Task pre-migration scheduling method for computing power network

Abstract

The application discloses a task pre-migration scheduling method for a computing power network, and belongs to the field of communication.The method comprises the following steps: obtaining relevant task parameters in the computing power network, and reserving computing power resources in a server network in advance through a task pre-migration mode.Based on the multi-dimensional capabilities of communication, calculation and storage of the computing power server, the total cost of executing tasks in the network is optimized, and a migration decision is made by using a distributed scheduling optimization algorithm.The application can be used for task scheduling decision in the computing power network, and can adjust service arrangement of the computing power network, so as to reduce the total cost of network execution.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and more specifically to a task pre-migration scheduling method for computing networks. Background Technology

[0002] Computing power networks, as a novel network technology solution for addressing the unified supply of multi-level computing resources (cloud computing, edge computing, and endpoint computing), act as a scheduling hub distributing resource information such as computing power, storage, and algorithms from service nodes, thereby achieving efficient resource utilization and orchestration. The core idea of ​​computing power networks is to deeply integrate distributed computing resources, combining them into a new network architecture that unifies the scheduling of multi-dimensional resources such as transmission, storage, and computing.

[0003] Currently, research on computing resources and their scheduling can be broadly divided into two directions: one is to build flexible and agile computing infrastructure, actively introduce technologies such as heterogeneous computing, continuously enhance computing power, and release the value of computing power; the other is to create a more ubiquitous computing power distribution, allowing computing resources to evolve ubiquitously from the cloud to the computing center, from the center to the edge and the end side, and to form a more ubiquitous multi-dimensional computing power layout by developing edge computing and end computing, and to achieve efficient scheduling of computing resources through integrated resource orchestration and multi-cloud collaboration and cloud-edge coordination.

[0004] Task migration addresses the mobility management challenges arising from QoS degradation due to user mobility in edge computing. When a user moves away from their original server, the memory, storage, and other data currently in use on the original server are transferred to the new server. After the transfer is complete, the original server shuts down its virtual machines (VMs), and the new MEC server starts up VMs to continue providing services to the user. While service migration incurs migration costs when tasks are moved between different compute nodes, it brings task execution closer to the user, significantly reducing communication latency between the user and the server.

[0005] Most current research on task-computing resource scheduling focuses on optimizing task offloading strategies, neglecting the advantages and efficiency of distributed computing resource orchestration in computing networks. However, most computation migration applications involve passive task migration when users leave the server's execution range. If task prediction and service pre-migration could be performed, integrating different distributed computing resources across the network to collaboratively complete computational tasks, efficiency could be significantly improved.

[0006] Meanwhile, current task offloading decisions mostly focus on optimizing execution latency and energy consumption, neglecting a comprehensive evaluation of the server. In real-world applications, it's crucial to consider not only task latency and energy consumption optimization but also server storage costs and reliability, among other factors. A multi-dimensional evaluation of the server is essential for achieving optimal offloading decisions. In actual computing scenarios, due to server congestion and other factors, actual latency may not follow the deterministic result of dividing the total number of tasks by the computing capacity, potentially resulting in random computational latency. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the purpose of this invention is to consider the task allocation situation in the current server network, integrate the multi-layered differentiated computing and communication resources in the computing power network according to business needs, provide a scheduling optimization algorithm for task pre-migration decision-making in the computing power network, and design a computing power scheduling algorithm to efficiently service orchestrate computing power demand and computing power resources, thereby improving computing power utilization efficiency.

[0008] A scheduling optimization algorithm for task pre-migrating decision-making in computing power networks, comprising the following steps:

[0009] 1) Obtain N tasks σ = [σ1, σ2, ..., σ] in the computing power network. N The parameters include tasks already executed in the network and tasks about to arrive in the network for unloading and execution. Task parameters include: task σ k Data volume l k Total task execution delay constraint d k Obtain the M servers E = [E1, E2, ..., E] in the computing power network. M The parameters include each server E i computing power f i Calculate power storage space Server E i and E j Inter-task migration transmission bandwidth Migration power

[0010] 2) Initialize the Q-value table Q(α(0),β(0)) as a zero matrix.

[0011] 3) Obtain the original unloading state α(t) of the task already in the network at time t = {α ik (t)} M×N α ik (t) represents the task σ at time t. k Unloaded to server E i The proportion of subtasks. For the task σ currently being executed on the network. kα(t) satisfies Arrival probability η k (t) = 1. For tasks that are not computed within the network, σ... k' α(t) satisfies α ik' (t) = 0, arrival probability η k' (t)∈[0,1].

[0012] 4) At time t, perform pre-migration on tasks already executed in the network β(t)={β ijk (t)} M×M×N ,β ijk (t) represents the task σ at time t. k From server E i Transmitted to server E j The proportion of subtasks. For the task σ currently being executed on the network. k The pre-migration action satisfies Let β be the set of all possible pre-transfer actions for the task. With probability ∈ , randomly select action β(t) at time t from action set β, and simultaneously select the optimal action β(t) = arg max with probability (1-∈ ). β∈β Q(α(t),β).

[0013] 5) At time t, perform the task pre-migration action β(t) and update the task unloading state α(t+1) = α(t) + β(t), where α(t) and β(t) represent the task unloading state and task pre-migration action at time t. For tasks newly arriving in the network at time t+1, unload them to the server with the lowest cost based on the server's idle state to perform calculations, and update α(t+1) simultaneously.

[0014] 6) Consider the multi-dimensional costs of computation, communication, and storage in task execution within a computing network:

[0015] For time t, the k-th task is on server E i The computation time required to execute the subtasks of loading and unloading.

[0016]

[0017] Where, θ ik (t)=α ik (t)·l k / f i , Γ k It is task σ k The computational reliability requirements.

[0018] The migration cost for all tasks in the computing power network is:

[0019]

[0020] in, It is the migration delay generated during the task pre-migration at time t.

[0021] The execution cost of all tasks in the computing power network at time t is:

[0022]

[0023] Where, ξ i For server E i Storage unit cost. Optimize server task offloading decisions with the goal of minimizing total execution cost.

[0024] 7) Calculate the cost reduction benefit R. The specific calculation formula is as follows:

[0025] 7.1) Due to the pre-migration action β(t), some computing resources in the computing network are reserved in advance. Therefore, tasks newly arriving in the network at time t+1 will be unloaded to the server with the lowest cost based on the server's idle status to perform computation and obtain the updated α. new (t+1). In contrast, when no pre-migration action is performed, the task unloading status is updated to α based on the idle computing resources at time t and the newly arrived tasks at time t+1. 0 (t+1).

[0026] The benefits of pre-migration are:

[0027]

[0028] in:

[0029] The execution of the pre-migration action β(t) results in a change ΔE in the total cost of the original task. cost (t);

[0030] ΔE cost (t)=E cost (α ori (t+1))-E cost (α(t));

[0031] The migration cost E generated by the pre-migration action β(t) mig (t);

[0032] The execution of the pre-migration action β(t) results in a reduction in execution cost E for the new task at arrival probability η. gain (α(t+1))E gain (α(t+1))=E cost (α new (t+1))-E cost (α 0(t+1));

[0033] 7.2) If the computation delay of the task exceeds the delay limit after performing the pre-transfer action. Or the server's task offload volume exceeds its maximum storage capacity. If this does not meet the network requirements, then the profit R = 0.

[0034] 8) Update the Q value. The specific calculation formula is as follows:

[0035]

[0036] Where the discount revenue V(α(t)) = argmax β Q(α(t),β(t)); γ is the learning rate, representing how fast learning occurs; γ is the discount factor.

[0037] 9) Repeat steps 4-8 at each time step, adjusting the original task unloading ratio to reserve computing resources for upcoming tasks. The unloading status at the next time step is updated to α(t+1)=α ori (t+1)+α new (t+1).

[0038] The beneficial effects of this invention are:

[0039] This invention focuses on the task-computing resource relationship in computing power networks. By obtaining the number of tasks, computational load, and latency constraints that are about to arrive in the network in advance, and by performing task pre-migration, the proportion of task offloading in the network is adjusted, thereby steadily reducing the total execution cost of the network. Attached Figure Description

[0040] Figure 1 This paper compares the performance of the present invention with two comparative algorithms (random allocation and greedy algorithm) under different task numbers, with fixed server parameters when the network is relatively idle (total number of tasks / maximum network storage < 60%).

[0041] Figure 2 This paper compares the performance of the present invention with two comparative algorithms (random allocation and greedy algorithm) under different task numbers, with fixed server parameters and a relatively congested network (total task volume / maximum network storage > 80%). Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the present invention.

[0043] A scheduling optimization algorithm for task pre-migrating decision-making in computing power networks, comprising the following steps:

[0044] 1) Obtain N tasks σ = [σ1, σ2, ..., σ] in the computing power network. N The parameters include tasks already executed in the network and tasks about to arrive in the network for unloading and execution. Task parameters include: task σ k Data volume l k Total task execution delay constraint d k Obtain the M servers E = [E1, E2, ..., E] in the computing power network. M The parameters include each server E i computing power f i Calculate power storage space Server E i and E j Inter-task migration transmission bandwidth Migration power

[0045] 2) Initialize the Q-value table Q(α(0),β(0)) as a zero matrix.

[0046] 3) Obtain the original unloading state α(t) of the task already in the network at time t = {α ik (t)} M×N α ik (t) represents the task σ at time t. k Unloaded to server E i The proportion of subtasks. For the task σ currently being executed on the network. k α(t) satisfies Arrival probability η k (t) = 1. For tasks that are not computed within the network, σ... k' α(t) satisfies α ik' (t) = 0, arrival probability η k' (t)∈[0,1].

[0047] 4) At time t, perform pre-migration on tasks already executed in the network β(t)={β ijk (t)} M×M×N ,β ijk (t) represents the task σ at time t. k From server E i Transmitted to server E j The proportion of subtasks. For the task σ currently being executed on the network. k The pre-migration action satisfies Let β be the set of all possible pre-transfer actions for the task. With probability ∈ , randomly select action β(t) at time t from action set β, and simultaneously select the optimal action β(t) = arg max with probability (1-∈ ). β∈βQ(α(t),β).

[0048] 5) At time t, perform the task pre-migration action β(t) and update the task unloading state α(t+1) = α(t) + β(t), where α(t) and β(t) represent the task unloading state and task pre-migration action at time t. For tasks newly arriving in the network at time t+1, unload them to the server with the lowest cost based on the server's idle state to perform calculations, and update α(t+1) at the same time.

[0049] 6) Considering the multi-dimensional costs of computation, communication, and storage in task execution within a computing network, the steps are as follows:

[0050] ① Calculate the migration delay generated during task pre-migration at time t. Based on the pre-migration amount of tasks between servers, the migration latency is calculated as follows:

[0051] Migration costs are calculated as follows:

[0052]

[0053] ② The computational delay generated when the server executes tasks Considering the server's computational failure, we introduce an uncertainty following a Poisson distribution:

[0054]

[0055] Where θ is defined as the computational delay under error-free conditions, through θ ik (t)=α ik (t)·l k / f i The calculation yields the following result. Based on the probability density function: Acquire the computational reliability requirements of the task Γ k The computation delay is calculated as follows:

[0056]

[0057] The total cost of executing further task k can be calculated using the following formula:

[0058]

[0059] in For time t, the k-th task is on server E i The computation time required to execute the subtasks of loading and unloading. For time t, the k-th task starts from server E. i Upgrade to server E j The required migration latency, ξ i For server Ei Storage unit cost.

[0060] Optimize server task offloading decisions with the goal of minimizing total execution cost.

[0061] 7) Calculate the cost reduction benefit R:

[0062] ① Due to the pre-migration action β(t), some computing resources in the computing network are reserved in advance. Therefore, tasks newly arriving in the network at time t+1 will be unloaded to the server with the lowest cost based on the server's idle status, and the updated α(t+1) will be obtained. In contrast, without the pre-migration action, the task unloading status is updated to α based on the idle computing resources at time t and the tasks newly arriving at time t+1. 0 (t+1).

[0063] Considering the probability of reaching the task η, the specific formula for calculating the reward is as follows:

[0064]

[0065]

[0066] in:

[0067] The execution of the pre-migration action β(t) results in a change ΔE in the total cost of the original task. cost (t)

[0068] ΔE cost (t)=E cost (α ori (t+1))-E cost (α(t));

[0069] The migration cost E generated by the pre-migration action β(t) mig (t);

[0070] The execution of the pre-migration action β(t) results in a reduction in execution cost E for the new task at arrival probability η. gain (α(t+1))E gain (α(t+1))=E cost (α new (t+1))-E cost (α 0 (t+1));

[0071] ② If the computation delay of the task exceeds the delay limit after performing the pre-transfer action. Or the server's task offload volume exceeds its maximum storage capacity. If this does not meet the network requirements, the return R(t) = 0.

[0072] 8) Update the Q value. The specific calculation formula is as follows:

[0073]

[0074] Where the discount revenue V(α(t)) = argmax β Q(α(t),β(t)); γ is the learning rate, representing how fast learning occurs; γ is the discount factor.

[0075] 9) Repeat steps 4-8 at each time step, adjusting the original task unloading ratio to reserve computing resources for upcoming tasks. The unloading status at the next time step is updated to α(t+1)=α ori (t+1)+α new (t+1).

[0076] exist Figure 1 , Figure 2 In this study, a simple scenario with M=3 and N=4 was selected, where two tasks were already computed in the computing network. The discount factor γ was set to 0.8, the learning rate to 0.7, and the task arrival probability η to 0.8. The total execution cost of the network tasks was compared after the algorithm converged. The following greedy algorithm and random allocation algorithm were selected as benchmark performance comparisons: the greedy algorithm compares server parameters and task latency requirements, prioritizing tasks with higher latency requirements to servers with stronger computing power; the random allocation algorithm employs a random offloading strategy for tasks while meeting latency and storage constraints.

[0077] Figure 1 This paper compares the performance of the proposed invention with two comparative algorithms (random allocation and greedy algorithm) under different task numbers, with fixed server parameters and a relatively idle network (total tasks / maximum network storage < 60%). The decision algorithm based on task pre-migration proposed in this invention can effectively reduce the total execution cost, and when the number of tasks increases, the proposed algorithm can further reduce the total execution cost within the limits of server capacity.

[0078] Figure 2 This paper compares the performance of this invention with two other algorithms (random allocation and greedy algorithm) under relatively congested network conditions (total tasks / maximum network storage > 80%), with fixed server parameters and varying numbers of tasks. The greedy algorithm selects tasks based on available network resources. However, under network congestion, the available resources unloaded by tasks using the greedy algorithm are difficult to utilize, resulting in poor performance. The task pre-migration proposed in this invention can effectively reduce the total cost of task execution.

[0079] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalent technical solutions.

Claims

1. A task pre-migration scheduling method for a computing power oriented network, characterized in that, The steps are as follows: 1) Obtain N tasks σ = [σ1, σ2, ..., σ] in the computing power network. N The parameters include tasks already executed in the network and tasks about to arrive in the network for unloading and execution. Task parameters include: task σ k Data volume l k Total task execution delay constraint d k ; Obtain the M servers E = [E1, E2, ..., E] in the computing power network. M The parameters include those for each server E. i computing power f i Calculate power storage space Server E i and E j Inter-task migration transmission bandwidth Migration power 2) Initialize the Q-value table Q(α(0),β(0)) as a zero matrix; 3) Obtain the original unloading state α(t) of the task already in the network at time t = {α ik (t)} M×N α ik (t) represents the task σ at time t. k Unloaded to server E i The proportion of subtasks; for tasks σ currently being executed on the network. k α(t) satisfies Arrival probability η k (t) = 1; for tasks that are not computed on the network, σ k' α(t) satisfies α ik' (t) = 0, arrival probability η k' (t)∈[0,1]; 4) At time t, perform pre-migration on tasks already executed in the network β(t)={β ijk (t)} M×M×N ,β ijk (t) represents the task σ at time t. k From server E i Transmitted to server E j The proportion of subtasks; for tasks σ currently being executed on the network. k The pre-migration action satisfies The set of all possible pre-transfer actions is β; the action β(t) at time t is randomly selected from the action set β with probability ∈ , and the optimal action β(t) = argmax is selected with probability (1-∈ ). β∈β Q(α(t),β); 5) Update the unloading status of existing tasks already in the network based on the task pre-migration action β(t) at time t: Calculate the cost reduction benefit R(t) after the action is performed, and update the Q value. The calculation formula is as follows: where the discounted reward V(a(t)) = argmax β Q(a(t), b(t)); is the learning rate, indicating how fast the learning is; and g is the discount factor. 6) Repeat steps 3)-5) at each time, adjust the original task offloading ratio, reserve computing resources for the arriving tasks, and update the offloading state at the next time as α(t+1) = α ori (t+1) + α new (t+1).

2. The method of claim 1, wherein, The profit R(t) in step 5): 2.1) Perform pre-migration action β(t), part of the computing resources in the computing network is reserved in advance, and the tasks newly arrived in the network at time t+1 are unloaded to the server with the minimum cost according to the idle state of the server to perform calculation, and updated α new (t+1) is obtained; As a control, define when the pre-migration action is not performed, according to the idle computing resource at time t and the newly arrived task at time t+1, the task offloading state is updated to a 0 (t+1); The benefits of pre-migration are: R(t) = ΔE cost (t) + E mig (t) + E gain (α new (t + 1)) in: The execution of the pre-migration action β(t) results in a change ΔE in the total cost of the original task cost (t) ΔE cost (t) = E cost (α ori (t + 1) - E cost (α(t)); Migration cost E resulting from pre-migration action β(t) mig (t); Execution of the pre-migration action β(t) results in a reduction E in the execution cost obtained by the new task with arrival probability η gain (α(t+1)) E gain (α(t+1)) = E cost (α new (t+1))-E cost (α 0 (t+1)) ; 2.2) The execution cost of all tasks in the computing power network at time t is where ξ i is the server E i storage unit cost, is the computing time required by the kth task at time t to execute the subtasks offloaded to server E i . where θ ik (t) = α ik (t) · l k / f i , Γ k is the computational reliability requirement of task σ k ; The migration cost for all tasks in the computing power network is: wherein, is the migration delay generated at the time of pre-migration of the task at time t, The total benefits of pre-migration include: If the computing time delay of the task is greater than the time delay limit after performing the pre-migration action Or the task offloading amount of the server is greater than the maximum storage amount Then it does not meet the network requirement, at this time the benefit R(t) = 0.

Images