Small particle resource allocation method based on spn in f6g computing power optical network

Abstract

The application provides a small particle resource allocation method based on SPN in a F6G computing power optical network, and belongs to the technical field of resource allocation in a computing power optical network.The application comprises the following steps: using an SDN control layer to realize real-time sensing of the system state of the entire F6G computing power optical network, including the network link connection state, the time slot resource occupation state on the link and the resource state of the computing power node, and constructing a global virtual topology view; a user terminal sends a computing power service request, and the SDN control layer analyzes the computing power service request; according to the computing power service request and the global virtual topology view, target computing power node selection, alternative path calculation and time slot resource allocation are sequentially performed, and a decision scheme is generated; an execution instruction is issued according to the decision scheme, and service data is routed and forwarded to the target computing power node. The application realizes the collaborative allocation of computing power resources and network resources, significantly improves the success rate of routing optimization, minimizes the transmission cost under the requirement of service delay, and improves the network performance.

Description

Technical Field

[0001] This invention relates to the technical field of resource allocation in optical computing networks, and more particularly to a small-granularity resource allocation method based on SPN in an F6G optical computing network. Background Technology

[0002] With the rapid emergence of new intelligent services such as AR / VR, metaverse, and autonomous driving, network data volume has exploded, posing a significant challenge to integrating widely distributed computing and network resources to meet the differentiated needs of computing power services. Computing power networks, as an innovative network architecture, have emerged to address this challenge. By connecting and interconnecting distributed computing nodes, they enable the coordinated scheduling of computing and network resources, flexibly allocating resources on demand across the cloud, edge, and endpoint according to business needs. The underlying infrastructure is an optical network. The sixth-generation fixed communication network (F6G) computing power optical network, evolved from this, is a network form that deeply integrates optical communication technology with computing power. It can carry the efficient transmission and real-time processing of massive tasks, providing a highly reliable and intelligently scheduled transport pipeline for computing power services. Meanwhile, the rapid development of 5G networks has driven the transformation of vertical industries from simple and singular to diverse and complex. Application scenarios such as government dedicated lines, financial dedicated lines, and large enterprise dedicated lines place higher demands on the isolation and granularity of network slicing. In these scenarios, the minimum bandwidth requirement for services can reach 2Mbps, and services with bandwidth below 10Mbps account for a large proportion. However, current slicing technologies are mostly at the Gbit / s level, resulting in low carrying efficiency. To address this issue, Fine Granularity Unit (FGU) technology for Slice Packet Networks (SPN) has emerged. FGU not only inherits the high-efficiency Ethernet core of SPN but also integrates its fine-grained slicing technology into the overall SPN architecture, providing low-cost, refined, and hard-isolated small-granularity transport pipelines. Through FGU technology, the granularity of hard slices can be refined from 5Gbps to 10Mbps, thereby precisely meeting the urgent needs of 5G+ vertical industry applications and leased line services for carrying differentiated services with low bandwidth, high isolation, and high security.

[0003] The development of computing power optical networks needs to adapt to the business requirements of scenarios such as government dedicated lines, financial dedicated lines and large enterprise dedicated lines. Researching small-granularity resource allocation methods in computing power optical networks is of great significance for improving network resource utilization efficiency and meeting diversified business needs.

[0004] SPN small-granularity technology inherits the high-efficiency Ethernet core of SPN and integrates fine-grained slicing technology into the overall SPN architecture through a hierarchical design, providing a low-cost, refined, and hard-isolated small-granularity transport pipeline. An SPN architecture supporting small-granularity technology is as follows:Figure 4 As shown, the system includes a Slicing Packet Layer (SPL), a Slicing Channel Layer (SCL), a Slicing Transport Layer (STL), a frequency / time synchronization module, and an integrated SDN management and control module. The SPL corresponds to the traditional L2 and L3 layers, and its main function is to route packet data, using an SDN-based controller to establish and manage tunnels. The SCL establishes a slicing layer above the L1 layer, enabling time-slot-based slice transmission. The STL uses Flexible Ethernet (FlexE) technology to decompose services and allocate channels at ports, carrying physical layer services. The frequency / time synchronization module provides high-precision clock synchronization, ensuring accurate alignment of small-granular time slots between nodes. The integrated SDN management and control module is responsible for network-wide status awareness and centralized policy distribution.

[0005] However, traditional SPN network slicing is geared towards single-channel communication load balancing. Applying it directly to F6G computing power optical networks presents the main problem of fragmented allocation of computing and network resources. Because traditional SPN only focuses on network link status and lacks awareness of computing node load, services are easily routed to nodes with unobstructed networks but already full computing power, resulting in severe computational queuing delays.

[0006] Patent application number 202211711824.0 discloses a sliced ​​packet network and a time slot adjustment method for the sliced ​​packet network. The method includes: a source device sending first request information to a destination device via a node device, wherein a first request flag is set in the first request information, and a reserved field in the first request information carries time slot adjustment information; the source device receiving first response information sent by the destination device via the node device, wherein a first response flag is set in the first response information; in response to the source device receiving the first response information within a preset time period, the source device sending control information to the destination device via the node device, wherein an enable flag is set in the control information. This invention solves the technical problem of low adjustment accuracy in time slot adjustment of sliced ​​packet networks in the prior art, but it does not consider flexible time slot allocation for services of different priorities. Summary of the Invention

[0007] To address the technical problem that the isolation and granularity of existing network slicing cannot meet the requirements of application scenarios such as government dedicated lines, financial dedicated lines, and large enterprise dedicated lines, this invention proposes a small-granularity resource allocation method based on SPN in F6G computing power optical networks. Based on the small-granularity technology of SPN, routing decisions and time slot resource allocation are performed to meet service latency requirements while reducing transmission costs and improving network performance.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: a small-granularity resource allocation method based on SPN in an F6G computing power optical network, the steps of which are as follows:

[0009] Step 1: Utilize the SDN control layer to perceive the system status of the entire F6G computing power optical network in real time, including network link connectivity, time slot resource occupancy on the links, and resource status of computing power nodes, and construct a global virtual topology view;

[0010] Step 2: The user terminal sends a computing power service request, and the SDN control layer analyzes the computing power service request;

[0011] Step 3: Based on the computing power service request and the global virtual topology view, the SDN control layer sequentially selects the target computing power node, calculates alternative paths, and allocates time slot resources to generate a decision plan;

[0012] Step 4: The SDN control layer issues execution instructions based on the decision scheme and routes and forwards the business data to the target computing power node.

[0013] The steps of this invention are as follows: First, the computing power optical network monitors the system status in real time, including network link connectivity, time slot resource occupancy on the links, and resource status of computing power nodes, to obtain a comprehensive view of the entire system and provide global data support for joint scheduling of computing power and the network. Second, the user terminal sends a computing power service request to the SDN control layer. The SDN control layer selects the computing power node with the lightest load that meets the computing power requirements based on the current network and computing power node status, and then uses the KSP algorithm to select... There are 10 alternative paths, which are traversed in order. The system allocates time slot resources based on priority along a chosen path. If allocation is successful, a solution is output; otherwise, time slot resources are allocated along the next alternative path until a decision is reached. This collaborative allocation of computing and network resources significantly improves the success rate of route optimization. Subsequently, time slot resources are allocated to services according to this decision, routing services from the source node to the target computing node. This establishes an end-to-end hard-isolated transport pipeline for small-granularity services, ensuring high reliability of service transmission. Finally, for computing services with latency requirements, while meeting link latency requirements, it also checks whether time slots are evenly distributed. If the actual allocated links and time slots cannot meet the required latency, the service is blocked, fundamentally guaranteeing deterministic low latency for already accessed services. This invention minimizes transmission costs and improves network performance while meeting service latency requirements.

[0014] Preferably, the SDN control layer uses the BGP-LS protocol to collect routing and link information of the underlying physical devices in real time to obtain the connectivity status of the network topology; it uses network telemetry technology to collect the time slot resource occupancy on the SPN link and the real-time load utilization of the physical resources of each computing node with millisecond-level accuracy; and it aggregates the collected multi-dimensional status data to construct a global virtual topology view that includes the time slot status and computing power weight of the entire network.

[0015] Preferably, the SDN control layer is centrally deployed in the core cloud. The SDN control layer interacts with the transport network layer, edge layer and cloud layer devices through standard interfaces to collect status and distribute configurations. The SDN control layer receives computing power service requests from the user terminal layer.

[0016] The global virtual topology view uses a weighted undirected graph. It means that, among them, This represents the set of SPN nodes in the SPN network. Represents the set of SPN links in an SPN network; connects SPN nodes. The SPN link is Where i and j are node numbers; in an SPN link, if a pair of cables transmits directional data, then there is an SPN link. , These represent the forward link and the backward link, respectively.

[0017] Only one direction needs to be allocated for computing power service requests; only one 5G bandwidth is available for small-granularity services in each link.

[0018] Preferably, the total number of all service requests in the computing power service request is The collection of computing power service requests Indicates; the first [transmission] sent by the user terminal A service request is represented as ,in, Indicates the first The source node of each service request. For the first The transmission bandwidth of each service request. It is the first CPU computing resources for each service request It is the first GPU computing resources for each service request It is the first Storage resources required for each service request Representing the Maximum tolerable latency for each service request.

[0019] Preferably, the currently available CPU resources, GPU resources, and remaining storage resources of the selected target computing node are greater than or equal to the CPU computing resources. GPU computing resources and storage resources ;No. A service request originates from the source node. The total network transmission and computation latency to the target node is less than or equal to ;

[0020] The SDN control layer calls the constructed global virtual topology view to obtain the real-time status of computing nodes, and filters them in combination with the computing and storage resources of computing service requests: computing nodes whose remaining resources cannot meet the computing service requirements are eliminated; the comprehensive load utilization rate of the remaining computing nodes is calculated, and the computing node with the lowest comprehensive load utilization rate and the most balanced resource distribution is selected as the target computing node for computing service requests.

[0021] Preferably, the alternative path is implemented as follows: the SDN control layer uses a global virtual topology graph as a basis, defining the physical distance of the underlying fiber optic link as the transmission cost; it runs the KSP algorithm, and when the condition of the first... Transmission bandwidth of each service request and maximum tolerable latency Under the constraints, calculate the minimum transmission cost between the source node and the target computing power node. Alternative routes;

[0022] Time slot resource allocation involves identifying and reserving micro-level idle sub-time slots on each underlying link within the selected alternative paths to carry computing power service requests; the SDN control layer traverses these sub-time slots in ascending order of transmission cost. There are several alternative paths; on the currently traversed alternative paths, a time slot allocation algorithm based on a conflict hash table and business priority is executed to achieve even time slot distribution for small-granularity computing power services.

[0023] Preferably, the implementation method of the time slot allocation algorithm based on conflict hash table and service priority includes:

[0024] (1) Hash pre-allocation and collision hash table construction: count all small-granularity computing power service requests carried on the current candidate path, select the extremely small-granularity services with a required number of time slots of 1 or 2 from the computing power service request set, put them into set B and do not process them for the time being; for other large-bandwidth computing power service requests, calculate the ideal time slot interval according to their bandwidth requirements, and calculate the set of sub-time slot numbers to be allocated according to the hash algorithm; build a hash table for the computing power service requests that generate collisions to obtain a collision hash table, use the time slot number that causes the collision as the key, and use the list of computing power service request IDs that compete for the time slot number that causes the collision as the value;

[0025] (2) Conflict checking and priority-based time slot preemption: Traverse the conflict hash table, for the current conflict time slot number, call the priority evaluation function to calculate the priority of each computing power service request on the conflict time slot, and allocate the conflict time slot to the service with the highest priority; for other services with lower priority, find the unoccupied free time slot closest to the conflict time slot number and pre-allocate it, and delete the current key-value pair in the conflict hash table.

[0026] (3) Iterative convergence: The above conflict checking and reallocation process is executed repeatedly until the length of the business list corresponding to all slot numbers in the conflict hash table is no greater than 1 or the maximum iteration limit is reached.

[0027] (4) Minimal granular service allocation: For computing power service requests requiring 2 time slots, the first available time slot is searched in ascending order from the two sides of the time slot number and descending order from the two sides of the time slot number as the allocated time slot number; for services requiring 1 time slot, an available time slot number is randomly selected for allocation.

[0028] Preferably, the priority is related to the time slot interval of the computing power service request, the position of the previously allocated time slot of the computing power service request, and the interval between the current time slot position and the ideal time slot position;

[0029] The priority of the normalized computing power service request r is:

[0030] ;

[0031] in, For the set of computing power service requests, This is the minimum interval constant that can be achieved by computing power service requests. The ideal time slot interval for computing power service request r; It is the interval between the current conflicting timeslot number n and the previously allocated timeslot number of the computing power service request r; This represents the ideal (i+1)th time slot number that should be placed after allocating i time slots to computing power service request r according to the ideal interval; Here are the normalization coefficients, and .

[0032] Preferably, the decision-making scheme includes the selected target computing node number, the end-to-end data forwarding path sequence, and the specific sub-time slot number allocated to the computing service request for each hop link on the transmission path;

[0033] If the current alternative path cannot complete the allocation due to time slot conflicts, the allocation on this alternative path fails, and the system automatically switches to the next alternative path to re-allocate time slot resources until a decision is successfully reached; if all alternative paths have been traversed... If the alternative paths still cannot meet the latency or bandwidth requirements of the computing power service request, the computing power service request will be rejected.

[0034] Preferably, the SDN control layer, based on the generated decision scheme, issues routing forwarding table entries and time slot configuration instructions to the underlying transmission network layer devices and the target computing power nodes through the standard southbound interface. The physical layer transmission devices, based on the received time slot configuration instructions, complete the hard slicing configuration of the time slot resources of the underlying cross-connect matrix. The data flow of computing power service requests follows the physical path sequence specified by the decision scheme, and is routed from the source node to the target computing power node for computation in the end-to-end physical-level hard-isolated bearer pipeline.

[0035] Record and output the performance metrics involved in the entire F6G computing power optical network resource allocation process, including: end-to-end average latency, routing success rate, network transmission cost, and average resource utilization and load balancing of computing power nodes.

[0036] The beneficial effects of this invention are as follows: This invention applies the small-granularity technology (FGU) of SPN to the resource allocation of optical computing networks, constructs an SPN optical computing network architecture, and mathematically models the computing services and SPN network. The inputs are computing service requests, current network and node states, and the outputs are target nodes, routes, and time slot resource allocation schemes, achieving highly complex multi-dimensional joint scheduling of computing network resources. When small-granularity services need to be transmitted in the F6G optical computing network, the joint computing network load awareness mechanism constructed in this invention, as well as the time slot scheduling algorithm based on a collision hash table and a service priority evaluation function, are used to perform flexible small-granularity resource allocation. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the overall process in this invention.

[0039] Figure 2 This is a flowchart of the priority-based time slot allocation algorithm in this invention.

[0040] Figure 3 This is a diagram of the SPN computing network architecture in this invention.

[0041] Figure 4 This is a schematic diagram of the hierarchical SPN network model in this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, a small-granularity resource allocation method based on SPN in an F6G computing power optical network includes the following steps:

[0044] Step 1: The F6G computing power optical network utilizes the SDN control layer to perceive the system status of the entire F6G computing power optical network in real time, including network link connectivity, time slot resource occupancy on the links, and resource status of computing nodes. The network status in the F6G computing power optical network system status is constantly changing and also serves as a constraint on the algorithm.

[0045] like Figure 3 As shown, the SPN+ computing power optical network architecture proposed in this invention mainly consists of three layers: the SDN control layer, the edge layer, and the user terminal layer. The SDN control layer, acting as the centralized management brain, implements global resource management, routing planning, and task scheduling by executing the proposed computing network load joint perception mechanism and the priority time slot scheduling algorithm based on a conflict hash table. The edge layer consists of edge nodes with limited computing and storage capabilities. The user terminal layer includes a large number of user terminals with computing tasks; all small-granularity computing power service requests in the entire F6G computing power optical network system are generated and initiated in real time by the user terminal devices according to actual needs.

[0046] To achieve the aforementioned control and obtain a comprehensive view of the entire system, the SDN control layer establishes a multi-dimensional real-time state awareness mechanism. For monitoring computing nodes and network links: the SDN control layer uses the BGP-LS protocol to collect routing and link information of underlying physical devices in real time, thereby obtaining the connectivity status of the network topology; it employs network telemetry technology to collect the time slot resource occupancy on SPN links with millisecond-level precision, as well as the real-time load utilization of physical resources such as CPU, GPU, and memory of each computing node. The SDN control layer aggregates the collected multi-dimensional state data to construct a global virtual topology view containing the entire network's time slot status and computing power weights. In subsequent applications, this comprehensive view will serve as the basic data input into the multi-dimensional resource joint scheduling algorithm to support core decision-making processes such as target node evaluation, KSP pathfinding, and hash time slot allocation.

[0047] Figure 3 The transport network layer and its nodes constitute the underlying optical network foundation of SPN. Its main function is to provide widespread access for user terminals and to handle high-speed, hard-isolated forwarding of computing power service data within the network. The cloud layer and its nodes primarily provide massive, centralized computing and storage resources to handle complex or heavy-load computing power services that cannot be handled by edge layer resources. In the overall architecture, the SDN control layer is logically centrally deployed in the core cloud. It interacts with the transport network layer, edge layer, and cloud layer devices through standard interfaces to collect status and distribute configurations, and receives computing power service requests from the user terminal layer. By abstracting the attributes of the physical layer structure, the SDN control layer obtains a virtual topology map, effectively shielding the heterogeneity of the underlying hardware devices, enabling subsequent routing planning and time slot scheduling algorithms to be executed efficiently on a standardized mathematical model.

[0048] Using a weighted undirected graph This represents an SPN network, where, This represents the set of SPN nodes in the SPN network. This represents the set of SPN links in an SPN network. Let's assume the nodes connecting to the SPN are... The SPN link is Where i and j are node numbers. In an SPN link, a pair of cables transmits directional data, hence the existence of an SPN link. Here, "for" and "back" represent forward and backward, respectively. These two directions are absolute values ​​marked when the topology model is built, where "for" is... ,back is And it has node numbers. In this invention, for computing power service requests, only one direction needs to be allocated; path selection and time slot allocation can be completely separated. In engineering applications, the link bandwidth carried by the SPN network is typically above 100G, consisting of service types such as regular services, large-granularity slices, and small-granularity slices. Regular slices and large-granularity slices typically occupy n×5G bandwidth, while small-granularity services are multiplexed within a single 5G bandwidth. Each 5G bandwidth, after deducting overhead, has a total of 480 10M sub-time slots available for service use. To simplify the problem model, it is assumed that only one 5G bandwidth is available for small-granularity services in each link. The constructed weighted undirected graph... It will serve as the underlying mathematical model for the SDN control layer to perform global routing planning and time slot scheduling.

[0049] Step 2: The user terminal sends a computing power service request. The SDN control layer analyzes the computing power service request and makes decisions based on the requirements of the computing power service request and the overall resource status of the computing network, thereby enabling the selection of computing power nodes, routing, and the specification of time slot allocation schemes.

[0050] The computing power service request describes the required computing resources and latency requirements. The computing power service request issued by the user terminal can be described as follows: The total number of all service requests in the computing power service request is... The first one sent by the user terminal A service request can be represented as... ,in, Indicates the first The source node of each service request. For the first The transmission bandwidth of each service request. It is the first CPU computing resources for each service request It is the first GPU computing resources for each service request It is the first The storage resources required for each service request, and Representing the The maximum tolerable latency for each service request. The set of computing power service requests is used for... express.

[0051] The aforementioned service request parameters represent the multi-dimensional resource requirements of computing power services, while the system status perceived by the SDN control layer in step 1 represents the overall resource supply of the computing network. When making decisions, the SDN control layer requires that the currently available CPU, GPU, and remaining storage resources of the selected target computing power node be greater than or equal to the CPU computing power resources. GPU computing resources and storage resources Meanwhile, the business originates from the source node. The total network transmission and computation latency to the target node must be less than or equal to .

[0052] Based on the above constraints, the SDN control layer first parses the service request vector. First, multi-dimensional demand features are extracted. Second, the demand features are imported into a global virtual topology view constructed based on a weighted undirected graph. The computing power nodes that meet the conditions are evaluated and screened through a joint perception mechanism of computing network load. Finally, the KSP algorithm is called for pathfinding, and the priority time slot scheduling algorithm based on a conflict hash table is used to complete the micro-resource allocation. Finally, a complete decision scheme containing the target node, transmission path and precise time slot number is obtained.

[0053] Step 3: The SDN control layer executes joint computing and network scheduling decisions: For each computing power service request parsed in Step 2, the SDN control layer sequentially performs target computing power node selection, alternative path calculation, and time slot resource allocation, ultimately generating a decision scheme. The specific implementation steps are as follows:

[0054] Step 3.1: Target Computing Node Evaluation and Selection. The SDN control layer invokes the computing network load joint awareness mechanism, i.e., the constructed weighted undirected graph, incorporating the real-time status of the computing nodes obtained in Step 1, and combining it with the computing and storage requirements of the computing service requests in Step 2. The process involves screening. First, nodes whose remaining resources cannot meet business needs are eliminated. Then, the overall load utilization of the remaining candidate nodes is calculated. Finally, the node with the lowest overall load utilization and the most balanced resource distribution is selected as the target computing power node for this business.

[0055] Step 3.2: Calculation of Alternative Paths Based on KSP. After determining the target computing node, the SDN control layer uses the global virtual topology constructed in Step 1 as a basis, defining the physical distance of the underlying fiber optic links as the transmission cost. Then, the KSP algorithm is run to meet the service bandwidth requirements. and maximum tolerable latency Under strict constraints, calculate the minimum transmission cost between the source node and the target computing node. There are several alternative paths.

[0056] Step 3.3: Path Traversal and Time Slot Resource Allocation. Select... After selecting alternative paths, time slot allocation is based on the specific physical path. Specifically, path planning determines the macroscopic set of SPN links the service will traverse, while time slot allocation involves finding and reserving microscopic idle sub-time slots for carrying the service on each underlying link within the selected path; the two complement each other. The SDN control layer traverses the paths in ascending order of transmission cost. A backup path is established. On the currently traversed link, a time slot allocation algorithm based on a collision hash table and business priority is executed to achieve even time slot distribution for small-granularity computing power services. The specific algorithm flow is as follows:

[0057] like Figure 2 As shown, the allocation of time slot resources specifically includes the following steps:

[0058] (1) Hash pre-allocation and collision hash table construction. Count all small-granularity computing power service requests carried on the current link. First, select the extremely small-granularity services with a required number of time slots of 1 or 2 from the service set and put them into set B for temporary processing; for other large-bandwidth computing power services, calculate the required number of sub-time slots based on their bandwidth requirements. Then calculate the ideal time slot interval. Then, a set of sub-slot numbers to be allocated is pre-calculated according to the hash algorithm, and a hash function is used to uniquely identify the business request. Perform hashing to generate the starting offset. ; starting offset Starting from the ideal time slot interval The step size is determined by the formula. Calculate the complete set of sub-slot numbers, where After pre-allocation is completed, a hash table is created for the services that generate conflicts to obtain a conflict hash table. The time slot number that caused the conflict is used as the key, and the list of computing power service request IDs competing for that time slot number is used as the value.

[0059] (2) Conflict checking and priority-based time slot preemption. The hash table constructed above is traversed to handle conflicts. For the current conflicting time slot (the key in the hash table), the priority evaluation function is called to calculate the priority of each service in that conflicting time slot. Assign the conflicting time slot to a priority level. For the highest priority service, other lower priority services are pre-allocated by finding the nearest unoccupied free time slot to the conflicting time slot number. After the time slot allocation and the rearrangement of low-priority services are completed, the current key-value pair is deleted. This proximity-based pre-placement strategy can minimize the deviation of low-priority computing power service requests from their ideal time slot interval, thereby limiting the deterioration of their end-to-end latency jitter.

[0060] Priority is related to the time slot interval of the computing power service request, the position of the previously allocated time slot for this computing power service request, and the interval between the current time slot position and the ideal time slot position. The calculation function for normalized priority is as follows:

[0061]

[0062] in, This is the minimum interval constant that a computing power service request can achieve, and it is related to the bandwidth requirements of the computing power service request. In this invention, the bandwidth requirements of small-granularity services range from 1 to 100 Mbps, so the required number of time slots ranges from 1 to 10. . The ideal time slot interval for computing power service request r. It is the interval between the current conflicting timeslot number n and the previously allocated timeslot number of the computing power service request r. This represents the ideal slot number where the (i+1)th slot should be placed after allocating i slots to the computing power service request r according to the ideal interval. Here are the normalization coefficients, and we have Priority of computing power service request r The higher the priority, the more likely the conflicting time slots should be allocated to that service r. This priority evaluation function quantifies the severity of each computing power service request deviating from its ideal time slot distribution, ensuring that the computing power service request with the most severe jitter receives priority time slot compensation, thereby guaranteeing the fairness of global time slot allocation.

[0063] (3) Iterative convergence. The above conflict checking and re-preplacement process is executed repeatedly until the length of the business list corresponding to all time slot numbers in the hash table is no greater than 1, there are no conflicts, or the maximum iteration limit set by the system is reached.

[0064] (4) Minimal Granular Service Allocation. After allocating all services with a time slot requirement of 1 or 2, for services requiring 2 time slots, the first available time slot is searched in ascending order from both sides of the time slot number and descending order from the largest to the smallest, and then used as the allocated time slot number. This bidirectional shrinking search strategy can effectively utilize the edge fragmented time slots at both ends of the link and avoid excessive segmentation of the middle continuous time slots. For services requiring 1 time slot, an available time slot number is randomly selected for allocation. Since the impact of minimal granular computing power service requests on the overall transmission jitter is minimal, this random gap-filling allocation strategy maximizes the utilization rate of fragmented time slots in the underlying SPN link and does not interfere with the already established time slot distribution state of high-bandwidth computing power service requests.

[0065] The micro-time slot resource allocation method proposed in this invention uses the ideal time slot interval for each computing power service request. Using a hash table as the baseline anchor point, collisions are precisely located, and a priority function is used to dynamically correct the degree of deviation. This ensures that the multiple sub-slots allocated to each computing power service request are strictly evenly distributed throughout the entire SPN multiframe cycle. This mechanism fundamentally eliminates the queuing congestion caused by multiple computing power service requests arriving simultaneously at the cross-node, avoids the generation of additional multiframe delays, and provides microsecond-level, zero-jitter deterministic transmission guarantees for the computing power optical network.

[0066] Step 3.4: Decision Scheme Output and Rollback Mechanism. If time slot allocation is successfully completed on the current traversal path, the traversal immediately stops, and the final decision scheme is generated. This decision scheme specifically includes three core configurations: the selected target computing node number, the end-to-end data forwarding path sequence, and the specific sub-time slot number allocated to this service for each hop on the path. If the current path cannot complete allocation due to time slot conflicts, allocation on this path fails, and the system automatically switches to the next alternative path to re-allocate time slot resources until a successful decision scheme is obtained; if the traversal is complete... If the alternative paths still cannot meet the latency or bandwidth requirements of the service, the service will be rejected.

[0067] Step 4: The SDN control layer issues execution instructions and routes and forwards service data. Based on the final decision scheme generated in Step 3, the SDN control layer issues routing and forwarding table entries and time slot configuration instructions to the underlying transmission network layer devices and the target computing power nodes through the standard southbound interface. The physical layer transmission devices, based on the received time slot configuration instructions, complete the hard slicing configuration of the time slot resources in the underlying cross-connect matrix. This establishes a dedicated transmission channel with deterministic bandwidth and strict latency guarantees for this small-granularity computing power service at the physical layer, completely eliminating bandwidth contention and node buffer queuing interference during multi-service concurrency. Subsequently, the data stream requesting the computing power service will be routed and forwarded reliably and with zero jitter from the source node to the target computing power node for computation within the end-to-end physical-level hard-isolated bearer pipeline, following the physical path sequence specified by the final decision scheme.

[0068] Step 5: Performance Metric Recording and Visualization Evaluation. Record and output the core system performance metrics involved in the entire multi-dimensional resource joint scheduling process of the computing network, specifically including: end-to-end average latency, routing success rate, network transmission cost, and average resource utilization and load balancing of computing nodes. Transmit the recorded data to the management and control system for visualization charts and trend analysis, thereby intuitively obtaining a global resource load heatmap and algorithm scheduling efficiency evaluation report of the current F6G computing optical network. This provides network administrators with intuitive data support and decision-making basis for daily operation and maintenance, fault diagnosis, and subsequent network architecture expansion and computing node deployment.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A small-granularity resource allocation method based on SPN in an F6G computing power optical network, characterized in that, The steps are as follows: Step 1: Utilize the SDN control layer to perceive the system status of the entire F6G computing power optical network in real time, including network link connectivity, time slot resource occupancy on the links, and resource status of computing power nodes, and construct a global virtual topology view; Step 2: The user terminal sends a computing power service request, and the SDN control layer analyzes the computing power service request; Step 3: Based on the computing power service request and the global virtual topology view, the SDN control layer sequentially selects the target computing power node, calculates alternative paths, and allocates time slot resources to generate a decision plan; Step 4: The SDN control layer issues execution instructions based on the decision scheme and routes and forwards the service data to the target computing power node; The SDN control layer calls the constructed global virtual topology view to obtain the real-time status of computing power nodes, and filters them in combination with the computing power resources and storage resources of computing power service requests: eliminating computing power nodes whose remaining resources cannot meet the computing power service requirements; calculating the comprehensive load utilization rate of the remaining computing power nodes, and selecting the computing power node with the lowest comprehensive load utilization rate and the most balanced resource distribution as the target computing power node for computing power service requests. The alternative path is implemented as follows: the SDN control layer uses a global virtual topology graph as a basis, defining the physical distance of the underlying fiber optic links as the transmission cost; it runs the KSP algorithm, and when the condition of the first... Transmission bandwidth of each service request and maximum tolerable latency Under the constraints, calculate the minimum transmission cost between the source node and the target computing power node. Alternative routes; Time slot resource allocation involves identifying and reserving micro-level idle sub-time slots on each underlying link within the selected alternative paths to carry computing power service requests; the SDN control layer traverses these sub-time slots in ascending order of transmission cost. There are several alternative paths; on the currently traversed alternative paths, a time slot allocation algorithm based on a conflict hash table and business priority is executed to achieve even time slot distribution for small-granularity computing power services.

2. The small-granularity resource allocation method based on SPN in the F6G computing power optical network according to claim 1, characterized in that, The SDN control layer uses the BGP-LS protocol to collect routing and link information of the underlying physical devices in real time to obtain the connectivity status of the network topology; it uses network telemetry technology to collect the time slot resource occupancy on the SPN link and the real-time load utilization of the physical resources of each computing node with millisecond-level precision; and it aggregates the collected multi-dimensional status data to construct a global virtual topology view that includes the time slot status and computing power weight of the entire network.

3. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 2, characterized in that, The SDN control layer is centrally deployed in the core cloud. The SDN control layer interacts with the transport network layer, edge layer and cloud layer devices through standard interfaces to collect status and distribute configurations. The SDN control layer receives computing power service requests from the user terminal layer. The global virtual topology view uses a weighted undirected graph. It means that, among them, This represents the set of SPN nodes in the SPN network. Represents the set of SPN links in an SPN network; connects SPN nodes. The SPN link is Where i and j are node numbers; in an SPN link, if a pair of cables transmits directional data, then there is an SPN link. , These represent the forward link and the backward link, respectively. Only one direction needs to be allocated for computing power service requests; only one 5G bandwidth is available for small-granularity services in each link.

4. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 2 or 3, characterized in that, The total number of all service requests in the computing power service request is The collection of computing power service requests Indicates; the first [transmission] sent by the user terminal A service request is represented as ,in, Indicates the first The source node of each service request. For the first The transmission bandwidth of each service request. It is the first CPU computing resources for each service request It is the first GPU computing resources for each service request It is the first Storage resources required for each service request Representing the Maximum tolerable latency for each service request.

5. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 4, characterized in that, The selected target computing node currently has available CPU resources, GPU resources, and remaining storage resources that are all greater than or equal to its CPU computing resources. GPU computing resources and storage resources ;No. A service request originates from the source node. The total network transmission and computation latency to the target node is less than or equal to .

6. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 5, characterized in that, The implementation method of the time slot allocation algorithm based on conflict hash table and business priority includes: (1) Hash pre-allocation and collision hash table construction: count all small-granularity computing power service requests carried on the current candidate path, select the extremely small-granularity services with a required number of 1 or 2 time slots from the computing power service request set and put them into set B for temporary processing; for other large-bandwidth computing power service requests, calculate the ideal time slot interval according to the bandwidth requirement, and calculate the set of sub-time slot numbers to be allocated according to the hash algorithm; build a hash table for the computing power service requests that generate collisions to obtain a collision hash table, use the time slot number that causes the collision as the key, and use the list of computing power service request IDs that compete for the time slot number that causes the collision as the value; (2) Conflict checking and priority-based time slot preemption: Traverse the conflict hash table, for the current conflict time slot number, call the priority evaluation function to calculate the priority of each computing power service request on the conflict time slot, and allocate the conflict time slot to the service with the highest priority; for other services with lower priority, find the unoccupied free time slot closest to the conflict time slot number and pre-allocate it, and delete the current key-value pair in the conflict hash table. (3) Iterative convergence: The above conflict checking and reallocation process is executed repeatedly until the length of the business list corresponding to all slot numbers in the conflict hash table is no greater than 1 or the maximum iteration limit is reached. (4) Minimal granular service allocation: For computing power service requests requiring 2 time slots, the first available time slot is searched in ascending order from the two sides of the time slot number and descending order from the two sides of the time slot number as the allocated time slot number; for services requiring 1 time slot, an available time slot number is randomly selected for allocation.

7. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 6, characterized in that, The priority is related to the time slot interval of the computing power service request, the position of the previously allocated time slot of the computing power service request, and the interval between the current time slot position and the ideal time slot position; The priority of the normalized computing power service request r is: ； in, For the set of computing power service requests, This is the minimum interval constant that can be achieved by computing power service requests. The ideal time slot interval for computing power service request r; It is the interval between the current conflicting timeslot number n and the previously allocated timeslot number of the computing power service request r; This represents the ideal (i+1)th time slot number that should be placed after allocating i time slots to computing power service request r according to the ideal interval; Here are the normalization coefficients, and .

8. The SPN-based small-granularity resource allocation method in the F6G computing power optical network according to claim 6 or 7, characterized in that, The decision-making scheme includes the selected target computing node number, the end-to-end data forwarding path sequence, and the specific sub-time slot number allocated to each hop link on the transmission path for computing service requests. If the current alternative path cannot complete the allocation due to time slot conflicts, the allocation on this alternative path fails, and the system automatically switches to the next alternative path to re-allocate time slot resources until a decision is successfully reached; if all alternative paths have been traversed... If the alternative paths still cannot meet the latency or bandwidth requirements of the computing power service request, the computing power service request will be rejected.

9. The SPN-based small-granularity resource allocation method in an F6G computing power optical network according to any one of claims 5-7, characterized in that, Based on the generated decision scheme, the SDN control layer issues routing forwarding table entries and time slot configuration instructions to the underlying transmission network layer devices and target computing power nodes through the standard southbound interface. The physical layer transmission devices complete the hard slicing configuration of time slot resources in the underlying cross-matrix according to the received time slot configuration instructions. The data flow of computing power service requests is routed and forwarded from the source node to the target computing power node for computation in the end-to-end physical-level hard-isolated bearer pipeline along the physical path sequence specified by the decision scheme. Record and output the performance metrics involved in the entire F6G computing power optical network resource allocation process, including: end-to-end average latency, routing success rate, network transmission cost, and average resource utilization and load balancing of computing power nodes.

Images