An elastic resource scheduling method and system for a computing power network

Abstract

The application discloses an elastic resource scheduling method for a computing power network, comprising the following steps: the computing power network obtains statistical information of each computing power center in all computing power centers in the computing power network and channel states of a communication network between any two computing power centers through a routing layer; the computing power network calculates service supply pressure degrees of the computing power centers according to the statistical information of each computing power center, and classifies the computing power centers according to the service supply pressure degrees, namely, the computing power centers are divided into resource shortage sides or resource sufficient sides; and the computing power network inputs the statistical information of the resource shortage sides and the resource sufficient sides and the channel states of the communication network between any two computing power centers into a pre-constructed cross-regional queue model to obtain a global cross-regional queue set of the computing power network. The application can solve the technical problem that the existing centralized resource scheduling method is difficult to respond to demand changes and perform elastic scheduling of resources in a timely manner when a computing node fails or a business volume surges.

Description

Technical Field

[0001] This invention belongs to the field of computing power network technology, and more specifically, relates to a method and system for elastic resource scheduling for computing power networks. Background Technology

[0002] With the emergence of intelligent computing scenarios such as artificial intelligence and autonomous driving, the demand for computing resources is growing exponentially. To meet users' diverse needs for "anytime, anywhere, and on demand" in emerging services, computing networks have emerged. Because computing networks need to adapt to changes in different business requirements, including sudden increases or decreases in computing volume and changes in task types, elastic resource scheduling can better meet these changes. Furthermore, through effective scheduling and management of computing resources, the impact of node failures on the computing network can be avoided, improving user experience and reducing maintenance costs. Therefore, an elastic resource scheduling method is of great significance for improving the overall efficiency, cost-effectiveness, and business adaptability of computing networks.

[0003] There are two main existing resource scheduling methods for computing power networks: the first is a centralized resource scheduling method, in which a centralized scheduling node designs the optimal solution based on the computing resource requirements of the user task, allocates computing resources, and establishes the corresponding network connection; the second is a distributed resource scheduling method, in which the computing power center's computing capacity and network status are published to the network as routing information. After a user submits a task to the computing power network, the routing protocol automatically selects a suitable computing node in the computing power center to execute the task based on the task requirements and node resource information.

[0004] However, the above methods all have some drawbacks that cannot be ignored:

[0005] First, because all resource requests and allocations in the centralized resource scheduling method have to go through the centralized scheduling node, the burden on the centralized scheduling node is too great, making it difficult to respond to changes in demand and perform elastic resource scheduling in a timely manner when computing nodes fail or business volume surges.

[0006] Secondly, the two resource scheduling methods mentioned above are mainly designed by considering the timely delivery of user tasks and the load balancing of computing nodes. They do not further optimize the energy consumption and maintenance costs of resource allocation, thus resulting in high energy consumption and maintenance costs of resource allocation.

[0007] Third, because distributed resource scheduling discovers computing resource nodes through routing within a local area network, it is difficult to discover computing resource nodes in other local area networks. When computing resources in a certain physical domain are scarce, there may still be many available computing resource nodes in other physical domains isolated from that region, resulting in uneven distribution of users and waste of computing resources in multiple cross-physical domain computing centers.

[0008] Fourth, the two resource scheduling methods mentioned above often directly assign tasks to suitable computing nodes, resulting in weak security of user data and low privacy of business operations. Summary of the Invention

[0009] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for elastic resource scheduling in computing power networks. Its purpose is to solve the following technical problems: First, existing centralized resource scheduling methods suffer from excessive burden on centralized scheduling nodes, making it difficult to respond promptly to changes in demand and perform elastic resource scheduling when computing nodes fail or workload surges. Second, the lack of further optimization of energy consumption and maintenance costs in resource allocation during scheduling algorithm design leads to high energy consumption and maintenance costs. Third, existing distributed resource scheduling methods struggle to discover computing resource nodes within other local area networks, resulting in uneven user distribution and wasted computing resources across multiple computing power centers in different physical domains. Fourth, directly assigning tasks to suitable computing nodes compromises user data security and business privacy.

[0010] To achieve the above objectives, according to one aspect of the present invention, an elastic resource scheduling method for computing power networks is provided, comprising the following steps:

[0011] (1) The computing power network obtains statistical information of each computing power center in the computing power network and the channel status of the communication network between any two computing power centers through the routing layer.

[0012] (2) The computing power network calculates the service supply pressure of each computing power center based on the statistical information of each computing power center obtained in step (1), and classifies the computing power center according to the service supply pressure, that is, classifying it as a resource-scarce party or a resource-sufficient party.

[0013] (3) The computing power network inputs the statistical information of the resource-scarce party and the resource-sufficient party obtained in step (2) and the channel status of the communication network between any two computing power centers obtained in step (1) into the pre-constructed cross-regional queue model to obtain the global cross-regional queue set of the computing power network.

[0014] (4) Obtain the first cross-regional queue in the global cross-regional queue set obtained in step (3) and connect all the computing centers contained in the first cross-regional queue to each other through the wide area network in a cross-domain interconnection manner, so as to schedule the computing resources of all computing centers in the first cross-regional queue together in the cloud manner for use by the task queue of the resource-scarce computing center in the first cross-regional queue.

[0015] (5) For each remaining cross-region queue in the global cross-region queue set, repeat step (4) above until all remaining cross-region queues in the global cross-region queue set have been processed.

[0016] Preferably, the i-th computing center The statistics include the resource usage of the computing center. Task queue Task queue The j-th task Data volume Resource requirements Submission time Time occupied Start time End time Maximum execution latency and execution status (like ,but and Both are -1, if ,but The time recorded during task deployment, and has And the energy consumption of a unit computing resource in a computing center per unit time for computation. Where i∈[1, the total number of computing centers W in the computing power network], j∈[1, the task queue] [Total number of tasks in];

[0017] The channel status of the communication network between computing centers includes channel bandwidth B and signal-to-noise power. Energy consumption per unit of data transmitted through the channel Channel gain And factors affecting the channel transmission process, I.

[0018] Preferably, step (2) includes the following sub-steps:

[0019] (2-1) Initialize i=1;

[0020] (2-2) Based on the i-th computing center Task queue The execution status of each task is recorded, and the task is placed in either the execution queue or the waiting queue. The j-th task Execution status Then the task Add to the execution queue, if the task queue The j-th task Execution status Then the task Add to the waiting queue;

[0021] (2-3) Check if the waiting queue is empty. If it is, proceed directly to step (2-7). Otherwise, initialize the i-th computing center. Real-time available resources and set the current time. Then proceed to steps (2-4);

[0022] (2-4) Traverse all tasks in the execution queue to obtain the i-th computing center. The task with the shortest completion time Where p is the task with the shortest completion time in the task queue. The sequence number in the current time Set as the task with the smallest end time End time Remove the task with the shortest completion time from the execution queue and set... Then proceed to steps (2-5);

[0023] (2-5) Iterate through all tasks in the waiting queue to find the task with the longest submission time. Where q is the task with the longest submission time in the task queue. The sequence number in the database determines the real-time availability of resources. Is it greater than or equal to the task? resource requirements If so, then set , , Move the task with the longest submission time from the waiting queue to the execution queue and proceed to step (2-6); otherwise, return to step (2-4).

[0024] (2-6) Determine if the waiting queue is empty. If it is, proceed directly to step (2-7); otherwise, return to step (2-4).

[0025] (2-7) The i-th computing center To obtain its own level of service supply and judge Is it greater than the preset threshold? If so, then the i-th computing center Mark the party as having scarce resources, then proceed to step (2-8); otherwise, continue with the judgment. If the value is 0, then the i-th computing center will be... Mark the party as having sufficient resources and proceed to step (2-8); otherwise, proceed directly to step (2-8).

[0026] (2-8) Determine if i is equal to the total number W of computing centers in the computing power network. If yes, the process ends; otherwise, set... Then return to step (2-2);

[0027] Preferably, step (3) includes the following sub-steps:

[0028] (3-1) Select the computing center with the greatest service supply pressure from the resource-scarce parties obtained in step (2). And traverse the computing centers in areas with scarce resources. The task queue is used to move all tasks whose end time is greater than the maximum execution delay to the pending task queue, in order to obtain the computing power of resource-scarce computing centers. The queue of tasks to be executed, where u is the sequence number of the computing center with the greatest service supply pressure in the computing network;

[0029] (3-2) Based on the resource-scarce computing center obtained in step (3-1) The task queue to be executed and the resource-sufficient parties initialize the particle swarm and its parameters, and assign fitness to each particle in the particle swarm. and fitness Perform initialization to obtain the initialized particle swarm;

[0030] (3-3) Set the initial state of the external archive set to an empty set, update the external archive set according to the particle swarm obtained in step (3-2), initialize the particle crowding distance of the particles in the external archive set, and initialize the iterator t=1;

[0031] (3-4) Set t=t+1, and assign the position encoding vector and fitness of each particle in the particle swarm initialized in step (3-2). and fitness Perform an update and utilize the updated fitness of each particle. and fitness Update the external archive set to obtain the updated external archive set;

[0032] (3-5) Determine whether the number of particles in the external archive set is greater than or equal to the preset maximum number of particles H. If yes, proceed to step (3-6); otherwise, return to step (3-4).

[0033] (3-6) Determine whether t has reached the preset maximum number of iterations, or whether the number of particles in the external archive set no longer changes. If so, proceed to step (3-7); otherwise, return to step (3-4).

[0034] (3-7) Obtain the particle with the most balanced fitness in the external archive set, and decode the position encoding vector of the particle to obtain the computing power center containing the resource-scarce side. A global cross-regional queue collection.

[0035] (3-8) The computing power center containing resource-scarce parties obtained in step (3-7) The cross-regional queues are stored in the global cross-regional queue set of the computing power network (which is initially empty), and will be used by computing power centers in resource-scarce areas. Remove from the resource-scarce side and determine whether there are still computing centers in the resource-scarce side. If so, return to step (3-1); otherwise, the process ends, thus obtaining the global cross-regional queue set of the computing network.

[0036] Preferably, the fitness of the particles in step (3-2) The specific calculation process first involves creating a particle swarm based on the preset total number of particles R in the cross-regional queue model. Each particle in the swarm represents a cross-regional queue scheme. Then, a position encoding vector is generated for each particle in the swarm, with the encoding length set to the resource-scarce computing center. The total number of tasks in the task queue is N, and each value of the position encoding vector is a random integer, ranging from [1 to the total number of computing centers M in the resource-sufficient area]. Then, the update rate of each particle in the particle swarm is initialized to 0. Finally, each particle in the particle swarm is decoded and its fitness is calculated. and fitness ;

[0037] Particle fitness The calculation formula is as follows:

[0038]

[0039] in, The k-th value of the position encoding vector x of the particles in the particle swarm (representing the resource-scarce computing center) The k-th task in the pending task queue should be deployed to the resource-sufficient party. (one computing center) For the resource-sufficient party The serial number of each computing center in the computing network (where) ∈[1, M]), For resource-scarce computing centers In the computing power network, the first One computing center Energy consumption for transmitting a unit of data over a channel. For resource-scarce computing centers The k-th task in the queue of tasks to be executed is located in the resource-scarce computing center. The sequence number in the task queue. For resource-scarce computing centers The amount of data for the k-th task in the queue of tasks to be executed. For the first in the computing power network One computing center The energy consumption of a unit of computing resources per unit of time for computation. For resource-scarce computing centers The resource requirements of the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The time taken by the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The data is transmitted to the computing network. One computing center Data transmission rate, Calculations were performed using Shannon's formula:

[0040]

[0041] in, For resource-scarce computing centers In the computing power network, the first One computing center Channel gain between For resource-scarce computing centers In the computing power network, the first One computing center Data transmission between channels is affected by factors in the surrounding environment.

[0042] Particle fitness The specific calculation process is as follows: each value in the particle's position encoding vector is placed into a set, duplicate values ​​are filtered out, and the particle's fitness is... The value is the total number of values ​​in the set.

[0043] Preferably, step (3-3) specifically involves:

[0044] First, set the external archive set to an empty set;

[0045] Then, obtain the first particle from the particle swarm initialized in step (3-2) and determine the fitness of that particle. and fitness If any of the particles in the particle swarm is less than the corresponding value of the remaining particles in the external archive set, then the first particle is placed into the external archive set; otherwise, the first particle is returned to the particle swarm. Then, for each remaining unprocessed particle in the particle swarm, the above process is repeated until all particles have been processed, thus completing the filling of the external archive set.

[0046] Then, all particles in the external archive are sorted according to their fitness. Sort the particles from smallest to largest and calculate the particle crowding distance for each particle in the external archive set. The particle with the smallest particle crowding distance is taken as the globally optimal particle in the t-th iteration.

[0047] The particle crowding distance is obtained using the following formula:

[0048]

[0049] in, The fitness of the external archive set in the t-th iteration The z-th smallest particle, The fitness of the external archive set in the t-th iteration The z-1 smallest particle, The fitness of the external archive set in the t-th iteration The (z+1)th smallest particle Represents the fitness of the external archive set in the t-th iteration. The particle crowding distance of the z-th smallest particle, and z∈[1,R].

[0050] Preferably, step (3-4) specifically involves:

[0051] First, update the position encoding vector of each particle in the particle swarm initialized in step (3-2) according to the following formula;

[0052]

[0053] in, and This represents the update velocity and position encoding vector of the particle at the t-th iteration. This represents the position encoding vector of the historical best individual of the particle at the t-th iteration. If the particle has not entered the external archive set, then the historical best individual is the one that maximizes the particle's fitness during its historical iterations. The smallest particle; otherwise, a particle stored in an external archive set. Let represent the position encoding vector of the globally optimal particle at the t-th iteration, where r and R are random values ​​within [0,1], and w, c1, and c2 are weight values. Since the position encoding vector values ​​of particles in the particle swarm may be non-integer during the particle swarm iteration update process, the position encoding vector values ​​of particles in the particle swarm are converted to integers according to the rounding principle when actually calculating the fitness (the integer is taken as 1 when it is greater than the total number of computing centers M contained in the resource-sufficient side, and similarly, the integer is taken as M when it is less than 1).

[0054] Then, the fitness of the particle is determined based on the updated position encoding vector of each particle in the particle swarm. and fitness Perform an update to obtain the updated particle swarm (fitness). and fitness The calculation formula is shown in step (3-2));

[0055] Then, the first particle is retrieved from the external archive set, and its fitness is determined. and fitness If any of the first particles in the outer archive is less than the corresponding value of each particle in the updated particle swarm, then the first particle is returned to the outer archive set; otherwise, the first particle is added to the updated particle swarm. Then, for each of the remaining unprocessed particles in the outer archive set, the above process is repeated until all particles have been processed, thus eliminating particles in the outer archive set that no longer meet the Pareto optimality condition.

[0056] Then, the first particle is obtained from the updated particle swarm, and its fitness is determined. and fitness If any one of the particles in the first set is less than the corresponding value of each particle in the outer set, then the first particle is added to the outer set; otherwise, the first particle is returned to the updated particle group. Then, for each remaining unprocessed particle in the updated particle group, the above process is repeated until all particles have been processed, thus completing the filling of the outer set.

[0057] Finally, retrieve the first particle from the external archive set and determine whether the particle satisfies the time delay constraint. (in For resource-scarce computing centers If the maximum execution latency of the k-th task in the queue of pending tasks is found, then the first particle is returned to the external archive set; otherwise, the first particle is added to the updated particle swarm. Then, for each remaining unprocessed particle in the external archive set, the above process is repeated until all particles have been processed, and finally the updated external archive set is obtained.

[0058] Preferably, step (3-7) specifically involves first, setting the fitness of all particles in the external archive set. Normalize and take the average Fitness of all particles Normalize and take the average Then follow the formula Calculate the deviation of all particles, and select the particle with the smallest deviation in the external archive as the particle with the most balanced fitness. Then, decode the position encoding vector of this particle and obtain the k-th value of the position encoding vector. (The corresponding computing center's serial number in the resource-sufficient area) through The decoding mapping is used to determine the index of the computing center in the computing network. Then, the obtained indices are placed into a set (duplicate indices need to be removed). Next, the computing centers corresponding to the indices in the set are placed into a cross-regional queue. Finally, the resource-scarce computing centers selected in step (3-1) are... Add them to the cross-region queue to obtain the final global cross-region queue set.

[0059] According to another aspect of the present invention, an elastic resource scheduling system for computing power networks is provided, comprising:

[0060] The first module, located in the computing power network, is used to obtain statistical information of each computing power center in the computing power network through the routing layer, as well as the channel status of the communication network between any two computing power centers.

[0061] The second module, located in the computing power network, is used to calculate the service supply pressure of each computing power center based on the statistical information of each computing power center obtained from the first module, and to classify the computing power center according to the service supply pressure, that is, to divide it into resource-scarce parties or resource-sufficient parties.

[0062] The third module, located in the computing power network, is used to input the statistical information of resource-scarce and resource-sufficient parties obtained from the second module, as well as the channel status of the communication network between any two computing power centers obtained from the first module, into a pre-built cross-regional queue model to obtain a global cross-regional queue set of the computing power network.

[0063] The fourth module, located in the computing power network, is used to obtain the first cross-regional queue from the global cross-regional queue set obtained by the third module, and connect all computing power centers contained in the first cross-regional queue to each other through a wide area network in a cross-domain interconnection manner, so as to schedule the computing resources of all computing power centers in the first cross-regional queue together in a cloud manner for use by the task queues of the computing power centers with scarce resources in the first cross-regional queue.

[0064] The fifth module, located in the computing network, repeats the fourth module for each remaining cross-regional queue in the global cross-regional queue set until all remaining cross-regional queues in the global cross-regional queue set have been processed.

[0065] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0066] 1. The present invention employs step (2), which calculates the service supply pressure of each computing center and sets a threshold based on actual demand and expected resource scheduling effect. When the service supply pressure of a computing center exceeds the threshold, it is classified as a service resource shortage party and triggers elastic resource scheduling. This allows the computing network to avoid the burden of unnecessary resource scheduling operations on the computing network and respond to demand changes in a timely manner. Therefore, it can solve the technical problem of difficulty in responding to demand changes and performing elastic resource scheduling in a timely manner when computing nodes fail or business volume surges.

[0067] 2. This invention employs steps (3-2) to (3-4), which incentivizes the particle swarm to evolve in a direction that reduces energy consumption and maintenance costs through fitness design. It also expands the solution space of the particle swarm using an external archive set mechanism and Pareto optimal conditions. This allows the resulting cross-regional queues to minimize energy consumption from cross-physical domain data transmission and execution, and minimize the number of computing centers covered in each cross-regional queue, under the constraint of completing all tasks that resource-scarce parties cannot complete on time (tasks with expected start times greater than user latency requirements). This improves user experience and minimizes the cost of cross-domain interconnection maintenance, thus solving the technical problem of high energy consumption and maintenance costs in resource allocation.

[0068] 3. Since the present invention adopts steps (3) to (4), it is based on cross-regional queues and unifies resources across multiple management domains into the same resource pool area through cross-domain interconnection, which reduces user waiting time, improves user experience, balances the resource utilization of each computing center, and indirectly improves the overall resource utilization of the computing network. Therefore, it can solve the technical problems of uneven user distribution and waste of computing resources in multiple computing centers across physical domains.

[0069] 4. Since the present invention adopts steps (4) to (5), the resource scheduling is achieved only through cross-regional queues, rather than directly scheduling user tasks. The scheduling right of the task is still in the hands of the user, which has good data privacy and business security. Therefore, it can solve the technical problem of weak security of user data and weak privacy of business. Attached Figure Description

[0070] Figure 1 This is a flowchart of the elastic resource scheduling method for computing power networks according to the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0072] The basic idea of ​​this invention is as follows: First, the computing power network calculates the service supply pressure of each computing power center. Only when the pressure exceeds a pre-set threshold is the center classified as a service resource shortage party, triggering elastic resource scheduling to respond promptly to changes in demand. Second, a cross-regional queue model (particle swarm optimization model) is used to solve the cross-regional queues of service resource shortage parties. A fitness incentive is designed to encourage the particle swarm to evolve in a direction that reduces energy consumption and maintenance costs. An external archive set mechanism and Pareto optimality conditions are used to expand the solution space of the particle swarm, thereby finding the global optimal solution as much as possible, improving user experience and minimizing the cost of cross-domain interconnection maintenance. Third, based on the cross-regional queues, resources across multiple management domains are unified in the same resource pool area through cross-domain interconnection, thereby balancing the resource utilization of each computing power center. Finally, resource scheduling is achieved only through cross-regional queues, rather than directly scheduling user tasks. Users choose whether to submit tasks across domains, thus ensuring data privacy and business security.

[0073] like Figure 1 As shown, this invention provides an elastic resource scheduling method for computing power networks, comprising the following steps:

[0074] (1) The computing power network obtains statistical information of each computing power center in the computing power network and the channel status of the communication network between any two computing power centers through the routing layer.

[0075] Specifically, the i-th computing center The statistics include the resource usage of the computing center. Task queue Task queue The j-th task Data volume Resource requirements Submission time Time occupied Start time End time Maximum execution latency and execution status (like ,but and Both are -1, if ,but The time recorded during task deployment, and has And the energy consumption of a unit computing resource in a computing center per unit time for computation. Where i∈[1, the total number of computing centers W in the computing power network], j∈[1, the task queue] [Total number of tasks in];

[0076] The channel status of the communication network between computing centers includes channel bandwidth B and signal-to-noise power. Energy consumption per unit of data transmitted through the channel Channel gain And factors affecting the channel transmission process, I.

[0077] (2) The computing power network calculates the service supply pressure of each computing power center based on the statistical information of each computing power center obtained in step (1), and classifies the computing power center according to the service supply pressure, that is, classifying it as a resource-scarce party or a resource-sufficient party.

[0078] Step (2) includes the following sub-steps:

[0079] (2-1) Initialize i=1;

[0080] (2-2) Based on the i-th computing center Task queue The execution status of each task is recorded, and the task is placed in either the execution queue or the waiting queue. The j-th task Execution status Then the task Add to the execution queue, if the task queue The j-th task Execution status Then the task Add to the waiting queue;

[0081] (2-3) Check if the waiting queue is empty. If it is, proceed directly to step (2-7). Otherwise, initialize the i-th computing center. Real-time available resources and set the current time. Then proceed to steps (2-4);

[0082] (2-4) Traverse all tasks in the execution queue to obtain the i-th computing center. The task with the shortest completion time Where p is the task with the shortest completion time in the task queue. The sequence number in the current time Set as the task with the smallest end time End time Remove the task with the shortest completion time from the execution queue and set... Then proceed to steps (2-5);

[0083] (2-5) Iterate through all tasks in the waiting queue to find the task with the longest submission time. Where q is the task with the longest submission time in the task queue. The sequence number in the database determines the real-time availability of resources. Is it greater than or equal to the task? resource requirements If so, then set , , Move the task with the longest submission time from the waiting queue to the execution queue and proceed to step (2-6); otherwise, return to step (2-4).

[0084] This sub-step, based on the principle of fairness, deploys service resources according to the order of task submission time, which can improve user experience and avoid situations where tasks with high resource requirements are constantly preempted by tasks with low resource requirements, resulting in them being unable to obtain service resources.

[0085] (2-6) Determine if the waiting queue is empty. If it is, proceed directly to step (2-7); otherwise, return to step (2-4).

[0086] (2-7) The i-th computing center To obtain its own level of service supply and judge Is it greater than the preset threshold? If so, then the i-th computing center Mark the party as having scarce resources, then proceed to step (2-8); otherwise, continue with the judgment. If the value is 0, then the i-th computing center will be... Mark the party as having sufficient resources and proceed to step (2-8); otherwise, do nothing and proceed directly to step (2-8).

[0087] The formula for calculating the level of service availability in this sub-step is as follows: .

[0088] (2-8) Determine if i is equal to the total number W of computing centers in the computing power network. If yes, the process ends; otherwise, set... Then return to step (2-2);

[0089] The advantage of this step is that only when the service supply pressure of the computing center exceeds the pre-set threshold will it be classified as a service resource shortage party and trigger elastic resource scheduling. This allows the computing network to avoid the burden of unnecessary resource scheduling operations and respond to changes in demand in a timely manner.

[0090] (3) The computing power network inputs the statistical information of the resource-scarce party and the resource-sufficient party obtained in step (2) and the channel status of the communication network between any two computing power centers obtained in step (1) into the pre-constructed cross-regional queue model to obtain the global cross-regional queue set of the computing power network.

[0091] Specifically, the cross-regional queue model is a particle swarm algorithm model. The state of each particle in the particle swarm is described by a set of position encoding vectors and velocity vectors, which correspond to the feasible solutions to the problem and their direction of motion in the search space in the parameter space. After multiple iterations, the particles in the particle swarm continuously learn the globally optimal solution they discover and the historical individual optimal solutions to achieve a globally optimal search, and finally output the optimal solution found, which is the optimal cross-regional queue.

[0092] Step (3) includes the following sub-steps:

[0093] (3-1) Select the computing center with the greatest service supply pressure from the resource-scarce parties obtained in step (2). And traverse the computing centers in areas with scarce resources. The task queue is used to move all tasks whose end time is greater than the maximum execution delay to the pending task queue, in order to obtain the computing power of resource-scarce computing centers. The queue of tasks to be executed, where u is the sequence number of the computing center with the greatest service supply pressure in the computing network;

[0094] Specifically, the computing power center of the resource-scarce side in this step refers to the computing power center of the resource-scarce side; similarly, the computing power center of the resource-sufficient side mentioned in this article refers to the computing power center of the resource-sufficient side.

[0095] (3-2) Based on the resource-scarce computing center obtained in step (3-1) The task queue to be executed and the resource-sufficient parties initialize the particle swarm and its parameters, and assign fitness to each particle in the particle swarm. and fitness Perform initialization to obtain the initialized particle swarm;

[0096] Specifically, this step involves first creating a particle swarm based on the preset total number of particles R in the cross-regional queue model, where each particle represents a cross-regional queue scheme; then generating a position encoding vector for each particle in the particle swarm, with the encoding length set to the resource-scarce computing center. The total number of tasks in the task queue is N, and each value of the position encoding vector is a random integer, ranging from [1 to the total number of computing centers M in the resource-sufficient area]. Then, the update rate of each particle in the particle swarm is initialized to 0. Finally, each particle in the particle swarm is decoded and its fitness is calculated. and fitness ;

[0097] Particle fitness The calculation formula is as follows:

[0098]

[0099] in, The k-th value of the position encoding vector x of the particles in the particle swarm (representing the resource-scarce computing center) The k-th task in the pending task queue should be deployed to the resource-sufficient party. (one computing center) For the resource-sufficient party The serial number of each computing center in the computing network (where) ∈[1, M]), For resource-scarce computing centers In the computing power network, the first One computing center Energy consumption for transmitting a unit of data over a channel. For resource-scarce computing centers The k-th task in the queue of tasks to be executed is located in the resource-scarce computing center. The sequence number in the task queue. For resource-scarce computing centers The amount of data for the k-th task in the queue of tasks to be executed. For the first in the computing power network One computing center The energy consumption of a unit of computing resources per unit of time for computation. For resource-scarce computing centers The resource requirements of the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The time taken by the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The data is transmitted to the computing network. One computing center Data transmission rate, Calculations were performed using Shannon's formula:

[0100]

[0101] in, For resource-scarce computing centers In the computing power network, the first One computing center Channel gain between For resource-scarce computing centers In the computing power network, the first One computing center Data transmission between channels is affected by factors in the surrounding environment.

[0102] Particle fitness The specific calculation process is as follows: each value in the particle's position encoding vector is placed into a set, duplicate values ​​are filtered out, and the particle's fitness is... The value is the total number of values ​​in the set;

[0103] (3-3) Set the initial state of the external archive set to an empty set, update the external archive set according to the particle swarm obtained in step (3-2), initialize the particle crowding distance of the particles in the external archive set, and initialize the iterator t=1;

[0104] Specifically, this step involves first setting the external archive set to an empty set;

[0105] Then, obtain the first particle from the particle swarm initialized in step (3-2) and determine the fitness of that particle. and fitness If any of the particles in the particle swarm is less than the corresponding value of the remaining particles in the external archive set, then the first particle is placed into the external archive set; otherwise, the first particle is returned to the particle swarm. Then, for each remaining unprocessed particle in the particle swarm, the above process is repeated until all particles have been processed, thus completing the filling of the external archive set.

[0106] Then, all particles in the external archive are sorted according to their fitness. Sort the particles from smallest to largest and calculate the particle crowding distance for each particle in the external archive set. The particle with the smallest particle crowding distance is taken as the globally optimal particle in the t-th iteration.

[0107] The particle crowding distance is obtained using the following formula:

[0108]

[0109] in, The fitness of the external archive set in the t-th iteration The z-th smallest particle, The fitness of the external archive set in the t-th iteration The z-1 smallest particle, The fitness of the external archive set in the t-th iteration The (z+1)th smallest particle Represents the fitness of the external archive set in the t-th iteration. The particle crowding distance of the z-th smallest particle, and z∈[1,R].

[0110] (3-4) Set t=t+1, and assign the position encoding vector and fitness of each particle in the particle swarm initialized in step (3-2). and fitness Perform an update and utilize the updated fitness of each particle. and fitness Update the external archive set to obtain the updated external archive set;

[0111] Specifically, this step involves updating the position encoding vector of each particle in the particle swarm initialized in step (3-2) according to the following formula;

[0112] The specific formula for updating the position encoding vector of a particle in a particle swarm is as follows:

[0113]

[0114] in, and This represents the update velocity and position encoding vector of the particle at the t-th iteration. This represents the position encoding vector of the historical best individual of the particle at the t-th iteration. If the particle has not entered the external archive set, then the historical best individual is the one that maximizes the particle's fitness during its historical iterations. The smallest particle; otherwise, a particle stored in an external archive set. Let represent the position encoding vector of the globally optimal particle at the t-th iteration, where r and R are random values ​​within [0,1], and w, c1, and c2 are weight values. Since the position encoding vector values ​​of particles in the particle swarm may be non-integer during the particle swarm iteration update process, the position encoding vector values ​​of particles in the particle swarm are converted to integers according to the rounding principle when actually calculating the fitness (the integer is taken as 1 when it is greater than the total number of computing centers M contained in the resource-sufficient side, and similarly, the integer is taken as M when it is less than 1).

[0115] Then, the fitness of the particle is determined based on the updated position encoding vector of each particle in the particle swarm. and fitness Perform an update to obtain the updated particle swarm (fitness). and fitness The calculation formula is shown in step (3-2));

[0116] Then, the first particle is retrieved from the external archive set, and its fitness is determined. and fitness If any of the first particles in the outer archive is less than the corresponding value of each particle in the updated particle swarm, then the first particle is returned to the outer archive set; otherwise, the first particle is added to the updated particle swarm. Then, for each of the remaining unprocessed particles in the outer archive set, the above process is repeated until all particles have been processed, thus eliminating particles in the outer archive set that no longer meet the Pareto optimality condition.

[0117] Then, the first particle is obtained from the updated particle swarm, and its fitness is determined. and fitness If any one of the particles in the first set is less than the corresponding value of each particle in the outer set, then the first particle is added to the outer set; otherwise, the first particle is returned to the updated particle group. Then, for each remaining unprocessed particle in the updated particle group, the above process is repeated until all particles have been processed, thus completing the filling of the outer set.

[0118] Finally, retrieve the first particle from the external archive set and determine whether the particle satisfies the time delay constraint. (in For resource-scarce computing centers If the maximum delay of the execution of the k-th task in the queue of tasks to be executed is found, then the first particle is put back into the external archive set; otherwise, the first particle is put into the updated particle swarm. Then, for each remaining unprocessed particle in the external archive set, the above process is repeated until all particles are processed, and finally the updated external archive set is obtained.

[0119] The advantages of steps (3-2) to (3-4) above are that the fitness of the design can incentivize the particle swarm to evolve in the desired direction of reducing energy consumption and maintenance costs, and the external archive set mechanism and Pareto optimal conditions are used to expand the solution space of the particle swarm. This allows the final cross-regional queues to minimize the energy consumption of cross-physical domain data transmission and execution, and minimize the computing power centers covered in each cross-regional queue, under the constraint of completing all tasks that the resource-scarce parties cannot complete on time (tasks whose expected start time is greater than the user's latency requirement) on schedule. This achieves the effect of improving user experience and minimizing the cost required for cross-domain interconnect maintenance.

[0120] (3-5) Determine whether the number of particles in the external archive set is greater than or equal to the preset maximum number of particles H. If yes, proceed to step (3-6); otherwise, return to step (3-4).

[0121] Specifically, the preset maximum number of particles H ranges from 20 to 50, preferably 35;

[0122] (3-6) Determine whether t has reached the preset maximum number of iterations, or whether the number of particles in the external archive set no longer changes. If so, proceed to step (3-7); otherwise, return to step (3-4).

[0123] Specifically, the preset maximum number of iterations ranges from 100 to 4000, preferably 600;

[0124] (3-7) Obtain the particle with the most balanced fitness in the external archive set, and decode the position encoding vector of the particle to obtain the computing power center containing the resource-scarce side. A global cross-regional queue collection.

[0125] Specifically, this step involves first setting the fitness of all particles in the external archive set. Normalize and take the average Fitness of all particles Normalize and take the average Then follow the formula Calculate the deviation of all particles, and select the particle with the smallest deviation in the external archive as the particle with the most balanced fitness. Then, decode the position encoding vector of this particle and obtain the k-th value of the position encoding vector. (The corresponding computing center's serial number in the resource-sufficient area) through The decoding mapping is used to determine the index of the computing center in the computing network. Then, the obtained indices are placed into a set (duplicate indices need to be removed). Next, the computing centers corresponding to the indices in the set are placed into a cross-regional queue. Finally, the resource-scarce computing centers selected in step (3-1) are... Add them to the cross-region queue to obtain the final global cross-region queue set.

[0126] For example, the selected computing center for resource-scarce areas is The particle's position encoding vector is {1,1,3,2}, and the decoded index is {5,5,9,7}. After removing duplicate values, the resulting set is {5,7,9}. Therefore, the final global cross-region queue set is [...]. , , , ].

[0127] (3-8) The computing power center containing resource-scarce parties obtained in step (3-7) The cross-regional queues are stored in the global cross-regional queue set of the computing power network (which is initially empty), and will be used by computing power centers in resource-scarce areas. Remove from the resource-scarce side and determine whether there are still computing centers in the resource-scarce side. If so, return to step (3-1); otherwise, the process ends, thus obtaining the global cross-regional queue set of the computing network.

[0128] (4) Obtain the first cross-regional queue in the global cross-regional queue set obtained in step (3) and connect all the computing centers contained in the first cross-regional queue to each other through the wide area network in a cross-domain interconnection manner, so as to schedule the computing resources of all computing centers in the first cross-regional queue together in the cloud manner for use by the task queue of the resource-scarce computing center in the first cross-regional queue.

[0129] The advantages of steps (3) to (4) above are that, based on cross-regional queues, resources across multiple management domains are unified in the same resource pool area through cross-domain interconnection, which reduces user waiting time, improves user experience, balances the resource utilization of each computing center, and indirectly improves the overall resource utilization of the computing network.

[0130] (5) For each remaining cross-region queue in the global cross-region queue set, repeat step (4) above until all remaining cross-region queues in the global cross-region queue set have been processed.

[0131] The advantage of steps (4) to (5) above is that it only achieves resource scheduling through cross-regional queues, rather than directly scheduling user tasks. The scheduling right of the task is still in the hands of the user, which has better data privacy and business security.

[0132] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for elastic resource scheduling in computing power networks, characterized in that, Includes the following steps: (1) The computing power network obtains the statistical information of each computing power center in the computing power network and the channel status of the communication network between any two computing power centers through the routing layer; (2) The computing power network calculates the service supply pressure of each computing power center based on the statistical information of each computing power center obtained in step (1), and classifies the computing power center according to the service supply pressure, that is, classifying it as a resource-scarce party or a resource-sufficient party. (3) The computing power network inputs the statistical information of resource-scarce and resource-sufficient parties obtained in step (2) and the channel status of the communication network between any two computing power centers obtained in step (1) into the pre-constructed cross-regional queue model to obtain the global cross-regional queue set of the computing power network; step (3) includes the following sub-steps: (3-1) Select the computing center with the greatest service supply pressure from the resource-scarce parties obtained in step (2). And traverse the computing centers in areas with scarce resources. The task queue is used to move all tasks whose end time is greater than the maximum execution delay to the pending task queue, in order to obtain the computing power of resource-scarce computing centers. The queue of tasks to be executed, where u is the sequence number of the computing center with the greatest service supply pressure in the computing network; (3-2) Based on the resource-scarce computing center obtained in step (3-1) The task queue to be executed and the resource-sufficient parties initialize the particle swarm and its parameters, and assign fitness to each particle in the particle swarm. and fitness Perform initialization to obtain an initialized particle swarm; the fitness of the particles in step (3-2) The specific calculation process first involves creating a particle swarm based on the preset total number of particles R in the cross-regional queue model. Each particle in the swarm represents a cross-regional queue scheme. Then, a position encoding vector is generated for each particle in the swarm, with the encoding length set to the resource-scarce computing center. The total number of tasks in the task queue is N, and each value of the position encoding vector is a random integer, ranging from [1 to the total number of computing centers M in the resource-sufficient area]. Then, the update rate of each particle in the particle swarm is initialized to 0. Finally, each particle in the particle swarm is decoded and its fitness is calculated. and fitness ; Particle fitness The calculation formula is as follows: ； in, The k-th value of the position encoding vector x of a particle in the particle swarm represents the computing power center of the resource-scarce party. The k-th task in the pending task queue should be deployed to the resource-sufficient party. One computing center, For the resource-sufficient party The serial number of each computing center in the computing network, among which ∈[1, M], For resource-scarce computing centers In the computing power network, the first One computing center Energy consumption for transmitting a unit of data over a channel. For resource-scarce computing centers The k-th task in the queue of tasks to be executed is located in the resource-scarce computing center. The sequence number in the task queue. For resource-scarce computing centers The amount of data for the k-th task in the queue of tasks to be executed. For the first in the computing power network One computing center The energy consumption of a unit of computing resources per unit of time for computation. For resource-scarce computing centers The resource requirements of the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The time taken by the k-th task in the queue of tasks to be executed. For resource-scarce computing centers The data is transmitted to the computing network. One computing center Data transmission rate, Calculations were performed using Shannon's formula: ； in, For resource-scarce computing centers In the computing power network, the first One computing center Channel gain between For resource-scarce computing centers In the computing power network, the first One computing center Data transmission between channels is affected by factors in the surrounding environment; Particle fitness The specific calculation process is as follows: each value in the particle's position encoding vector is placed into a set, duplicate values ​​are filtered out, and the particle's fitness is... The value is the total number of values ​​in the set; (3-3) Set the initial state of the external archive set to an empty set, update the external archive set according to the particle swarm obtained in step (3-2), initialize the particle crowding distance of the particles in the external archive set, and initialize the iterator t=1; (3-4) Set t=t+1, and assign the position encoding vector and fitness of each particle in the particle swarm initialized in step (3-2). and fitness Perform an update and utilize the updated fitness of each particle. and fitness Update the external archive set to obtain the updated external archive set; (3-5) Determine whether the number of particles in the external archive set is greater than or equal to the preset maximum number of particles H. If yes, proceed to step (3-6); otherwise, return to step (3-4). (3-6) Determine whether t has reached the preset maximum number of iterations, or whether the number of particles in the external archive set no longer changes. If so, proceed to step (3-7); otherwise, return to step (3-4). (3-7) Obtain the particle with the most balanced fitness in the external archive set, and decode the position encoding vector of the particle to obtain the computing power center containing the resource-scarce side. A global cross-region queue collection; (3-8) The computing power center containing resource-scarce parties obtained in step (3-7) The cross-regional queues are stored in a global cross-regional queue set within the computing power network. This global cross-regional queue set is initially empty, and then the computing power centers of resource-scarce regions are included. Remove from the resource-scarce side and determine whether there are still computing centers in the resource-scarce side. If so, return to step (3-1); otherwise, the process ends, thus obtaining the global cross-regional queue set of the computing network. (4) Obtain the first cross-regional queue in the global cross-regional queue set obtained in step (3) and connect all the computing centers contained in the first cross-regional queue to each other through the wide area network in a cross-domain interconnection manner, so as to schedule the computing resources of all computing centers in the first cross-regional queue together in the cloud manner for use by the task queue of the resource-scarce computing center in the first cross-regional queue. (5) For each remaining cross-region queue in the global cross-region queue set, repeat step (4) above until all remaining cross-region queues in the global cross-region queue set have been processed.

2. The elastic resource scheduling method for computing power networks according to claim 1, characterized in that, The i-th computing center The statistics include the resource usage of the computing center. Task queue Task queue The j-th task Data volume Resource requirements Submission time Time occupied Start time End time Maximum execution latency and execution status And the energy consumption of a unit computing resource in a computing center per unit time for computation. Where i∈[1, the total number of computing centers W in the computing power network], j∈[1, the task queue] [Total number of tasks in], if ,but and Both are -1, if ,but The time recorded during task deployment, and has ; The channel status of the communication network between computing centers includes channel bandwidth B and signal-to-noise power. Energy consumption per unit of data transmitted through the channel Channel gain And factors affecting the channel transmission process, I.

3. The elastic resource scheduling method for computing power networks according to claim 1 or 2, characterized in that, Step (2) includes the following sub-steps: (2-1) Initialize i=1; (2-2) Based on the i-th computing center Task queue The execution status of each task is recorded, and the task is placed in either the execution queue or the waiting queue. The j-th task Execution status Then the task Add to the execution queue, if the task queue The j-th task Execution status Then the task Add to the waiting queue; (2-3) Check if the waiting queue is empty. If it is, proceed directly to step (2-7). Otherwise, initialize the i-th computing center. Real-time available resources and set the current time. Then proceed to steps (2-4); (2-4) Traverse all tasks in the execution queue to obtain the i-th computing center. The task with the shortest completion time Where p is the task with the shortest completion time in the task queue. The sequence number in the current time Set as the task with the smallest end time End time Remove the task with the shortest completion time from the execution queue and set... Then proceed to steps (2-5); (2-5) Iterate through all tasks in the waiting queue to find the task with the longest submission time. Where q is the task with the longest submission time in the task queue. The sequence number in the database determines the real-time availability of resources. Is it greater than or equal to the task? resource requirements If so, then set , , Move the task with the longest submission time from the waiting queue to the execution queue and proceed to step (2-6); otherwise, return to step (2-4). (2-6) Determine if the waiting queue is empty. If it is, proceed directly to step (2-7); otherwise, return to step (2-4). (2-7) The i-th computing center To obtain its own level of service supply and judge Is it greater than the preset threshold? If so, then the i-th computing center Mark the party as having scarce resources, then proceed to step (2-8); otherwise, continue with the judgment. If the value is 0, then the i-th computing center will be... Mark the party as having sufficient resources and proceed to step (2-8); otherwise, proceed directly to step (2-8). (2-8) Determine if i is equal to the total number W of computing centers in the computing power network. If yes, the process ends; otherwise, set... Then return to step (2-2).

4. The elastic resource scheduling method for computing power networks according to claim 3, characterized in that, Step (3-3) specifically involves: First, set the external archive set to an empty set; Then, obtain the first particle from the particle swarm initialized in step (3-2) and determine the fitness of that particle. and fitness If any of the particles in the particle swarm is less than the corresponding value of the remaining particles in the external archive set, then the first particle is placed into the external archive set; otherwise, the first particle is returned to the particle swarm. Then, for each remaining unprocessed particle in the particle swarm, the above process is repeated until all particles have been processed, thus completing the filling of the external archive set. Then, all particles in the external archive are sorted according to their fitness. Sort the particles from smallest to largest and calculate the particle crowding distance for each particle in the external archive set. The particle with the smallest particle crowding distance is taken as the globally optimal particle in the t-th iteration. The particle crowding distance is obtained using the following formula: ； in, The fitness of the external archive set in the t-th iteration The z-th smallest particle, The fitness of the external archive set in the t-th iteration The z-1 smallest particle, The fitness of the external archive set in the t-th iteration The (z+1)th smallest particle, Represents the fitness of the external archive set in the t-th iteration. The particle crowding distance of the z-th smallest particle, and z∈[1,R].

5. The elastic resource scheduling method for computing power networks according to claim 4, characterized in that, Steps (3-4) are as follows: First, update the position encoding vector of each particle in the particle swarm initialized in step (3-2) according to the following formula; ； in, and This represents the update velocity and position encoding vector of the particle at the t-th iteration. This represents the position encoding vector of the historical best individual of the particle at the t-th iteration. If the particle has not entered the external archive set, then the historical best individual is the one that maximizes the particle's fitness during its historical iterations. The smallest particle; otherwise, a particle stored in an external archive set. Let represent the position encoding vector of the globally optimal particle at the t-th iteration, where r and R are random values ​​in [0,1], and w, c1, and c2 are weight values. Since the position encoding vector of a particle in the particle swarm may take non-integer values ​​during the iterative update of the particle swarm, the position encoding vector of a particle in the particle swarm is converted to an integer according to the rounding principle when calculating the fitness. When the integer is greater than the total number of computing centers M contained in the resource-sufficient side, it is taken as 1. Similarly, when the integer is less than 1, it is taken as M. Then, the fitness of the particle is determined based on the updated position encoding vector of each particle in the particle swarm. and fitness Perform an update to obtain the updated particle swarm; Then, the first particle is retrieved from the external archive set, and its fitness is determined. and fitness If any of the first particles in the outer archive is less than the corresponding value of each particle in the updated particle swarm, then the first particle is returned to the outer archive set; otherwise, the first particle is added to the updated particle swarm. Then, for each of the remaining unprocessed particles in the outer archive set, the above process is repeated until all particles have been processed, thus eliminating particles in the outer archive set that no longer meet the Pareto optimality condition. Then, the first particle is obtained from the updated particle swarm, and its fitness is determined. and fitness If any one of the particles in the first set is less than the corresponding value of each particle in the outer set, then the first particle is added to the outer set; otherwise, the first particle is returned to the updated particle group. Then, for each remaining unprocessed particle in the updated particle group, the above process is repeated until all particles have been processed, thus completing the filling of the outer set. Finally, retrieve the first particle from the external archive set and determine whether the particle satisfies the time delay constraint. ,in For resource-scarce computing centers If the maximum delay of the execution of the k-th task in the queue of pending tasks is found, the first particle is put back into the external archive set; otherwise, the first particle is put into the updated particle swarm. Then, for each remaining unprocessed particle in the external archive set, the above process is repeated until all particles have been processed, and finally the updated external archive set is obtained.

6. The elastic resource scheduling method for computing power networks according to claim 5, characterized in that, Steps (3-7) specifically involve first determining the fitness of all particles in the external archive set. Normalize and take the average Fitness of all particles Normalize and take the average Then follow the formula Calculate the deviation of all particles, and select the particle with the smallest deviation in the external archive as the particle with the most balanced fitness. Then, decode the position encoding vector of this particle and obtain the k-th value of the position encoding vector. pass The decoding mapping is used to determine the index of the computing center in the computing network. Then, the multiple indexes are placed into a set. Next, the computing centers corresponding to the indexes in the set are placed into a cross-regional queue. Finally, the resource-scarce computing centers selected in step (3-1) are... Add them to the cross-region queue to obtain the final global cross-region queue set.

7. A flexible resource scheduling system for computing power networks, implemented based on the flexible resource scheduling method for computing power networks as described in claim 1, characterized in that, The elastic resource scheduling system includes: The first module, located in the computing power network, is used to obtain statistical information of each computing power center in the computing power network through the routing layer, as well as the channel status of the communication network between any two computing power centers; The second module, located in the computing power network, is used to calculate the service supply pressure of each computing power center based on the statistical information of each computing power center obtained from the first module, and to classify the computing power center according to the service supply pressure, that is, to divide it into resource-scarce parties or resource-sufficient parties. The third module, located in the computing power network, is used to input the statistical information of resource-scarce and resource-sufficient parties obtained from the second module, as well as the channel status of the communication network between any two computing power centers obtained from the first module, into a pre-built cross-regional queue model to obtain a global cross-regional queue set of the computing power network. The fourth module, located in the computing power network, is used to obtain the first cross-regional queue from the global cross-regional queue set obtained by the third module, and connect all computing power centers contained in the first cross-regional queue to each other through a wide area network in a cross-domain interconnection manner, so as to schedule the computing resources of all computing power centers in the first cross-regional queue together in a cloud manner for use by the task queues of the computing power centers with scarce resources in the first cross-regional queue. The fifth module, located in the computing network, repeats the fourth module for each remaining cross-regional queue in the global cross-regional queue set until all remaining cross-regional queues in the global cross-regional queue set have been processed.

Images