A hybrid backup MEC system and a burst task offloading method and system

Abstract

The application discloses a mixed backup MEC system and a burst task offloading method and system, and relates to the technical field of edge computing and task offloading. The method comprises the following steps: describing the arrival process of various tasks in the burst task flow of the MEC system by using a MAP; based on the arrival process of various tasks, sequentially solving the system average time delay and the edge layer average power; based on the system average time delay and the edge layer average power, constructing a double-layer Stackelberg game model, and solving the double-layer Stackelberg game model based on the iterative algorithm of ADMM and SCA to obtain an optimal task offloading strategy and an optimal frequency pricing strategy. The application can not only meet the differentiated QoS requirement of classified tasks, but also improve the reliability and fault tolerance of the MEC system.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and task offloading technology, and in particular to a hybrid backup MEC system and a method and system for offloading bursty tasks. Background Technology

[0002] MEC, as an emerging computing paradigm, places computing and data processing functions at the network edge, providing faster and more reliable computing resources for various applications. With the rise of smart devices and the Internet of Things (IoT), MEC systems are becoming one of the key technologies for solving the needs of data processing and low-latency applications. As IoT homes become increasingly diverse and intelligent, IoT terminal devices generate a large number of different types of tasks, often with bursty characteristics. Especially under conditions of network congestion or unreliable computing resources, this can lead to degraded task processing performance or excessive system load. Therefore, efficiently handling classification tasks with bursty characteristics has become a challenge. Summary of the Invention

[0003] The purpose of this invention is to provide a hybrid backup MEC system and a method and system for offloading bursty tasks, which can not only meet the requirements of differentiated QoS for classified tasks, but also improve the reliability and fault tolerance of the MEC system.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A hybrid backup MEC system and a method for unloading bursty tasks, comprising:

[0006] The arrival process of various tasks in the burst task flow of the MEC system is described using MAP; the terminal device layer of the MEC system includes multiple terminal devices with the same configuration, and each terminal device is equipped with a scheduler, battery, local CPU, transmission unit and cache component;

[0007] Based on the arrival process of various tasks, the system average latency and edge layer average power are solved sequentially; the system average latency includes the average latency of a task receiving service on the local processor, the average latency of latency-tolerant tasks at the edge layer, and the average latency of latency-tolerant tasks at the edge layer.

[0008] Based on the system's average latency and the edge layer's average power, a two-layer Stackelberg game model is constructed. The two-layer Stackelberg game model is then solved using iterative algorithms based on ADMM and SCA to obtain the optimal task offloading strategy and the optimal frequency pricing strategy.

[0009] This invention also provides a hybrid backup MEC system and a burst task offloading system, comprising:

[0010] The task flow description unit is used to describe the arrival process of various tasks in the burst task flow of the MEC system using MAP; the terminal device layer in the MEC system includes multiple terminal devices with the same configuration, and each terminal device is equipped with a scheduler, battery, local CPU, transmission unit and cache component;

[0011] The average latency and average power calculation unit is used to solve the system average latency and edge layer average power sequentially based on the arrival process of various tasks; the system average latency includes the average latency of a task receiving service on the local processor, the average latency of latency-tolerant tasks at the edge layer, and the average latency of latency-tolerant tasks at the edge layer.

[0012] The optimal task offloading strategy and optimal frequency pricing strategy solution unit is used to construct a two-layer Stackelberg game model based on the average system latency and the average power of the edge layer, and solve the two-layer Stackelberg game model based on the iterative algorithms of ADMM and SCA to obtain the optimal task offloading strategy and optimal frequency pricing strategy.

[0013] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0014] This invention discloses a hybrid backup MEC system and a method and system for burst task offloading. The method includes using a Map to describe the arrival process of various tasks in the burst task flow of the MEC system; based on the arrival process of various tasks, sequentially solving for the system average latency and the edge layer average power; constructing a two-layer Stackelberg game model based on the system average latency and the edge layer average power, and solving the two-layer Stackelberg game model based on the iterative algorithms of ADMM and SCA to obtain the optimal task offloading strategy and the optimal frequency pricing strategy. This invention can meet the requirements of differentiated QoS for classified tasks and improve the reliability and fault tolerance of the MEC system. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the hybrid backup MEC system and the method for unloading sudden tasks according to the present invention.

[0017] Figure 2This is a schematic diagram of the MEC system architecture based on hybrid backup in this embodiment. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a hybrid backup MEC system and a method and system for offloading bursty tasks, which can not only meet the requirements of differentiated QoS for classified tasks, but also improve the reliability and fault tolerance of the MEC system.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figures 1-2 As shown, this invention provides a hybrid backup MEC system and a method for unloading bursty tasks, including:

[0022] Step 100: Use MAP to describe the arrival process of various tasks in the burst task flow of the MEC system; the terminal device layer of the MEC system includes multiple terminal devices with the same configuration, and each terminal device is equipped with a scheduler, battery, local CPU, transmission unit and cache component.

[0023] Step 200: Based on the arrival process of various tasks, solve the system average latency and edge layer average power in sequence; the system average latency includes the average latency of a task receiving services on the local processor, the average latency of latency-tolerant tasks at the edge layer, and the average latency of latency-tolerant tasks at the edge layer.

[0024] Step 300: Based on the average system latency and the average power of the edge layer, construct a two-layer Stackelberg game model, and solve the two-layer Stackelberg game model based on the iterative algorithms of ADMM and SCA to obtain the optimal task offloading strategy and the optimal frequency pricing strategy.

[0025] Based on the above steps, the following embodiments are provided to describe the above steps in detail.

[0026] As a specific implementation method, such as Figure 1As shown, this embodiment will explore the burst task offloading strategy and performance study of MEC system based on hybrid backup. Hybrid backup is a redundancy mechanism that combines hot backup and warm backup, which can not only meet the differentiated QoS requirements of classified tasks, but also improve the reliability and fault tolerance of MEC system. By formulating a reasonable task offloading strategy under the hybrid backup mechanism, this embodiment aims to optimize the performance of MEC system when facing classified burst tasks, including reducing the latency of latency-sensitive tasks, increasing the throughput of latency-tolerant tasks, and enhancing the fault tolerance of the system.

[0027] To address the heterogeneous reliability requirements of various tasks in MEC systems, a classified task offloading strategy based on hybrid backup is proposed for MEC architecture. At the terminal layer, a random service system based on preemption priority is established as the local computing mechanism model. At the edge layer, random service systems based on PM hot standby and PM warm standby are established as edge computing mechanism models for latency-tolerant and latency-sensitive tasks, respectively. A three-dimensional continuous-time Markov chain based on the preemption priority mechanism is used to perform steady-state analysis on the local computing mechanism model. A six-dimensional continuous-time Markov chain based on the retry mechanism is used to perform steady-state analysis on the PM hot standby mechanism model. A seven-dimensional continuous-time Markov chain based on the "loss mechanism" is used to perform steady-state analysis on the PM warm standby mechanism model. Using the results of the above steady-state analysis, system performance indicators, including throughput, average latency, average power, and reliability, are derived. Numerical experiments are designed to reveal the impact of strategy parameters on performance indicators, verifying the effectiveness and performance advantages of the proposed bursty task offloading strategy. Finally, a two-layer Stackelberg game model is constructed from the perspectives of the terminal device group task offloading strategy and the MEC service provider's frequency pricing strategy. Based on the iterative algorithms of ADMM and SCA, we analyze the optimal task offloading strategy and the optimal frequency strategy.

[0028] As mobile users increasingly demand various IoT smart services, the types of tasks generated by terminal devices are becoming more and more complex. Based on the differentiated requirements of task response performance, tasks generated by terminal devices can be categorized into latency-tolerant tasks and latency-sensitive tasks. This embodiment will employ a two-layer logical structure, including a terminal layer and an edge layer, to study offloading strategies for multiple types of tasks.

[0029] In the MEC system architecture described above, the terminal device layer considers multiple terminal devices with identical configurations. Each terminal device is equipped with a scheduler, battery, local CPU, transmission unit, cache, and other components. Sufficiently large dedicated caches are deployed on the local CPU and transmission unit for latency-tolerant and latency-sensitive tasks, respectively. The task scheduler is responsible for making offloading decisions for various task types. Under the scheduling of the task scheduler, newly generated tasks of various types are either assigned to the local CPU for computation or offloaded to the edge layer for computation via the transmission unit.

[0030] To meet the real-time requirements of latency-sensitive tasks, a preemptive priority service mechanism is introduced at the local CPU and transmission unit, and an access threshold is set for the latency-sensitive task cache. On the local CPU or transmission unit, latency-sensitive tasks can preempt the service resources occupied by latency-tolerant tasks. The preempted latency-tolerant tasks will then be returned to their dedicated cache to wait in the queue. When the number of tasks in the latency-sensitive task cache reaches the access threshold, the latency-sensitive tasks are discarded.

[0031] To alleviate service pressure at the edge layer and meet the response requirements of different types of tasks, MEC service providers have begun deploying different types of base stations, most commonly macro base stations (MBS) and small base stations (SBS). Since latency-tolerant tasks are less affected by latency, they also have a certain tolerance for system reliability. However, latency-sensitive tasks are more sensitive to latency and have very high requirements for system reliability. Therefore, to improve edge layer reliability and ensure effective resource allocation, and to avoid resource contention between latency-tolerant and latency-sensitive tasks, MBS and SBS service resources are integrated, and a hybrid backup-based service mechanism is designed at the edge layer. On one hand, multiple PM hot backup redundancy mechanisms are adopted on the MBS. In MBS cluster I, multiple high-performance servers (HPS) are deployed as master servers, and in MBS cluster II, multiple low-performance servers (LPS) are deployed as hot backup servers. On the other hand, a warm backup redundancy mechanism is adopted on SBS, with multiple HPS deployed as master servers in SBS cluster I and multiple LPS deployed as warm backup servers in SBS cluster II.

[0032] In the aforementioned MEC system, the task offloading strategy is as follows:

[0033] (1) After being offloaded, latency-tolerant tasks arrive at the edge layer via the transmission unit and first enter MBS cluster I. If an available HPS exists in MBS cluster I, the task is immediately computed and processed; if no available HPS exists in MBS cluster I and the number of tasks in the queue buffer has not reached the access threshold, the task is queued in the queue buffer according to the first-come, first-served rule; if the number of tasks in the queue buffer reaches the access threshold, the task will be offloaded to MBS cluster II. After being offloaded to MBS cluster II, the task will use a retry mechanism to wait for an opportunity to be computed and processed. If an available LPS exists in MBS cluster II and the number of tasks in the retry buffer has not reached the retry threshold, the task is computed and processed according to the retry rules; if no available LPS exists in MBS cluster II or the number of tasks in the retry buffer reaches the retry threshold, the task will be discarded.

[0034] If a serviced HPS fails, and the number of tasks in the retry cache has not reached the retry threshold, the latency-tolerant tasks currently in service will be migrated to MBS cluster II for computation. If the number of tasks in the retry cache reaches the retry threshold, the latency-tolerant tasks currently in service will be discarded. If a serviced LPS in MBS cluster II fails, the latency-tolerant tasks currently in service will be discarded.

[0035] (2) A warm standby redundancy mechanism is adopted in SBS, and a "loss-based" service mechanism is introduced. This saves edge service costs while ensuring the response performance of latency-sensitive tasks. When HPS in SBS cluster I and LPS in SBS cluster II are active, they can provide services at any time. If they are in a warm standby state, they need to be activated to provide services. When the number of active servers is too low, HPS in standby state is activated first. If the number of available HPS in SBS cluster I is insufficient, then available LPS in SBS cluster II is activated.

[0036] After being offloaded, latency-sensitive tasks arrive at the edge layer via the transmission unit and preferentially occupy active HPS in SBS cluster I. If no active HPS exists in SBS cluster I, active LPS in SBS cluster II will be used. When a server serving an SBS fails, if an idle HPS or LPS exists, the task will preferentially occupy an idle HPS in SBS cluster I. If no idle HPS or LPS exists, or if the number of failed servers in the SBS exceeds the failure threshold, the SBS will stop service, and latency-sensitive tasks will be discarded. Once the failed HPS or LPS is repaired, the system will switch to standby mode or resume normal operation.

[0037] The main steps described in this embodiment are as follows:

[0038] Step 1: In order to describe the correlation of the burst task flow, a MAP is used to characterize the arrival process of various tasks.

[0039] Assume the terminal layer contains D devices, and uses a set Indicates. Terminal device. The process of reaching the upper task depends on having V d The underlying Markov chain {Z} of states d (t), t≥0}. Therefore, the underlying Markov chain {Z} can be defined. d The state space of {(t), t≥0} Represents the underlying Markov chain {Z} at time t (t≥0). d The state z is in the state z(t), t≥0. d At this point, the underlying Markov chain {Z} d The residence time of the group (t), t≥0} follows a parameter σ. _d (σ _d The exponential distribution of >0). Assume the underlying Markov chain {Z d From state z (t), t≥0} d Transferred to At that time, the probability of no task arrival is The arrival probability of a delay-tolerant task is The arrival probability of a time-delay-sensitive task is definition This is the one-step state transition rate matrix when no task arrives. This is the one-step state transition rate matrix when a time-delay-tolerant task arrives. This is the one-step state transition matrix when a time-delay-sensitive task arrives.

[0040] Therefore, the underlying Markov chain {Z d The infinitesimal generator of the group (t), t≥0 can be expressed as And there is a steady-state probability. Satisfy the equilibrium equation With normalization conditions

[0041] Based on steady-state probability The terminal devices can be derived separately. Arrival rate of latency-tolerant tasks Arrival rate of latency-sensitive tasks As shown in the following formula.

[0042]

[0043] terminal equipment Both latency-tolerant and latency-sensitive tasks are probabilistically assigned to the local CPU or offloaded to the edge layer for computation. Let... and They represent terminal devices respectively. The probability that latency-tolerant and latency-sensitive tasks are assigned to the local CPU for computation and processing. Let and They represent terminal devices respectively. The probability that latency-tolerant and latency-sensitive tasks are offloaded to the edge layer, and the following relationship is satisfied. and say and This represents the offloading probability of latency-tolerant tasks and latency-sensitive tasks.

[0044] Step 2, calculate the average latency T of the task receiving services on the local CPU. loc .

[0045] The process of no task arrival on the local CPU can be handled by a matrix. The arrival process of a time-delay-tolerant task can be characterized by a matrix. The arrival process of time-delay-sensitive tasks can be characterized by a matrix. To depict.

[0046]

[0047] Assume the service time of the task on the local CPU follows the parameter. The exponential distribution is such that the arrival process and the service process are independent of each other.

[0048] Latency-sensitive tasks can preempt local CPU resources currently being processed on the local CPU by latency-tolerant tasks. The access threshold for the local cache of latency-sensitive tasks is set to... Therefore, the computational processing of tasks on the local CPU is modeled as a MAP with preemptive priority. tol ,MAP sen / M / 1 / ∞, Random service system.

[0049] Regarding local computation and uplink transmission processes, once the local CPU or transmission resources are occupied, newly arriving latency-tolerant tasks will queue in their dedicated cache. If the local CPU or transmission resources are occupied by latency-tolerant tasks, newly arriving latency-sensitive tasks will preempt the local CPU or transmission resources occupied by the latency-tolerant tasks, forcing the currently processing latency-sensitive tasks to return to their dedicated cache to wait. Similarly, if the local CPU or transmission resources are occupied by latency-sensitive tasks, newly arriving latency-sensitive tasks will also queue in their dedicated cache. Therefore, it can be deduced that for terminal devices... Average wait queue length in local cache for latency-tolerant and time-sensitive tasks and

[0050]

[0051] in, This represents the number of latency-tolerant tasks in the local execution unit at time t (t > 0). This represents the number of latency-sensitive tasks in the local execution unit at time t (t > 0). π dd It consists of steady-state probabilities Steady-state probability vector.

[0052] Using Little's theorem, the local average latency for latency-tolerant and latency-sensitive tasks is derived. and

[0053]

[0054] in This indicates the number of latency-sensitive tasks completed by the local CPU service.

[0055]

[0056] Step 3: Solve for the average latency of latency-tolerant tasks at the edge layer.

[0057] In MBS, the arrival processes of delay-tolerant and delay-tolerant tasks are represented by matrix D respectively. 0-tol With D 1-tol To depict.

[0058]

[0059] Deploying N on MBS cluster I M-HPS One HPS, deployed in MBS cluster II. M-LPS One LPS. Assume that the service time of a latency-tolerant task on the HPS in MBS cluster I follows a parameter μ. M-HPS (μM-HPS The exponential distribution (>0) in MBS cluster II, where the service time on the LPS follows a parameter μ. M-LPS (μ M-LPS It follows an exponential distribution with >0, and there exists μ. M-HPS >μ M-LPS .

[0060] Considering the fault and fault recovery behaviors of HPS in MBS cluster I and LPS in cluster II, the following assumptions are made. The busy-hour fault rate and idle-hour fault rate of HPS in MBS cluster I respectively follow the parameter α. Mb-HPS (α Mb-HPS >0) and α Mi-HPS (α Mi-HPS In a Poisson process where >0), the busy-hour failure rate and idle-hour failure rate of the LPS in MBS cluster II follow parameters α. Mb-LPS (α Mb-LPS >0) and α Mi-LPS (α Mi-LPS The Poisson process (>0). Any of the above-mentioned faults will trigger corresponding repair behaviors. The repair times of HPS in MBS cluster I and LPS in cluster II follow parameters β. HPS (β HPS >0) and β LPS (β LPS The retry time of a latency-tolerant task, when offloaded to MBS cluster II, follows an exponential distribution with parameter θ (θ > 0). Furthermore, the access thresholds for the queue cache and retry cache are set to K. que With K orb .

[0061] Let random variable X orb (t)=x orb (0≤x orb ≤K orb Let Y represent the number of latency-tolerant tasks in the retry cache of MBS cluster II at time t (t≥0). Mf-LPS =y Mf-LPS (0≤y Mf-LPS ≤N Mf-LPS Let Y represent the number of faults in LPS within MBS cluster I at time t (t≥0). Mb-LPS =y Mb-LPS (0≤y Mb-LPS ≤N M-LPS -y Mf-LPS Let Y represent the number of LPS occupied in MBS cluster II at time t (t≥0). Mf-HPS =y Mf-HPS (0≤y Mf-HPS ≤N M-HPSLet X represent the number of HPS failures in MBS cluster I at time t (t≥0). que (t)=x que (0≤x que ≤N M-HPS +K que -y Mf-HPS ) represents the number of latency-tolerant tasks in the queue buffer of MBS cluster I at time t (t≥0).

[0062] At the edge layer, the underlying Markov chains of multiple terminal devices aggregate into a new Markov chain {Z′(t), t≥0}={Z 1 (t),

[0063] Z 2 (t),…,Z D The state space is (t)}, and its state space is Therefore, a continuous-time Markov chain ξ can be constructed. tol (t), whose state space is Ω tol .

[0064] ξ tol (t)={(X orb (t),Y Mf-LPS (t),Y Mb-LPS (t),Y Mf-HPS (t),X que (t),Z′(t)),t≥0}

[0065]

[0066] As t→∞, the MBS calculation process will enter a steady state. Let τ x y y y x z′ ξ represents a continuous-time Markov chain tol The steady-state probability of (t).

[0067]

[0068] (x orb ,y Mf-LPS ,y Mb-LPS ,y Mf-HPS ,x que ,z′)∈Ω tol

[0069] For edge computing, when all available HPS in MBS cluster I are occupied, and the queue cache in MBS cluster I has not reached its access threshold, latency-tolerant tasks will enter the queue cache to wait. Therefore, the average waiting length of latency-tolerant tasks in the MBS queue cache at the edge layer can be derived.

[0070]

[0071] When latency-tolerant tasks are offloaded to the edge layer, they may also enter the MBS cluster II retry cache for LPS computation and processing. Therefore, the average wait length of latency-tolerant tasks in the MBS retry cache can be derived.

[0072]

[0073] Using Little's theorem, the average delay T of a tolerant task on HPS in MBS cluster I can be derived. que-tol The average delay T on LPS in MBS cluster II orb-tol .

[0074]

[0075] Based on the above results, the average latency T of latency-tolerant tasks can be derived. tol .

[0076]

[0077] Where p que p represents the probability of delay-tolerant processing on MBS cluster I. orb This represents the probability of time-delay tolerant processing on MBS cluster II.

[0078]

[0079] Step 4: Solve for the average latency of latency-tolerant tasks at the edge layer.

[0080] In SBS, the arrival processes of non-delay-sensitive tasks and delay-sensitive tasks are represented by matrix D respectively. 0-sen With D 1-sen To depict.

[0081]

[0082] Deploy N on SBS cluster I S-HPS One HPS, deployed on SBS cluster II. S-LPS There are N LPS, and N satisfies S-HPS >N S-LPSBased on the k-out-of-(M+N):G warm standby mechanism, the normal operation of SBS depends on at least k normally functioning servers. Assume that the service time of latency-tolerant tasks processed by HPS in SBS cluster I follows a parameter μ. S-HPS (μ S-HPS The exponential distribution of >0, and the service time of latency-sensitive tasks processed by LPS in SBS cluster II follows a parameter μ. S-LPS (μ S-LPS It follows an exponential distribution with >0, and there exists μ. S-HPS >μ S-LPS Due to the low latency requirements of latency-sensitive tasks, dynamic voltage frequency adjustment technology can be considered to adjust the server's clock frequency, making μ... S-HPS ≥μ M-HPS .

[0083] Compared to hot standby, warm standby allows for three server states: active, warm standby, and faulty. Both active and warm standby HPS / LPS can fail. Therefore, the following assumptions are made: The busy-hour failure rates of HPS and LPS in the active state follow a parameter α. Sb-HPS (α Sb-HPS >0) and α Sb-LPS (α Sb-LPS The exponential distribution of >0) and the failure rate during idle time follow the parameter α. Si-HPS (α Si-HPS >0) and α Si-LPS (α Si-LPS The failure rates of HPS and LPS in backup mode follow an exponential distribution with parameter α > 0. Ss-HPS (α Ss-HPS >0) and α Ss-LPS (α Ss-LPS The exponential distribution of >0).

[0084] In warm standby mode, the failure rate of HPS and LPS is lowest because they operate at low speed and do not process latency-sensitive tasks. In active mode, when HPS and LPS are idle, although they are not currently processing latency-sensitive tasks, they are operating at high speed, resulting in the next lowest failure rate. In active mode, the failure rate of HPS and LPS is highest when they are busy. Furthermore, the burst intensity of tasks also affects the failure rate of HPS and LPS. Therefore, the HPS failure rate in different states satisfies α. Sb-HPS >α Si-HPS >α Ss-HPS The LPS failure rate under different conditions satisfies α Sb-LPS >α Si-LPS >α Ss-LPS .

[0085] Any of the above faults will trigger corresponding repair actions. The repair times for HPS and LPS respectively follow the parameter β. HPS (β HPS >0) and β LPS (β LPS The exponential distribution of >0).

[0086] Let random variable Y Sf-HPS =y Sf-HPS Let Y represent the number of HPS failures in SBS cluster I at time t (t≥0). Sf-LPS =y Sf-LPS Let Y represent the number of faults in the LPS in the edge SBS cluster II at time t (t≥0). Ss-HPS =y Ss-HPS Let Y represent the number of spare HPS in SBS cluster I at time t (t≥0). Ss-LPS =y Ss-LPS Let Y represent the number of spare LPS in SBS cluster II at time t (t≥0). Sb-HPS =y Sb-HPS Let Y represent the number of HPS in SBS cluster I at time t (t≥0). Sb-LPS =y Sb-LPS Let represent the number of LPS occupied in SBS cluster II at time t (t≥0). A seven-dimensional continuous-time Markov chain ξ is constructed by combining the continuous-time Markov chain {Z′(t),t≥0}. sen (t).

[0087] ξ sen (t)={(Y Sf-HPS (t),Y Sf-PS (t),Y Ss-HPS (t),Y Ss-LPS (t),Y Sb-HPS (t),Y Sb-LPS (t),Z′(t)),t≥0}

[0088] Among them, Ω sen ξ represents a continuous-time Markov chain sen The state space of (t) is such that the relations satisfied by each state are as follows:

[0089] Number of HPS failures in SBS cluster I y Sf-HPS Satisfies the condition. 0≤y Sf-HPS ≤N S-HPS The number of LPS faults in SBS cluster II. Sf-LPS Satisfying 0≤y Sf-LPS ≤min{N S-LPS N S-HPS+N S-LPS -ky Sf-HPS +1}; The number of spare HPS in SBS cluster I y Ss-HPS satisfy:

[0090]

[0091] The number of spare LPS in SBS cluster II y Ss-LPS Satisfy y Ss-LPS =max{0,N S-HPS +N S-LPS -ky Sf-HPS -y Sf-LPS -y Ss-HPS}; The number of HPS occupied in SBS cluster I y Sb-HPS satisfy:

[0092]

[0093] The amount of LPS occupied in SBS cluster II y Sb-LPS satisfy:

[0094]

[0095] As t→∞, the SBS calculation process will enter a steady state. Let τ sen-state ξ represents a continuous-time Markov chain sen Steady-state probability of (t):

[0096]

[0097] Y Ss-LPS (t)=y Ss-LPS ,Y Sb-HPS (t)=y Sb-HPS ,Y Sb-LPS (t)=y Sb-LPS ,

[0098] Z′(t)=z′},y Sf-HPS ,y Sf-PS ,y Ss-HPS ,y Ss-LPS ,y Sb-HPS ,y Sb-LPS ,z′∈Ω sen

[0099] Where, sen-state = (y Sf-HPS ,y Sf-PS ,y Ss-HPS ,y Ss-LPS ,y Sb-HPS ,y Sb-LPS ,z′).

[0100] Step 5: Solve for the average power of the edge layer.

[0101] Since MEC service providers use servers of different specifications when providing edge computing services, they can also use DVFS technology to adjust the server frequency.

[0102] make and Let represent the clock frequencies of the HPS in MBS cluster I during the running and idle states, respectively. and Let represent the clock frequencies of the LPS in MBS cluster II during the running and idle states, respectively. Using the same method as for calculating the average power consumption of the terminal device, the execution power of the HPS(LPS) is derived. With idle power

[0103] P MBS Defined as the energy consumption of MBS per unit time. MBS is composed of HPS and LPS. Therefore, P MBS It consists of the energy consumption of HPS per unit time on MBS cluster I and the energy consumption of LPS per unit time on MBS cluster II.

[0104]

[0105] Where ρ M-HPS Indicates HPS utilization rate in MBS cluster I; ρ M-LPS Indicates the LPS utilization rate in MBS cluster II;

[0106]

[0107] Similarly, let and Let these represent the clock frequencies of HPS in SBS cluster I during the running and idle states, respectively. and Let represent the clock frequencies of LPS in SBS cluster II during the running and idle states, respectively; similarly, derive the execution power of HPS(LPS). With idle power

[0108] P SBS Defined as the energy consumption per unit time of SBS. SBS is also composed of HPS and LPS. Therefore, P SBS It consists of the energy consumption of HPS per unit time on SBS cluster I and the energy consumption of LPS per unit time on SBS cluster II.

[0109]

[0110] Where, ρ S-HPS Indicates the utilization rate of HPS in SBS cluster I; ρ S-LPS This indicates the utilization rate of LPS in SBS cluster II;

[0111]

[0112]

[0113] Energy consumption level P of MEC service providers MEC Composed of the energy consumption per unit time of all servers on MBS and SBS: P MEC =P MBS +P SBS .

[0114] Step 6: Construction of the two-layer Stackelberg game problem.

[0115] On the one hand, all terminal devices want to offload their tasks to the edge layer as efficiently as possible, which can improve the QoS of mobile users. On the other hand, MEC service providers need to adjust server frequencies to meet the needs of end users while controlling their own power costs to maximize profits. Furthermore, the task offloading strategy p on terminal device d (d∈D) d The task offloading policy for other devices will be applied. -d The formulation of this system has an impact. Therefore, in this embodiment, the interaction between the terminal device and the MEC service provider is modeled as a Stackelberg game problem with multiple leaders and multiple followers.

[0116] MEC service providers, as leaders, determine their frequency pricing strategies in the first phase. The logarithmic utility function, based on the economic law of diminishing marginal utility, has been widely applied in mobile computing and wireless communications. Here, the logarithmic utility function will be used to construct a profit-maximizing problem for MEC service providers.

[0117] in, f is a vector consisting of clock frequencies.

[0118]

[0119] The constraints are satisfied:

[0120] f min ≤f<f max

[0121] 0.9≤CRA≤1

[0122] 0.9≤p S-run ≤1

[0123] In the second phase, the terminal device, acting as a follower, determines its task offloading strategy, and the follower's objective is represented as the terminal layer average latency F. delay Minimize the problem:

[0124]

[0125] Where p is the task unloading policy vector, p = (p d ,p -d ); Indicates terminal device Task unloading strategy, p -d Indicates the removal of terminal devices Task offloading strategies for other terminal devices.

[0126] The constraints are satisfied:

[0127]

[0128] Given a frequency pricing f from the MEC service provider, there is competition among terminal devices, which is a non-cooperative game. Terminal devices will compete for edge layer computing resources based on their own resource situation and formulate reasonable task offloading strategies. However, the formulated task offloading strategies will indirectly affect the formulation of task offloading strategies by other terminal devices.

[0129] Within the framework of game theory, both MEC service providers and terminal devices are assumed to be rational, aiming to maximize profits. Minimizing the average latency at the terminal layer affects the interests of other terminal devices and MEC service providers. When the number of MEC service providers and terminal devices is large, a centralized optimization scheme that simultaneously maximizes the profits of MEC service providers and minimizes the average latency of terminal devices is practically impossible. Frequency pricing strategies can be viewed as incentives provided by MEC service providers to terminal devices. Since MEC service providers can influence the average latency at the terminal layer by controlling server frequencies, they can further influence the task offloading strategies of terminal devices.

[0130] Therefore, the Stackelberg game described above can be rephrased as an incentive problem. While deriving the average latency at the terminal layer, the optimal frequency pricing strategy for MEC service providers is also presented.

[0131]

[0132] The constraints are satisfied:

[0133] p * =argminF delay (p,*)

[0134]

[0135] Step 7: Solving for the optimal task unloading strategy and the optimal pricing strategy.

[0136] The optimization problem in step 6 involves balancing the upper and lower layers, making it a standard Equilibrium Problem with Equilibrium Constraints (EPEC). In this scenario, it involves coordinating multiple conflicting interests: the average latency of the terminal devices and the interests of the MEC service provider. To maximize the interests of the MEC service provider while minimizing the average latency of the terminal devices, this embodiment will employ an iterative algorithm based on ADMM and SCA.

[0137] As another specific implementation method, such as Figure 2 As shown in the diagram, in the MEC system architecture, the terminal device layer considers multiple terminal devices with identical configurations. Each terminal device is equipped with a scheduler, battery, local CPU, transmission unit, cache, and other components. Sufficiently large dedicated caches are deployed on the local CPU and transmission unit for latency-tolerant and latency-sensitive tasks, respectively. The task scheduler is responsible for making offloading decisions for various task types. Under the scheduling of the task scheduler, newly generated tasks are either assigned to the local CPU for computation or offloaded to the edge layer for computation via the transmission unit.

[0138] To meet the real-time requirements of latency-sensitive tasks, a preemptive priority service mechanism is introduced at the local CPU and transmission unit, and an access threshold is set for the latency-sensitive task cache. On the local CPU or transmission unit, latency-sensitive tasks can preempt the service resources occupied by latency-tolerant tasks. The preempted latency-tolerant tasks will then be returned to their dedicated cache to wait in the queue. When the number of tasks in the latency-sensitive task cache reaches the access threshold, the latency-sensitive tasks are discarded.

[0139] To alleviate service pressure at the edge layer and meet the response requirements of different types of tasks, MEC service providers have begun deploying different types of base stations, most commonly macro base stations (SBS) and micro base stations (MBS). Since latency-tolerant tasks are less affected by latency, they also have a certain tolerance for system reliability. However, latency-sensitive tasks are more sensitive to latency and have very high requirements for system reliability. Therefore, to improve edge layer reliability and ensure effective resource allocation, and to avoid resource contention between latency-tolerant and latency-sensitive tasks, MBS and SBS service resources are integrated, and a hybrid backup-based service mechanism is designed at the edge layer. On one hand, multiple PM hot backup redundancy mechanisms are adopted on the MBS. In MBS cluster I, multiple high-performance servers (HPS) are deployed as master servers, and in MBS cluster II, multiple low-performance servers (LPS) are deployed as hot backup servers. On the other hand, a k-out-of-(M+N):G warm backup redundancy mechanism is adopted on SBS, with multiple HPS deployed as master servers in SBS cluster I and multiple LPS deployed as warm backup servers in SBS cluster II.

[0140] After being offloaded, latency-tolerant tasks arrive at the edge layer via the transmission unit and first enter MBS cluster I. If an available HPS exists in MBS cluster I, the task is immediately computed and processed. If no available HPS exists in MBS cluster I and the number of tasks in the queue buffer has not reached the access threshold, the task is queued in the queue buffer according to the first-come, first-served rule. If the number of tasks in the queue buffer reaches the access threshold, the task will be offloaded to MBS cluster II. After being offloaded to MBS cluster II, a retry mechanism will be used to wait for an opportunity to be computed and processed. If an available LPS exists in MBS cluster II and the number of tasks in the retry buffer has not reached the retry threshold, the task is computed and processed according to the retry rules. If no available LPS exists in MBS cluster II or the number of tasks in the retry buffer reaches the retry threshold, the task will be discarded.

[0141] If a serviced HPS fails, and the number of tasks in the retry cache has not reached the retry threshold, the latency-tolerant tasks currently in service will be migrated to MBS cluster II for computation. If the number of tasks in the retry cache reaches the retry threshold, the latency-tolerant tasks currently in service will be discarded. If a serviced LPS in MBS cluster II fails, the latency-tolerant tasks currently in service will be discarded.

[0142] SBS employs a k-out-of-(M+N):G warm standby redundancy mechanism and introduces a "loss-based" service mechanism. This ensures the response performance of latency-sensitive tasks while saving edge service costs. HPS in SBS cluster I and LPS in SBS cluster II can provide service at any time when they are active; if they are in warm standby, they need to be activated to provide service. When the number of active servers is too low, HPS in standby mode is activated first. If the number of available HPS in SBS cluster I is insufficient, then available LPS in SBS cluster II is activated.

[0143] After being offloaded, latency-sensitive tasks arrive at the edge layer via the transmission unit and preferentially occupy active HPS in SBS cluster I. If no active HPS exists in SBS cluster I, active LPS in SBS cluster II will be used. When a server serving an SBS fails, if an idle HPS or LPS exists, the task will preferentially occupy an idle HPS in SBS cluster I. If no idle HPS or LPS exists, or if the number of failed servers in the SBS exceeds the failure threshold, the SBS will stop service, and latency-sensitive tasks will be discarded. Once the failed HPS or LPS is repaired, the system will switch to standby mode or resume normal operation.

[0144] Compared with the prior art, the technical solution of the present invention has obvious beneficial effects.

[0145] Addressing the heterogeneous reliability requirements of various tasks in MEC systems, this invention presents an offloading strategy and performance analysis for bursty tasks in MEC systems based on hybrid backup. At the terminal layer, a random service model with preemption priority is established based on the bursty characteristics of tasks. At the edge layer, hot standby and warm backup mechanisms are proposed for latency-tolerant and latency-sensitive tasks, respectively. Performance indicators such as throughput, average latency, average power, and reliability are derived within a matrix geometric solution framework. Numerical experiments analyze the impact of task offloading probability and correlation coefficient on system performance and verify the effectiveness of the multi-type bursty task offloading strategy under the hybrid backup mechanism. A two-layer Stackelberg game problem is established, treating the terminal layer device group as followers and the MEC service provider as the leader. Finally, an iterative algorithm based on ADMM and SCA is used to solve for the optimal task offloading strategy and optimal pricing strategy, and the effectiveness of the proposed algorithm is verified.

[0146] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0147] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A hybrid backup MEC system and a method for unloading bursty tasks, characterized in that, include: The arrival process of various tasks in the burst task flow of the MEC system is described using MAP; the terminal device layer of the MEC system includes multiple terminal devices with the same configuration, and each terminal device is equipped with a scheduler, battery, local CPU, transmission unit and cache component; Based on the arrival process of various tasks, the system average latency and edge layer average power are solved sequentially; the system average latency includes the average latency of a task receiving service on the local processor, the average latency of latency-tolerant tasks at the edge layer, and the average latency of latency-tolerant tasks at the edge layer. Based on the system's average latency and the edge layer's average power, a two-layer Stackelberg game model is constructed. The two-layer Stackelberg game model is then solved using iterative algorithms based on ADMM and SCA to obtain the optimal task offloading strategy and the optimal frequency pricing strategy.

2. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The description of the arrival process of various tasks in the burst task flow of the MEC system using MAP specifically includes: Assume the terminal device layer contains D terminal devices, and uses a set Indicates; the terminal device The process of reaching the upper task depends on having V d The underlying Markov chain {Z} of states d (t), t≥0}; Therefore, the underlying Markov chain {Z} is defined. d The state space of {(t), t≥0} And by Represents the underlying Markov chain {Z} at time t (t≥0). d The state z is in the state z(t), t≥0. d At this point, the underlying Markov chain {Z} d The residence time of the group (t), t≥0} follows the parameter. The exponential distribution; assuming the underlying Markov chain {Z d From state z (t), t≥0} d Transferred to At that time, the probability of no task arrival is The arrival probability of a delay-tolerant task is The arrival probability of a time-delay-sensitive task is definition This is the one-step state transition rate matrix when no task arrives. This is the one-step state transition rate matrix when a time-delay-tolerant task arrives. This is the one-step state transition matrix when a time-delay-sensitive task arrives; Therefore, the underlying Markov chain {Z d The infinitesimal generator of the group (t), t≥0 is represented as And there is a steady-state probability. Satisfy the equilibrium equation With normalization conditions Based on steady-state probability Derive the terminal devices respectively Arrival rate of latency-tolerant tasks Arrival rate of latency-sensitive tasks As shown below: The terminal device Both latency-tolerant and latency-sensitive tasks are probabilistically assigned to the local CPU or offloaded to the edge layer for computation; let and They represent terminal devices respectively. The probability that latency-tolerant and latency-sensitive tasks are assigned to the local CPU for computation and processing; let and They represent terminal devices respectively. The probability that latency-tolerant and latency-sensitive tasks are offloaded to the edge layer, and the following relationship is satisfied. and This yields the offloading probability of latency-tolerant tasks. Probability of offloading latency-sensitive tasks 3. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The process for calculating the average latency of the task receiving services on the local processor is as follows: No task arrival process on the local CPU is handled by a matrix. The arrival process of a time-delay-tolerant task is characterized by a matrix. The arrival process of time-delay-sensitive tasks is characterized by a matrix. Depiction; Assume the service time of the task on the local CPU follows the parameter. The exponential distribution, and the arrival process and the service process are independent of each other; Latency-sensitive tasks preempt local CPU resources currently being processed on the local CPU from latency-tolerant tasks. The access threshold for the local cache of latency-sensitive tasks is set to [value missing]. Therefore, the computational processing of tasks on the local CPU is modeled as a MAP with preemptive priority. tol ,MAP sen / M / 1 / ∞, Random service system; For local computation and uplink transmission processes, once the local CPU or transmission resources are occupied, newly arriving latency-tolerant tasks will queue in the corresponding dedicated cache. If the local CPU or transmission resources are occupied by latency-tolerant tasks, newly arriving latency-sensitive tasks will preempt the local CPU or transmission resources occupied by the latency-tolerant tasks, causing the currently processed latency-sensitive tasks to return to the corresponding dedicated cache to queue. If the local CPU or transmission resources are occupied by latency-sensitive tasks, newly arriving latency-sensitive tasks will also queue in the corresponding dedicated cache. Therefore, the terminal device is derived. Average wait queue length in local cache for latency-tolerant and time-sensitive tasks and in, This represents the number of latency-tolerant tasks in the local execution unit at time t (t > 0). This represents the number of latency-sensitive tasks in the local execution unit at time t (t > 0). It consists of steady-state probabilities Steady-state probability vector; Using Little's theorem, the local average latency for latency-tolerant and latency-sensitive tasks is derived. and in, This indicates the number of latency-sensitive tasks completed by the local CPU service.

4. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The process for calculating the average latency of the latency-tolerant task at the edge layer is as follows: In MBS, the arrival processes of delay-free and delay-tolerant tasks are represented by matrix D, respectively. 0-tol With D 1-tol Characterization: Deploying N on MBS cluster I M-HPS One HPS, deployed in MBS cluster II. M-LPS One LPS; assuming that the service time of a latency-tolerant task on the HPS in MBS cluster I follows a parameter μ. M-HPS (μ M-HPS The exponential distribution (>0) in MBS cluster II, where the service time on the LPS follows a parameter μ. M-LPS (μ M-LPS It follows an exponential distribution with >0, and there exists μ. M-HPS >μ M-LPS ; Considering the fault and fault recovery behaviors of HPS in MBS cluster I and LPS in cluster II, the following assumptions are made: the busy-hour fault rate and idle-hour fault rate of HPS in MBS cluster I respectively follow the parameter α. Mb-HPS (α Mb-HPS >0) and α Mi-HPS (α Mi-HPS In a Poisson process where >0), the busy-hour failure rate and idle-hour failure rate of the LPS in MBS cluster II follow parameters α. Mb-LPS (α Mb-LPS >0) and α Mi-LPS (α Mi-LPS The Poisson process (>0) applies; any of the above-mentioned faults will trigger corresponding repair actions. The repair times of HPS in MBS cluster I and LPS in cluster II respectively follow the parameter β. HPS (β HPS >0) and β LPS (β LPS The retry time of a latency-tolerant task follows an exponential distribution with parameter θ (θ > 0); when the task is offloaded to MBS cluster II, its retry time follows an exponential distribution with parameter θ (θ > 0); in addition, the access thresholds for the queue cache and the retry cache are set to K respectively. que With K orb ; Let random variable X orb (t)=x orb (0≤x orb ≤K orb Let Y represent the number of latency-tolerant tasks in the retry cache of MBS cluster II at time t (t≥0). Mf-LPS =y Mf-LPS (0≤y Mf-LPS ≤N Mf-LPS Let Y represent the number of faults in LPS within MBS cluster I at time t (t≥0), and let the random variable Y be... Mb-LPS =y Mb-LPS (0≤y Mb-LPS ≤N M-LPS -y Mf-LPS Let Y represent the number of LPS occupied in MBS cluster II at time t (t≥0). Mf-HPS =y Mf-HPS (0≤y Mf-HPS ≤N M-HPS Let X represent the number of HPS failures in MBS cluster I at time t (t≥0), and let the random variable X be... que (t)=x que (0≤x que ≤N M-HPS +K que -y Mf-HPS ) represents the number of latency-tolerant tasks in the queue buffer of MBS cluster I at time t (t≥0); At the edge layer, the underlying Markov chains of multiple terminal devices aggregate into a new Markov chain {Z′(t), t≥0}={Z 1 (t),Z 2 (t),…,Z D The state space is (t)}, where is Therefore, construct a continuous-time Markov chain ξ tol (t), with state space Ω tol ; ξ tol (t)={(X orb (t),Y Mf-LPS (t),Y Mb-LPS (t),Y Mf-HPS (t),X que (t),Z′(t)),t≥0} As t→∞, the MBS calculation process will enter a steady state, let ξ represents a continuous-time Markov chain tol The steady-state probability of (t); (x orb ,and Mf-LPS ,and Mb-LPS ,and Mf-HPS ,x que ,z′)∈Ω tol For edge computing, when all HPS in MBS cluster I are occupied and the queue cache in MBS cluster I has not reached the access threshold, latency-tolerant tasks will enter the queue cache to wait in line. Therefore, the average waiting length of edge-layer latency-tolerant tasks in the MBS queue buffer is derived. When latency-tolerant tasks are offloaded to the edge layer, they may also enter the MBS cluster II retry cache for LPS computation and processing; therefore, the average wait queue length of latency-tolerant tasks in the MBS retry cache can be derived. Using Little's theorem, the average latency T of a tolerant task on HPS in MBS cluster I is derived. que-tol The average delay T on LPS in MBS cluster II orb-tol : Based on the above results, the average latency T of latency-tolerant tasks is derived. tol : Where, p que p represents the probability of delay-tolerant processing on MBS cluster I. orb This represents the probability of time-delay tolerant processing on MBS cluster II.

5. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The process for calculating the average latency of the latency-tolerant task at the edge layer is as follows: In SBS, the arrival processes of latency-free and latency-sensitive tasks are represented by matrix D respectively. 0-sen With D 1-sen Characterization: Deploy N on SBS cluster I S-HPS One HPS, deployed on SBS cluster II. S-LPS There are N LPS, and N satisfies S-HPS >N S-LPS Based on the k-out-of-(M+N):G warm standby mechanism, the normal operation of SBS depends on at least k normally functioning servers; assuming that the service time of latency-tolerant tasks processed by HPS in SBS cluster I follows a parameter μ. S-HPS (μ S-HPS The exponential distribution of >0, and the service time of latency-sensitive tasks processed by LPS in SBS cluster II follows a parameter μ. S-LPS (μ S-LPS It follows an exponential distribution with >0, and there exists μ. S-HPS >μ S-LPS Due to the low latency requirements of latency-sensitive tasks, dynamic voltage frequency adjustment technology is considered to adjust the server's clock frequency, so that μ S-HPS ≥μ M-HPS ; Compared to the hot standby mechanism, the warm standby mechanism divides server states into three categories: active, warm standby, and faulty. Failures can occur in both the active and warm standby states of HPS / LPS. Therefore, we assume that the busy-hour failure rates of HPS and LPS in the active state follow a parameter α. Sb-HPS (α Sb-HPS >0) and α Sb-LPS (α Sb-LPS The failure rate in the idle state follows an exponential distribution with parameter α > 0. Si-HPS (α Si-HPS >0) and α Si-LPS (α Si-LPS The failure rates of HPS and LPS in backup mode follow an exponential distribution with parameters α > 0. Ss-HPS (α Ss-HPS >0) and α Ss-LPS (α Ss-LPS The exponential distribution of >0); In warm standby mode, the failure rate of HPS and LPS is lowest because they operate at low speed and do not process latency-sensitive tasks. In active mode, when HPS and LPS are idle, although they are not currently processing latency-sensitive tasks, they are operating at high speed, resulting in the next lowest failure rate. In active mode, the failure rate of HPS and LPS is highest when they are busy. Furthermore, the burst intensity of tasks also affects the failure rate of HPS and LPS. Therefore, the HPS failure rate in different states satisfies α. Sb-HPS >α Si-HPS >α Ss-HPS The LPS failure rate under different conditions satisfies α Sb-LPS >α Si-LPS >α Ss-LPS ; Any of the above faults will trigger corresponding repair actions. The repair times for HPS and LPS respectively follow the parameter β. HPS (β HPS >0) and β LPS (β LPS The exponential distribution of >0); Let random variable Y Sf-HPS =y Sf-HPS Let Y represent the number of HPS failures in SBS cluster I at time t (t≥0). Sf-LPS =y Sf-LPS Let Y represent the number of faults in the LPS in the edge SBS cluster II at time t (t≥0). Ss-HPS =y Ss-HPS Let Y represent the number of spare HPS in SBS cluster I at time t (t≥0). Ss-LPS =y Ss-LPS Let Y represent the number of spare LPS in SBS cluster II at time t (t≥0). Sb-HPS =y Sb-HPS Let Y represent the number of HPS occupied in SBS cluster I at time t (t≥0). Sb-LPS =y Sb-LPS The number of LPSs occupied in SBS cluster II at time t (t≥0) is represented by the number of LPSs occupied. A seven-dimensional continuous-time Markov chain ξ is constructed by combining the continuous-time Markov chain {Z′(t),t≥0}. sen (t): ξ sen (t)={(Y Sf-HPS (t),Y Sf-PS (t),Y Ss-HPS (t),Y Ss-LPS (t),Y Sb-HPS (t),Y Sb-LPS (t),Z′(t)),t≥0} Among them, Ω sen ξ represents a continuous-time Markov chain sen The state space of (t) is such that the relations satisfied by each state are as follows: Number of HPS failures in SBS cluster I y Sf-HPS Satisfying 0≤y Sf-HPS ≤N S-HPS The number of LPS faults in SBS cluster II. Sf-LPS Satisfying 0≤y Sf-LPS ≤min{N S-LPS N S-HPS +N S-LPS -ky Sf-HPS +1}; The number of spare HPS in SBS cluster I y Ss-HPS satisfy: The number of spare LPS in SBS cluster II y Ss-LPS Satisfy y Ss-LPS =max{0,N S-HPS +N S-LPS -ky Sf-HPS -y Sf-LPS -y Ss-HPS }; The number of HPS occupied in SBS cluster I y Sb-HPS satisfy: The amount of LPS occupied in SBS cluster II y Sb-LPS satisfy: As t→∞, the SBS calculation process will enter a steady state. Let τ sen-state ξ represents a continuous-time Markov chain sen Steady-state probability of (t): wherein, sen-state=(y Sf-HPS ,y Sf-PS ,y Ss-HPS ,y Ss-LPS ,y Sb-HPS ,y Sb-LPS ,z').

6. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The process for calculating the average power of the edge layer is as follows: Since MEC service providers use servers of different specifications when providing edge computing services, DVFS technology is used to adjust the server frequency. make and Let these represent the clock frequencies of HPS in MBS cluster I during the running and idle states, respectively. and Let represent the clock frequencies of the LPS in MBS cluster II during the running and idle states, respectively; using the same method as for calculating the average power consumption of the terminal device, the execution power of HPS(LPS) is derived. With idle power Among them, P MBS This represents the energy consumption of MBS per unit time, and MBS is composed of HPS and LPS; therefore, P MBS It consists of the energy consumption of HPS per unit time on MBS cluster I and the energy consumption of LPS per unit time on MBS cluster II: Where ρ M-HPS Indicates HPS utilization rate in MBS cluster I; ρ M-LPS Indicates the LPS utilization rate in MBS cluster II; Similarly, let and Let these represent the clock frequencies of HPS in SBS cluster I during the running and idle states, respectively. and Let represent the clock frequencies of LPS in SBS cluster II during the running and idle states, respectively; similarly, derive the execution power of HPS(LPS). With idle power Among them, P SBS This represents the energy consumption of the SBS per unit time. Since the SBS is also composed of HPS and LPS, P... SBS It consists of the energy consumption of HPS per unit time on SBS cluster I and the energy consumption of LPS per unit time on SBS cluster II: Where, ρ S-HPS Indicates the utilization rate of HPS in SBS cluster I; ρ S-LPS This indicates the utilization rate of LPS in SBS cluster II; Energy consumption level P of MEC service providers MEC Composed of the energy consumption per unit time of all servers on MBS and SBS: P MEC =P MBS +P SBS .

7. The hybrid backup MEC system and the method for unloading bursty tasks according to claim 1, characterized in that, The construction process of the two-layer Stackelberg game model includes: terminal equipment Task unloading strategy p d The task offloading policy for other devices will be applied. -d The formulation of the rules has an impact, therefore, the interaction between terminal devices and MEC service providers is modeled as a multi-master, multi-slave Stackelberg game problem; MEC service providers, as leaders, determine their frequency pricing strategies in the first phase. The logarithmic utility function, based on the economic law of diminishing marginal utility and widely applied in mobile computing and wireless communications, is used to construct a profit-maximizing problem for MEC service providers. in, f is a vector consisting of clock frequencies. The constraints are satisfied: f min ≤f<f max 0.9≤CRA≤1 0.9≤p S-run ≤1 In the second phase, the terminal device, acting as a follower, determines its task offloading strategy, and the follower's objective is represented as the terminal layer average latency F. delay Minimize the problem: Where p is the task unloading policy vector, p = (p d ,p -d ); Indicates terminal device Task unloading strategy, p -d Indicates the removal of terminal devices Task offloading strategies for other terminal devices; The constraints are satisfied: Given a frequency pricing f from the MEC service provider, there is competition among terminal devices, which is a non-cooperative game. Terminal devices will compete for edge layer computing resources based on their own resource situation and formulate reasonable task offloading strategies. However, the formulated task offloading strategies will indirectly affect the formulation of task offloading strategies by other terminal devices. Therefore, the Stackelberg game described above can be rephrased as an incentive problem, yielding the optimal frequency pricing strategy for MEC service providers while deriving the average latency at the terminal layer: The constraints are satisfied: p * =argminF delay (p,*) 8. A hybrid backup MEC system and a burst task offloading system, characterized in that, include: The task flow description unit is used to describe the arrival process of various tasks in the burst task flow of the MEC system using MAP; the terminal device layer in the MEC system includes multiple terminal devices with the same configuration, and each terminal device is equipped with a scheduler, battery, local CPU, transmission unit and cache component; The average latency and average power calculation unit is used to solve the system average latency and edge layer average power sequentially based on the arrival process of various tasks; the system average latency includes the average latency of a task receiving service on the local processor, the average latency of latency-tolerant tasks at the edge layer, and the average latency of latency-tolerant tasks at the edge layer. The optimal task offloading strategy and optimal frequency pricing strategy solution unit is used to construct a two-layer Stackelberg game model based on the average system latency and the average power of the edge layer, and solve the two-layer Stackelberg game model based on the iterative algorithms of ADMM and SCA to obtain the optimal task offloading strategy and optimal frequency pricing strategy.

Images