution trajectory of the flow. At the same time of INT information synchronization, a double-threshold judgment mechanism is started in the ASIC, which is based on whether the duration of the flow exceeds the set time threshold (such as 100 milliseconds) and whether the cumulative data volume exceeds the byte threshold (such as 10 MB). When both thresholds are met, the flow is marked as an elephant flow, if only a short burst but the data volume fluctuates sharply, the flow is classified as a burst flow, if the transmission is continuous but the rate is low, the flow is identified as a long-tail flow, and the rest of the flow is defaulted as a normal mouse flow, completing the fine-grained classification and labeling of the flow type. The flow type classification result is mapped to the tenant identification field, and through the hardware logic lookup mechanism, each flow is divided into gold, silver, bronze, or normal levels according to its tenant level field, and the final result is written into the active flow table constructed by the on-chip SRAM of the ASIC chip. The active flow table is continuously maintained by the hardware module in a high-speed access manner, and the table entries include the flow identifier, tenant level, flow type, cumulative byte volume, latest timestamp, and path history, which reflect the active flow set and its state evolution handled by the current switching node in real time. The flow table has an automatic aging mechanism, which automatically removes the table entry corresponding to a flow when the active state of the flow is interrupted for more than a set time (such as 1 second), to maintain the availability and efficiency of the flow table resources. The traffic feature information is uploaded to the control plane as a reference for the deployment and function initialization of the Spine layer, Leaf layer, and ToR layer three-layer controllers.The Spine layer controller constructs a global traffic view and formulates a cross-Pod path strategy according to the traffic characteristics of the whole network, the Leaf layer controller executes an intermediate-granularity path planning and scheduling strategy in combination with the traffic state of the local domain, and the ToR layer controller directly obtains the fields in the traffic characteristic mapping table for local forwarding and scheduling response at the microsecond level. The deployment structure of the three-layer controller is dynamically optimized according to the flow density distribution, tenant level distribution and link state index in the active flow table.

[0027] In an embodiment, the process of deploying the Spine layer controller, the Leaf layer controller and the ToR layer controller of the data center network according to the traffic characteristic information can specifically include the following steps: performing control responsibility division on the data center network based on the traffic characteristic information to obtain a Spine layer control responsibility, a Leaf layer control responsibility and a ToR layer control responsibility; deploying a global coordinator instance in the Spine switch, deploying a domain controller instance in the Leaf switch, and integrating a local decision maker instance in the ToR switch control plane according to the Spine layer control responsibility, the Leaf layer control responsibility and the ToR layer control responsibility; connecting the global coordinator instance, the domain controller instance and the local decision maker instance and mapping the control object to different controllers to obtain the Spine layer controller, the Leaf layer controller and the ToR layer controller.

[0028] Specifically, the flow characteristic information is structured and analyzed. After receiving multi-dimensional data reported by the underlying ASIC chip of the switching node, including flow type, tenant level, path hop count, real-time bandwidth usage, flow table activity, and the like, key indicator sets of the current load structure of the entire data center network, tenant level distribution, link heat, and path coupling relationship are constructed, and systematic judgment of control granularity division and control responsibility sinking is made based on the key indicators. By analyzing the active flow density and flow type distribution map, elephant flow paths across Pods and medium and long-term load concentration areas are identified, and such cross-domain scheduling requirements are determined as the control responsibility scope of the Spine layer. For medium traffic and short-term burst loads concentrated in a single Pod, they belong to the local path scheduling responsibility scope of the Leaf layer. For short flows and mouse flows with high path change frequency and obvious packet-level scheduling requirements, fine scheduling and flow table maintenance are completed locally by the ToR layer because the scheduling process must fully utilize hardware cache and port queue states. According to the divided control levels and control targets, specific switching nodes are selected in the network physical structure for controller instance deployment. A global coordinator instance is deployed in the Spine switch to maintain and deduce the global path planning constraints, multi-Pod bandwidth distribution state, and global tenant SLA strategy. A domain controller instance is deployed in each Leaf switch to monitor the link utilization rate, select the path, adjust the QoS configuration, and issue short-term scheduling tasks. A local decision maker instance is integrated in the control plane of each ToR switch to quickly extract the port congestion state, update the hardware flow table, identify burst flows, and perform microsecond-level path switching. The three levels of controllers are connected and mapped to realize linkage through the asynchronous state synchronization, path instruction issuing, and state feedback mechanisms. All control objects in the network, including switching nodes, port resources, flow identifiers, path segments, and queue instances, are mapped to different controller instances through a hash function mapping mechanism to ensure that the control responsibilities are dispersed as needed. For example, the Spine layer global coordinator holds the control right of all gold medal tenant cross-Pod flows through consistent hash mapping, the Leaf layer domain controller is responsible for the local flow traffic engineering objects of bronze and silver medal tenants, and the ToR layer local decision maker has priority access to all normal tenants and burst flow path maintenance, fast sinking of rerouting, and other key objects.

[0029] In a specific embodiment, the process of performing step 102 can specifically include the following steps: Classify the gold medal tenant, silver medal tenant, bronze medal tenant and ordinary tenant according to the tenant identification field in the traffic characteristic information, and configure the mapping relationship of the virtual output queue in the ASIC chip; Based on the mapping relationship, allocate queues for the data traffic of each tenant level, and set weighted scheduling parameters for each virtual output queue; Based on the weighted scheduling parameters, perform bandwidth proportion calculation and transmission opportunity allocation to obtain a basic bandwidth allocation result; According to the basic bandwidth allocation result, perform dynamic bandwidth borrowing and monitor the utilization rate of each virtual output queue. When it is detected that the utilization rate of the virtual output queue is lower than the utilization rate threshold, perform bandwidth reallocation process to obtain a bandwidth quota scheme.

[0030] Specifically, the tenant identification bits in the VLAN tag field extracted by the parser in the switching node are taken as the classification basis, and different tenants are divided into four levels according to predefined rules through a hardware logic judgment mechanism. The priority field values of 7-6 are determined as gold medal tenants, 5-4 as silver medal tenants, 3-2 as bronze medal tenants, and 1-0 as ordinary tenants. Based on the classification results, 8 physical virtual output queues are configured in the ASIC chip of each switching device, and the tenant levels are one-to-one mapped to the queue numbers, i.e., gold medal tenants are mapped to queues 0 and 1, silver medal tenants are mapped to queues 2 and 3, bronze medal tenants are mapped to queues 4 and 5, and ordinary tenants are mapped to queues 6 and 7, to realize the structural basis of physical isolation and traffic classification scheduling. The output queues mapped by the tenant levels are set with exclusive weighted scheduling parameters for each group of queues, including scheduling weight and scheduling quantum value. In terms of scheduling weight, the gold medal tenant queue is given the highest priority, with a weight value of 4 and a corresponding scheduling quantum of 8192 bytes; the weight value of the silver medal tenant queue is 3, and the quantum is 6144 bytes; the weight of the bronze medal tenant queue is 2, and the quantum is 4096 bytes; and the ordinary tenant queue is only given a weight of 1, and the quantum is 2048 bytes. The scheduling parameters are written into the register group of the scheduling engine of the switching chip, and the data packet out-of-queue rhythm is controlled in proportion in the scheduling period, so that higher-level tenants obtain higher frequency of sending opportunities in the forwarding link. When entering the scheduling period, the system scheduler runs the weighted round-robin mechanism according to the weighted scheduling parameters, and performs proportional data scheduling operation based on the remaining bandwidth resources of the current period and the quantum values of each queue. The scheduling engine evaluates the current backlog of data packets of each queue, and judges whether the current round has scheduling qualification in combination with the quantum value. If it is satisfied, the scheduling flag bit is immediately sent to the out-port to control the data out-of-queue and update the scheduling debt value of the queue. Through multiple rounds of scheduling iteration, the basic bandwidth proportion allocation result of each tenant level in the same physical port is constructed by using the weighted round-robin method. According to the basic bandwidth allocation result, dynamic bandwidth borrowing is performed, which allows high-priority tenants to temporarily borrow unused scheduling opportunities when their own traffic increases and low-priority queues are idle. A bandwidth usage rate counter is configured in the ASIC chip for each queue, which is updated every 1 millisecond to detect the average utilization rate of each virtual output queue in real time. When it is detected that the utilization rate of a certain queue is continuously lower than a preset threshold (such as 50%), the bandwidth redistribution process is triggered, and the unused sending opportunities of the idle queue in the current scheduling period are redistributed to the higher-level queue according to the priority principle, to improve the sending rhythm of the latter. During the redistribution execution, the scheduling weight and quantum value of the target queue are dynamically adjusted, and the corresponding register content in the scheduling engine is temporarily updated.

[0031] In a specific embodiment, the process of performing step 103 can specifically include the following steps: In the Spine layer controller, the bandwidth quota scheme is input into a graph neural network model, and a graph data structure is constructed, the vertices of the graph data structure containing Spine switches, Leaf switches, ToR switch nodes and their node feature vectors, and the edges containing inter-switch physical links and their link feature vectors; In the three-layer graph neural network layer in the graph neural network model, feature aggregation and propagation calculation are performed on the graph data structure to obtain east-west traffic prediction data; Based on the east-west traffic prediction data, a path planning is performed to obtain a global path planning scheme.

[0032] Specifically, the modeling module of the graph neural network is started inside the Spine layer controller, and the bandwidth quota scheme is imported as the input basic parameter into the network state modeling process. The graph data structure is constructed according to the physical topology structure of the data center. The graph structure takes all switch devices in the entire data center as the node set and takes the physical link connection relationship as the edge set, forming the graph expression form of G=(V,E). The vertex set V contains all Spine switches, Leaf switches and ToR switch instances, and each vertex is attached to the node feature vector associated with it. The composition of the node feature vector includes: the port bandwidth utilization rate of the current node, the local virtual queue average length, the tenant level distribution structure, the scheduling quantum usage state, the cache overflow count, the forwarding table hit rate, the controller synchronization delay and other multi-dimensional indicators. All indicators are represented in the form of numerical vectors and mapped into a unified feature space to support the normalization processing of network training. For each edge in the graph, i.e., the physical link between switch nodes, a link feature vector is defined for the physical link, and the physical link is attached to the edge set E of the graph structure as an edge attribute. The link feature vector contains information such as total capacity, current utilization, average link delay, packet loss rate, congestion marker frequency, SLA violation times, tenant traffic proportion, path hop count and traffic directionality. The link monitor and queue sampling mechanism periodically collect and upload the information to the Spine layer controller for structured merging processing. After the vertex and edge feature vectors are bound, the entire graph data structure forms an input carrier for traffic prediction and is used as a training sample or prediction input for the graph neural network model. The graph neural network model in the Spine layer controller starts a three-layer graph convolution operation process. In the first layer of the graph neural network, the initial aggregation operation of the feature vector is completed through the standardized adjacency matrix and degree matrix. The feature vector of each switch node is combined with the feature vectors of its adjacent nodes, and the first layer intermediate representation is formed through the activation function mapping. In the second layer of graph convolution, the model continues to diffuse the feature propagation and multi-hop perception along the graph structure, so that the representation vector of each node not only contains local features, but also fuses the link situation and forwarding capacity of upstream and downstream switch nodes on the communication path, realizing the aggregated perception of regional traffic distribution trend of the network. In the third layer of the graph neural network, the aggregated high-dimensional vector is mapped to the target dimension, and the prediction results of the node level and the edge level are output. The node output represents the traffic density change trend of the node in a unit time, and the edge output represents the traffic estimation value of each link under different tenant levels in a given future time window, forming an east-west traffic prediction data set. The east-west traffic prediction data set is used as the input of the optimization model to start the global path planning module.The path planning module constructs a scheduling strategy evaluation system based on multiple objective functions, including four main objectives of minimizing end-to-end communication delay, minimizing path load standard deviation, maximizing SLA achievement rate, and maximizing network throughput efficiency, while imposing a series of hard constraint conditions such as link capacity limit, path hop limit, tenant level path constraint, etc. In the implementation of the path planning strategy, a genetic algorithm is used to search for a combined path scheme, and a scheduling individual is constructed including path combination, weight configuration, and node load division. Through population initialization, crossover and mutation, fitness evaluation and iterative evolution, the optimal path scheduling solution at the current time is converged in the global optimization process. The optimal path scheduling solution constitutes the global path planning scheme, which includes the ECMP path set, the mapping relationship between the path and the tenant level, the weight scheduling proportion, the expected load balancing index, and the delay margin identifier.

[0033] In a specific embodiment, the execution step is based on east-west traffic prediction data to perform path planning to obtain a global path planning scheme, which can specifically include the following steps: In combination with the east-west traffic prediction data, the objective functions including minimizing network total delay, maximizing load balancing degree, and maximizing tenant SLA satisfaction rate are set; The link capacity constraint, flow conservation constraint, and tenant SLA constraint are taken as constraint conditions, and the candidate ECMP path set of each node pair in the graph data structure is defined as the decision variable; According to the objective function, the constraint condition, and the decision variable, an initial path planning result is solved; According to the initial path planning result, the shortest delay path is allocated to the gold medal tenant, the load balanced path is allocated to the silver medal tenant and the bronze medal tenant under the SLA constraint, and the remaining path is allocated to the ordinary tenant, to obtain a global path planning scheme.

[0034] Specifically, an objective function of the optimization model is constructed based on the east-west traffic prediction data. The objective function is expressed in a multi-objective joint form, including: minimizing the total network latency, i.e., the sum of the total delay introduced by the transmission of each tenant data on each link after path allocation is as low as possible, ensuring fast response of high-opt tenant and critical business path; maximizing the degree of link load balancing, reducing the load distribution variance of all links in unit time, preventing path concentration and causing bottleneck effect; maximizing tenant SLA satisfaction rate, by evaluating the actual degree of achievement of service indicators such as delay, bandwidth and packet loss of each tenant under the planned path, improving tenant experience. At the same time, the constraint set of the optimization problem is constructed according to the network operation constraints, wherein the link capacity constraint requires that the total flow allocated to a certain link at any time should not exceed its bandwidth upper limit, to avoid queue overflow or data loss; the flow conservation constraint is limited for each source-destination node pair, to ensure that all input traffic must be forwarded to the destination node through the intermediate path, and path breakage, loop or flow loss phenomenon cannot occur; the tenant SLA constraint proposes differentiated constraint conditions according to the service quality requirements of different levels of tenants, requiring that the path delay of gold medal tenants should not be higher than the preset upper limit of their service level, silver and bronze tenants need to meet the corresponding minimum threshold in bandwidth, delay and packet loss control, and ordinary tenants use non-critical paths under the condition of resource allowance. All legal ECMP path sets in the graph data structure are defined as decision variables, and the path selection behavior between each pair of nodes is represented by variables. Each path in the path set has an adjustable traffic allocation weight, which is updated continuously in the optimization process until the optimal state of convergence is reached. Based on the multi-objective optimization model and the path decision variable set, a genetic algorithm is used as an optimization solving tool to carry out iterative search. In each generation of population, each individual represents a complete node pair path combination and traffic weight allocation scheme. The individual represents the strategy configuration by encoding the path number and weight value, and the fitness score of each individual is evaluated by the objective function to evaluate its scheduling efficiency and constraint satisfaction level under the current traffic prediction scenario. At the same time, the crossover operation and mutation mechanism are introduced to maintain population diversity and explore better solutions. Through continuous multiple generations of evolution and fitness screening, when the change trend of the fitness function tends to converge or reaches the maximum iteration number, the current optimal individual is extracted as the initial path planning result. On the basis of the initial path planning result, the path resource allocation strategy is executed combined with the tenant level information.For gold medal tenants, the path with the shortest latency in the predicted path set is selected as the main path, regardless of the path load balancing priority, to ensure the highest guarantee of its communication performance; for silver and bronze medal tenants, the path combination that meets the service delay and bandwidth requirements is selected according to the SLA threshold, and the path group with the most balanced load is selected from multiple candidate paths as the bearing path to maximize the path life cycle and avoid burst congestion; and for ordinary tenants, after the path resources of high-level tenants are configured, the remaining unsaturated link resources and suboptimal paths are used for traffic scheduling to preferentially guarantee global stability and resource waste, while not interfering with critical business communication links. The path planning results processed by the level difference are summarized as a structured global path planning scheme, and the path selection, scheduling proportion, service level adaptation label and expected congestion index between each pair of source-destination nodes are marked to form a path decision set.

[0035] In a specific embodiment, the process of performing step 104 can specifically include the following steps: In the Leaf layer controller, the global path planning scheme is combined with the link utilization vector, the queue length vector and the tenant SLA margin vector of the current network environment to construct a state space of a Markov decision process; Based on the state space, the ECMP path set and its weight distribution are defined as the action space of the reinforcement learning algorithm, and the reward function is calculated for the execution result of each candidate path weight action in the action space to obtain a reward value; The reward value is input into a deep Q network for path weight optimization to obtain a scheduling instruction; The scheduling instruction is sent to the ToR layer controller and synchronized through the Raft protocol to obtain a collaborative exchange control result.

[0036] Specifically, in the Leaf layer controller, the global path planning scheme periodically issued by the Spine layer controller is received and parsed, which contains the optional ECMP path set between each pair of source-destination nodes, the predicted delay value corresponding to the path, the link utilization rate boundary, the path level label and the initial reference of the scheduling weight. To construct the state space for reinforcement learning training and decision-making, the Leaf controller synchronously accesses the network environment monitoring data module maintained locally, from which the link utilization rate vector at the current time is extracted, i.e. the ratio data between the allocated capacity and the total capacity per unit time on each link; the queue length vector of each virtual output queue corresponding to the current port is extracted, reflecting the actual traffic accumulation degree under each tenant level; at the same time, the target SLA value defined in the service level contract of each tenant is extracted, and compared with the actual network performance indicators (such as delay, throughput) obtained by the tenant under the current path planning to construct the tenant SLA margin vector, reflecting the gap between the current service and the target SLA. After fusing the above three types of indicators with the path planning data, the Markov decision process state space is formed. Based on the state space, the action space in reinforcement learning is defined, and each action in the action space represents a weight distribution strategy for the paths in the current ECMP path set. The Leaf controller configures several groups of path weight distribution templates in the optional path set for each source-destination node pair, such as equal weight distribution, bandwidth proportion-based weighting, predicted delay inverse-based weighting, link congestion penalty factor-based weighting, tenant priority product-based weighting, etc., and each group of path combination and weight forms a specific action. Each time an action is selected from the action space for path scheduling attempt, and real-time network feedback data after action execution is collected for reward value calculation. The reward function is designed in a multi-objective fusion manner, considering the changes in port queuing delay, traffic packet loss rate, link load variance, and tenant SLA completion degree after scheduling, all of which are quantified after normalization to ensure fair weighting of indicators in the reward function under different scales. The reward value is used as the evaluation standard for the scheduling performance of the current action in the current state, which is fed back to the reinforcement learning module for Q value update. The state space vector is input as the input state, the action number is input as the input action, and the current reward value is input as the supervision signal to construct the deep Q network model in the Leaf controller. The deep Q network model consists of an input layer, multiple fully connected hidden layers and an output layer, uses ReLU activation function, mean square error loss function and Adam optimizer, stores training trajectories through experience replay mechanism and periodically samples state-action-reward-next state four-tuple from the sample pool for gradient descent training. The model uses a dual network structure of target network and main network to improve training stability, and introduces an ε-greedy strategy to control the dynamic balance between exploration and utilization, ensuring that the model can fully explore high-yield action combinations in the state space.When the training reaches a certain round or the model error converges to a set threshold, the DQN model outputs the action with the maximum Q value in the current state as the optimal path weight allocation strategy. The action with the maximum Q value is converted into a scheduling instruction format, including path ID, corresponding weight, tenant level, scheduling effective timestamp, and effective period information, and is packaged into a standardized control instruction structure. The Leaf layer controller sends the scheduling instruction to the corresponding ToR layer controller through the southbound control interface, and embeds the state synchronization mechanism of the Raft protocol in the transmission path. To ensure consistency and fault tolerance between controllers, the Leaf controller acts as the Leader node of the Raft protocol, encapsulates the latest scheduling instruction as a log entry, and broadcasts it to all slave ToR controller instances with a sequence number and version identifier. After receiving the instruction log, the ToR controller performs verification, application, and returns an acknowledgment response. When most ToR nodes complete the log application operation, the scheduling instruction is considered consistent and officially takes effect. The scheduling strategy after taking effect immediately updates the hardware flow table in the local switch chip, realizes the adjustment of path forwarding behavior, and thus completes the microscopic path scheduling operation based on deep reinforcement learning decision support in the data plane. An end-to-end closed-loop response is formed between the control layer and the execution layer, and a collaborative exchange control result is obtained.

[0037] In a specific embodiment, the process of inputting the reward value into the deep Q network to perform path weight optimization to obtain the scheduling instruction can specifically include the following steps: Input the reward value into the deep Q network to calculate the Q value prediction result of each candidate path weight action in the current network environment. Based on the Q value prediction result, apply the experience replay mechanism to generate Q value function weight parameters. According to the Q value function weight parameters, execute the ε-greedy strategy to select the optimal path weight for the current network environment from the candidate path weight actions, and convert the optimal path weight into the corresponding scheduling instruction.

[0038] Specifically, the Markov state vector contains the link utilization vector, the queue length vector of each virtual output queue, the tenant SLA margin vector, and the path set structure in the current global path planning, etc. After unified dimension expansion and normalization processing, the information is input into the input layer of the deep Q network, and then propagated forward along the multi-layer fully connected neural network. In each hidden layer, the ReLU activation function is used to enhance the non-linear expression ability, and the Q value prediction vector equal to the size of the action space is generated in the output layer, where each output unit corresponds to a path weight configuration action, and its value represents the expected return or cost performance of the current action in the current state. The Q value prediction result and the structured interaction record of state-action-reward-next state (i.e. four-tuple) are written into the experience replay buffer as a training sample. The experience replay buffer maintains a limited length of historical interaction sample pool in the form of a circular queue, and supports random sampling of training batches from the sample pool to prevent time correlation between samples from affecting learning stability. In each training cycle, multiple four-tuple samples are randomly sampled from the experience pool, and for each group of samples, a Q value updating process is performed. The output of the target Q network is used as the expected value reference in the current state, the deviation between the current main Q network prediction value and the expected value is measured by the target function such as mean square error, and the gradient update is performed on each weight parameter in the main network according to the error by the back propagation algorithm. Through continuous training iterations, the Q network gradually approaches the optimal action value function, enabling it to accurately predict the potential return of each action in any given state, thereby enabling the scheduling decision to have learning and generalization capabilities. As the network weight parameters gradually optimize and converge, according to the Q value function prediction result obtained in the current training cycle, the action selection process based on the ε-greedy strategy is started. The ε-strategy aims to balance the relationship between "exploration" and "exploitation" in path scheduling, i.e. in part of the scheduling period, the system performs random action selection with probability ε to ensure that the network can jump out of the local optimal policy area and try unassessed policy paths, while in the remaining 1-ε scheduling period, the path weight configuration action with the maximum Q value is directly selected as the scheduling decision. The value of ε is initially set to 1 and gradually converges to a low constant value (such as 0.1) according to an exponential decay function as the number of training rounds increases, realizing an early exploration-based and late utilization-based strategy conversion mechanism. When the Q network determines the optimal action in the current state, the path weight configuration represented by the optimal action is parsed into specific scheduling instruction content. Each action corresponds to a path set and its corresponding weight distribution ratio, for example, a source-destination node pair contains three optional ECMP paths, and the action corresponds to path weights {path 1: 60%, path 2: 30%, path 3: 10%}. According to the weight ratio, a scheduling instruction structure is constructed, which contains path ID, target weight, tenant level adaptation information, effective timestamp, priority label, and control policy version number, and is packaged into a standard format instruction conforming to the control channel protocol.

[0039] In a specific embodiment, the execution step sends the scheduling instruction to the ToR layer controller and synchronizes through the Raft protocol, and the process of obtaining the collaborative exchange control result can specifically include the following steps: The scheduling instruction is encapsulated into a Raft protocol transmission message using the Raft protocol; The first node and the second node are confirmed based on the Raft protocol transmission message, the scheduling instruction is sent from the first node to the second node, the second node triggers re-election if no heartbeat is received within a preset period, and control state synchronization information is obtained; When it is detected that the deviation between the local version number and the global version number in the control state synchronization information exceeds a preset threshold, the latest control state is obtained from the first node and the local state is updated, the hardware flow table cache and the path weight configuration in the ToR layer controller are updated, and the collaborative exchange control result is obtained.

[0040] Specifically, in the Leaf layer controller, path scheduling instructions optimized via deep Q network are generated, which contain key fields such as source-destination node pair, ECMP path set, path weight ratio, target tenant level, scheduling priority, control version number, and scheduling effective timestamp. The scheduling instructions are encapsulated as Raft transmission message format as part of the Raft protocol log entry. The Raft transmission message format contains the log index defined by Raft, the current term number, the source node ID, the target node ID, the log data length, the log load content (i.e. the scheduling instruction body), and the integrity check field. Based on the leader election and log replication mechanism of the Raft protocol, node role division and master-slave confirmation are performed among multiple instances of the ToR controller. Through the start of the Raft pre-vote and vote aggregation process, the first node in the current scheduling period is selected as the master node (Leader), and the remaining nodes are selected as the second node (Follower) in the passive state to receive and replicate the control log published by the first node. After the first node is ready, the first node broadcasts the encapsulated scheduling instruction log message to all second nodes through TCP or control bus-oriented transmission channels, and periodically (such as every 100 ms) sends a heartbeat packet (AppendEntries empty message) to the second node to maintain the effectiveness of the leadership. When the second node does not receive a heartbeat message from the first node within a complete preset period (such as 150 ms), the leader re-election process of the Raft protocol is automatically triggered, switching to the Candidate state, initiating a new round of term, and promoting control plane role update and responsibility migration, ensuring that the scheduling instructions can still be taken over by subsequent nodes and continue to be transmitted when the first node is disconnected or down, forming a stable state advancement mechanism for the control state. While the heartbeat period and leader election mechanism are running stably, a version vector management mechanism is maintained to track the consistency of the control state among ToR controller nodes. Each ToR node maintains a local version number and a global version number synchronized from the first node, and sets an allowed version deviation threshold for dynamically detecting synchronization delay and state drift. When a node finds that the absolute deviation between its local version number and the last received global version number exceeds the set δ (such as δ = 5), it initiates a state update process to apply to the current Leader node for pulling the latest control state data, which includes the latest path scheduling instructions, as well as the corresponding path weight configuration table, active flow table snapshot, queue state summary, and scheduling strategy version label. After the synchronization process is completed, the ToR controller immediately applies the received scheduling data to its local state machine, and updates the ASIC chip on-chip hardware flow table by accessing the local switch chip driver interface.Specifically, the controller constructs new flow table entries according to the path ID and weight value in the scheduling instruction, and replaces or inserts the new flow table entries into the hardware table structure of the chip, the flow table entries including a flow identifier, an egress port number, a path weight value, a priority level, a timeout counter, an entry condition mask and the like, and simultaneously updating the virtual output queue scheduling parameters bound thereto, to ensure that all traffic matching the path from the current time slice will be distributed to the target path according to the new weight proportion. To improve the dynamic response capability, the ToR controller synchronously refreshes the path state cache and the forwarding statistics module.

[0041] The data center and cloud computing exchange method in the embodiment of the application is described above, and the data center and cloud computing switch in the embodiment of the application is described below. The data center and cloud computing switch in the embodiment of the application is used to execute the data center and cloud computing exchange method.

[0042] In the embodiment of the application, the P4 programmable data plane is deployed in the programmable ASIC chip built in the data center and cloud computing switch, nanosecond-level hardware identification of east-west traffic of multiple tenants is realized, the traffic identification delay is greatly reduced compared with the traditional software processing mode, and the accuracy and real-time performance of traffic feature extraction are ensured based on the CRC32 hash function and the INT technology. A decentralized control system of the Spine layer, the Leaf layer and the ToR layer is constructed, the single-point failure risk and the scalability problem of the traditional centralized controller are effectively solved, the coordination of the second-level global policy formulation, the hundred-millisecond-level intra-domain path selection and the microsecond-level packet forwarding decision is realized through the hierarchical decision mechanism of different time scales, the differentiated bandwidth allocation and the quality of service isolation of the gold medal, silver medal, bronze medal and ordinary tenants are realized through the virtual output queue mapping and the weighted DRR scheduling algorithm based on the tenant level, the mutual interference between tenants is effectively avoided, and the utilization efficiency of network resources is improved through the dynamic bandwidth borrowing mechanism. The graph neural network model is deployed in the Spine layer controller, the topology relationship and the traffic propagation characteristics of the data center network are modeled, the accurate prediction of the east-west traffic mode in the future time window is realized, the reinforcement learning algorithm of the deep Q network is used in the Leaf layer controller, the Markov decision process and the multi-objective reward function are constructed, the adaptive optimization of the path weight is realized, the ECMP path allocation strategy can be dynamically adjusted according to the network state, the network load balancing degree and the overall performance are effectively improved. The hardware flow table cache and the dynamic path switching mechanism based on the congestion index are realized in the ToR switch, the fast response capability of path re-routing is completed, the network congestion and link failure are effectively coped with, and the continuity and real-time performance of data transmission are ensured. Through the Raft protocol and the version vector mechanism, the state synchronization and decision conflict problem in the distributed control system is solved, and the coordinated operation of the three-layer control architecture is ensured.

[0043] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described here.

[0044] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0045] The above-described and the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data center and cloud computing exchange method, characterized in that, include: The east-west data traffic is analyzed to obtain traffic characteristic information, and the Spine layer controller, Leaf layer controller and ToR layer controller of the data center network are deployed at the same time. Based on the tenant identifier field in the traffic characteristic information, a virtual output queue mapping is performed to obtain a bandwidth quota scheme. In the Spine layer controller, east-west traffic prediction is performed in conjunction with the bandwidth quota scheme to obtain a global path planning scheme; In the Leaf layer controller, a scheduling instruction is created according to the global path planning scheme, and the scheduling instruction is sent to the ToR layer controller for synchronization to obtain the cooperative exchange control result.

2. The data center and cloud computing exchange method according to claim 1, characterized in that, The process of parsing east-west data traffic to obtain traffic characteristic information, and simultaneously deploying the Spine layer controller, Leaf layer controller, and ToR layer controller of the data center network, includes: The hardware parser of the ASIC chip extracts the header fields of the received east-west data traffic to obtain a five-tuple information including source IP address, destination IP address, source port number, destination port number, protocol type, and tenant identifier field in VLAN tag; Based on the five-tuple information, a CRC32 hash function is calculated and a metadata field is embedded in the packet header using INT technology to obtain identifiable traffic data. The flow duration stamp and cumulative byte count in the identified traffic data are determined by a dual threshold to obtain the flow type classification result; The flow type classification result and the tenant identifier field are mapped hierarchically, and an active flow table record is established in the on-chip SRAM of the ASIC chip to obtain flow characteristic information; Deploy the Spine layer controller, Leaf layer controller, and ToR layer controller of the data center network based on the traffic characteristic information.

3. The data center and cloud computing exchange method according to claim 2, characterized in that, The deployment of the Spine layer controller, Leaf layer controller, and ToR layer controller for the data center network based on the traffic characteristic information includes: Based on the traffic characteristic information, the control responsibilities of the data center network are divided to obtain the control responsibilities of the Spine layer, Leaf layer, and ToR layer. Based on the Spine layer control responsibilities, the Leaf layer control responsibilities, and the ToR layer control responsibilities, deploy a global coordinator instance in the Spine switch, deploy a domain controller instance in the Leaf switch, and integrate a local decision-maker instance in the ToR switch control plane; The global coordinator instance, the domain controller instance, and the local decision-maker instance are connected, and the control objects are mapped to different controllers to obtain the Spine layer controller, the Leaf layer controller, and the ToR layer controller.

4. The data center and cloud computing exchange method according to claim 1, characterized in that, The process of mapping virtual output queues based on the tenant identifier field in the traffic feature information to obtain a bandwidth quota scheme includes: Based on the tenant identifier field in the traffic characteristic information, gold tenants, silver tenants, bronze tenants and ordinary tenants are classified into different levels, and the mapping relationship of virtual output queues is configured in the ASIC chip. Based on the mapping relationship, queues are allocated to the data traffic of each tenant level, and weighted scheduling parameters are set for each virtual output queue. Based on the weighted scheduling parameters, bandwidth ratios are calculated and transmission opportunities are allocated to obtain the basic bandwidth allocation results. Based on the basic bandwidth allocation results, dynamic bandwidth borrowing is performed and the utilization rate of each virtual output queue is monitored. When the utilization rate of a virtual output queue is detected to be lower than the utilization rate threshold, a bandwidth reallocation process is executed to obtain a bandwidth quota scheme.

5. The data center and cloud computing exchange method according to claim 1, characterized in that, In the Spine layer controller, east-west traffic prediction is performed in conjunction with the bandwidth quota scheme to obtain a global path planning scheme, including: In the Spine layer controller, the bandwidth quota scheme is input into the graph neural network model to construct a graph data structure. The vertices of the graph data structure contain Spine switches, Leaf switches, ToR switch nodes and their node feature vectors, and the edges contain physical links between switches and their link feature vectors. In the three-layer graph neural network model, feature aggregation and propagation calculations are performed on the graph data structure to obtain east-west flow prediction data. Based on the east-west traffic prediction data, a global path planning scheme is obtained.

6. The data center and cloud computing exchange method according to claim 5, characterized in that, The process of performing path planning based on the east-west traffic prediction data to obtain a global path planning scheme includes: Based on the east-west traffic prediction data, an objective function is set that includes minimizing total network latency, maximizing load balancing, and maximizing tenant SLA satisfaction rate. Link capacity constraints, traffic conservation constraints, and tenant SLA constraints are used as constraints, and the candidate ECMP path set for each node pair in the graph data structure is defined as a decision variable. The initial path planning result is obtained by solving the objective function, the constraints, and the decision variables. Based on the initial path planning results, the shortest latency path is assigned to gold tenants, load-balanced paths are assigned to silver and bronze tenants under SLA constraints, and the remaining paths are assigned to ordinary tenants, thus obtaining a global path planning scheme.

7. The data center and cloud computing exchange method according to claim 1, characterized in that, The process of creating scheduling instructions in the Leaf layer controller according to the global path planning scheme and sending the scheduling instructions to the ToR layer controller for synchronization to obtain the cooperative exchange control result includes: In the Leaf layer controller, the global path planning scheme is combined with the link utilization vector, queue length vector, and tenant SLA margin vector of the current network environment to construct the state space of the Markov decision process; Based on the state space, the ECMP path set and its weight assignment are defined as the action space of the reinforcement learning algorithm, and the reward value is obtained by calculating the execution result of the weight action of each candidate path in the action space using a reward function. The reward value is input into a deep Q-network for path weight optimization to obtain scheduling instructions; The scheduling instructions are sent to the ToR layer controller and synchronized via the Raft protocol to obtain the cooperative exchange control results.

8. The data center and cloud computing exchange method according to claim 7, characterized in that, The step of inputting the reward value into a deep Q-network for path weight optimization to obtain scheduling instructions includes: The reward value is input into a deep Q-network to calculate the Q-value prediction results of the weight actions of each candidate path in the current network environment. Based on the Q-value prediction results, an experience playback mechanism is applied to generate Q-value function weight parameters; The ε-greedy strategy is executed according to the Q-value function weight parameters to select the optimal path weight for the current network environment from the candidate path weight actions, and the optimal path weight is converted into the corresponding scheduling instruction.

9. The data center and cloud computing exchange method according to claim 8, characterized in that, The step of sending the scheduling command to the ToR layer controller and synchronizing via the Raft protocol to obtain the cooperative exchange control result includes: The scheduling instructions are encapsulated into Raft protocol transmission messages using the Raft protocol; Based on the Raft protocol, the first node and the second node are confirmed by transmitting messages. The scheduling instruction is sent from the first node to the second node. If the second node does not receive a heartbeat within a preset period, a re-election is triggered to obtain control status synchronization information. When the deviation between the local version number and the global version number in the control state synchronization information exceeds a preset threshold, the latest control state is obtained from the first node and the local state is updated. The hardware flow table cache and path weight configuration are updated in the ToR layer controller to obtain the collaborative exchange control result.

10. A data center and cloud computing switch, characterized in that, Used to perform the data center and cloud computing exchange method as described in any one of claims 1-7.

Images