A large model computing task offloading and semantic coding optimization method and system

By optimizing the unloading and semantic coding strategies of large-model computing tasks and selecting appropriate computing modes and coding strategies, the high latency and high energy consumption problems of large-model computing tasks in IoT scenarios are solved, communication performance is improved, and application of large-models is supported. .

CN117472466BActive Publication Date: 2024-11-05SHAOGUAN POWER SUPPLY BUREAU OF GUANGDONG POWER GRID CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311421470.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-10-30
Publication Date
2024-11-05
Estimated Expiration
2043-10-30

AI Technical Summary

Technical Problem

In the Internet of Things and intelligent application scenarios, the high latency and high energy consumption problems of large-model computing tasks lead to limited communication performance and the increased burden on the core network, making it difficult to effectively support the development and application of large-models.

Method used

By optimizing the unloading strategy and semantic coding strategy of large-scale computing tasks, selecting local computing mode or edge computing mode, uninstalling large-scale computing tasks, extracting low-dimensional semantic information for transmission, reducing communication delay and energy consumption overhead, and improving communication performance .

Benefits of technology

It realizes the extraction of low-dimensional semantic information from high-dimensional tasks of large models, reduces communication delay and energy consumption overhead, improves communication performance, and supports the development and application of large models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117472466B_ABST
    Figure CN117472466B_ABST
Patent Text Reader

Abstract

The present invention discloses a large model computing task unloading and semantic coding optimization method and system, which receive the large model computing task of the user end, obtain the system overall delay according to the local computing delay, edge computing delay and user transmission delay, obtain the system overall energy consumption according to the local computing energy consumption, edge computing energy consumption and user transmission energy consumption, calculate the system comprehensive cost according to the system overall delay and the system overall energy consumption, build a computing task unloading model, solve the computing task unloading model to obtain the optimal unloading strategy, obtain the final bandwidth and example resource amount according to the optimal unloading strategy, and optimize the semantic coding strategy with the optimal unloading strategy to obtain the optimal semantic rate. The method optimizes the large model computing task unloading strategy and the semantic coding strategy, so that the system can extract low-dimensional semantic information from the high-dimensional task of the large model for transmission, thereby reducing the communication delay and energy consumption overhead.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of machine learning and wireless communication technology, and in particular to a large model computing task offloading and semantic coding optimization method and system. Background Art

[0002] With the rapid development of artificial intelligence and machine learning, large models have attracted widespread attention from researchers. Large models refer to machine learning models with a large number of parameters and complex structures. These models can be applied to process large-scale data and complex problems. Traditional machine learning models, such as logistic regression, decision trees, and naive Bayes, are small in scale and can only process small amounts of data. Deep learning models, on the other hand, can contain millions of parameters and process massive amounts of data. Ultra-large-scale deep learning models can even reach tens of billions of parameters and require supercomputers for training. Compared with traditional models, large models have strong ability to process large-scale data. Large models can process massive amounts of data, thereby improving the accuracy and generalization of machine learning models. In addition, large models have strong ability to handle complex problems. Large models have higher complexity and greater flexibility, and can handle more complex problems. Furthermore, large models have higher accuracy and performance. Large models have more parameters and more complex structures, and can more accurately express data distribution and learn more complex features, thereby improving the accuracy and performance of the model. With the rapid evolution of 5G technology, more and more devices can connect to the Internet through wireless channels, further promoting the emergence and expansion of the Internet of Things (IoT). In the IoT network, each device communicates with other devices through interconnection technology, which provides beneficial support for the development of smart device applications (such as virtual reality, connected cars, and real-time live broadcasting). However, due to the large amount of data generated by these smart applications, the transmission of the core network has placed a heavy load. In addition, most smart applications also involve computationally intensive tasks, which further increases the burden of energy consumption.

[0003] In order to promote the development of smart device applications, mobile edge computing (MEC) has been identified as a revolutionary computing model. It allows smart devices to offload latency-sensitive and energy-intensive large-model computing tasks to nearby edge servers for accelerated computing to support their performance. This model not only helps to reduce local latency and energy consumption, but also helps to reduce the burden on the core network. Specifically, the network has the following significant features: local computing enables data to be processed instantly at the location where it is generated, and data analysis and processing are implemented on edge devices and servers, thereby quickly producing meaningful results. In addition, the MEC network helps to reduce the pressure on central cloud computing and delegate some computing tasks to the edge layer. At the same time, it also emphasizes data security, especially for applications involving privacy information, data does not need to be transmitted to remote cloud servers, which helps to improve overall security. MEC is suitable for a wide range of application scenarios, such as the Internet of Things, smart cities, industrial automation, self-driving cars, etc., providing efficient and low-latency computing solutions for these scenarios. Finally, the MEC network, with the advantages of distributed computing, brings computing resources close to the data source, providing efficient and fast solutions for various real-time applications and services. However, due to limited bandwidth resources and large-scale data transmission, the performance of MEC networks is limited. Summary of the invention

[0004] In order to solve the above technical problems, an embodiment of the present invention provides a large model computing task offloading and semantic coding optimization method. By optimizing the large model computing task offloading strategy and semantic coding strategy, the system can extract low-dimensional semantic information from the high-dimensional tasks of the large model for transmission, reducing communication latency and energy consumption overhead, thereby improving communication performance and supporting the development and application of large models.

[0005] A first aspect of an embodiment of the present invention provides a method for offloading large model computing tasks and optimizing semantic encoding, the method comprising:

[0006] Receive large model computing tasks from the user;

[0007] Select local computing mode or edge computing mode to offload large model computing tasks. If local computing mode is selected to offload large model computing tasks, local computing latency and local computing energy consumption are obtained. If edge computing mode is selected to offload large model computing tasks, edge computing latency and edge computing energy consumption are obtained.

[0008] The overall system delay is obtained based on the local computing delay, edge computing delay and user transmission delay, and the overall system energy consumption is obtained based on the local computing energy consumption, edge computing energy consumption and user transmission energy consumption;

[0009] The system comprehensive cost is calculated based on the system overall delay and system overall energy consumption. After building a computing task offloading model based on the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy.

[0010] The final bandwidth and instance resource amount are obtained according to the optimal offloading strategy, and the semantic encoding strategy is optimized using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode large model computing tasks according to the final bandwidth, instance resource amount and optimal semantic rate.

[0011] In this embodiment, a large model computing task is received from the user end, and a local computing mode or an edge computing mode is selected to unload the large model computing task. If the local computing mode is selected to unload the large model computing task, the local computing delay and the local computing energy consumption are obtained. If the edge computing mode is selected to unload the large model computing task, the edge computing delay and the edge computing energy consumption are obtained. The overall system delay is obtained according to the local computing delay, the edge computing delay and the user transmission delay. The overall system energy consumption is obtained according to the local computing energy consumption, the edge computing energy consumption and the user transmission energy consumption. The overall system delay and the overall system energy consumption are calculated to obtain the system comprehensive cost. The overall system cost is calculated according to the overall system delay and the overall system energy consumption. After the computing task offloading model is constructed, the computing task offloading model is solved to obtain the optimal offloading strategy. The final bandwidth and instance resource amount are obtained according to the optimal offloading strategy, and the semantic encoding strategy is optimized using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode the large model computing tasks according to the final bandwidth, instance resource amount and optimal semantic rate. This method optimizes the large model computing task offloading strategy and the semantic encoding strategy, so that the system can extract low-dimensional semantic information from the high-dimensional tasks of the large model for transmission, reducing the communication delay and energy consumption overhead, thereby improving the communication performance and supporting the development and application of large models.

[0012] In a possible implementation manner of the first aspect, if the local computing mode is selected to offload the large model computing task, the local computing delay and the local computing energy consumption are obtained, specifically:

[0013] If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are determined. The calculation formula for the local computing delay is:

[0014]

[0015] In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing latency;

[0016] The local computing energy consumption is:

[0017]

[0018] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency;

[0019] The local computing energy consumption is:

[0020]

[0021] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.

[0022] In a possible implementation manner of the first aspect, if the edge computing mode is selected to offload the large model computing task, the edge computing latency and edge computing energy consumption are obtained, specifically:

[0023] If the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are determined. The calculation formula for the edge computing delay is:

[0024]

[0025] In the formula, is the computing power allocated to user m by the CAP side, Indicates edge computing latency;

[0026] Edge computing consumes zero energy.

[0027] In a possible implementation manner of the first aspect, the overall system delay is obtained according to the local computing delay, the edge computing delay, and the user transmission delay, specifically:

[0028] Obtain the user transmission delay, where the calculation formula for the user transmission delay is:

[0029]

[0030] In the formula, represents the user transmission delay, rm represents the semantic rate, nm represents the number of neural networks, and I represents the average amount of semantic information contained in each neural network sent to the CAP end;

[0031] According to the local computing delay, edge computing delay and user transmission delay, the overall delay of each network is obtained. The calculation formula of the overall delay of each network is:

[0032]

[0033] In the formula, represents the user transmission delay, Tm represents the overall delay of each network, Indicates edge computing latency, represents the local computing delay, αm represents the binary offloading factor;

[0034] The overall system delay is obtained based on the overall delay of each network, where the overall system delay is:

[0035]

[0036] In the formula, T represents the overall system delay, m represents the Mth user, Represents a collection of users.

[0037] In a possible implementation manner of the first aspect, the overall energy consumption of the system is obtained according to the local computing energy consumption, the edge computing energy consumption, and the user transmission energy consumption, specifically:

[0038] Obtain the user transmission energy consumption, where the calculation formula of the user transmission energy consumption is:

[0039]

[0040] In the formula, Indicates the user transmission delay, represents the transmission power of user m, Indicates the user transmission energy consumption;

[0041] The overall delay of each network is obtained based on the local computing energy consumption, edge computing energy consumption and user transmission energy consumption. The calculation formula for the overall delay of each network is:

[0042]

[0043] In the formula, represents the user transmission energy consumption, E m represents the overall energy consumption of each network, represents the local computing energy consumption, α m represents the binary unloading factor;

[0044] The overall energy consumption of the system is obtained according to the overall energy consumption of each network, where the overall energy consumption of the system is:

[0045]

[0046] In the formula, E represents the overall energy consumption of the system, m represents the Mth user, Represents a collection of users.

[0047] In a possible implementation of the first aspect, the system comprehensive cost is calculated according to the system overall delay and the system overall energy consumption. After the computing task offloading model is constructed according to the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy, which is specifically:

[0048] The system comprehensive cost is calculated based on the system overall delay and system overall energy consumption, where the calculation formula for the system comprehensive cost is:

[0049] Φ=ηT+(1-η)E

[0050] In the formula, Φ represents the comprehensive cost of the system, T represents the overall system delay, E represents the overall system energy consumption, and η∈[0,1] is the weighting factor for evaluating system delay and energy consumption;

[0051] The computing task offloading model is constructed according to the comprehensive cost of the system, where the objective function and constraints of the computing task offloading model are:

[0052]

[0053] In the formula, Φ represents the comprehensive cost of the system, w CAP is the total system bandwidth, F CAP is the total computing power of the CAP side, is the minimum local computing power, is the maximum local computing power, m represents the Mth user, and M represents the user set;

[0054] The computing task offloading model is solved through the Markov decision process to obtain the optimal offloading strategy.

[0055] In a possible implementation of the first aspect, the final bandwidth and the amount of computing resources are obtained according to the optimal offloading strategy, specifically:

[0056] The optimal offloading strategy is calculated using the preset formula to obtain the final bandwidth and instance resource quantity. The preset formula is:

[0057]

[0058]

[0059]

[0060] In the formula, S o Indicates that it has been unloaded to the CAP end user set, S l represents the user set of local calculation, |·| is the set trace operation, Indicates local computing power, is the computing power allocated to user m by the CAP side, is the minimum local computing power, The maximum value of local computing power.

[0061] In a possible implementation manner of the first aspect, the semantic coding strategy is optimized by using the optimal unloading strategy to obtain the optimal semantic rate, specifically:

[0062] According to the optimal offloading strategy and semantic similarity, the semantic rate is calculated using a binary algorithm to obtain the optimal semantic rate.

[0063] A second aspect of an embodiment of the present invention provides a large model computing task offloading and semantic coding optimization system, the system comprising:

[0064] A receiving module, used for receiving large model computing tasks from the user end;

[0065] An unloading module is used to select a local computing mode or an edge computing mode to unload large model computing tasks. If the local computing mode is selected to unload large model computing tasks, the local computing delay and local computing energy consumption are obtained. If the edge computing mode is selected to unload large model computing tasks, the edge computing delay and edge computing energy consumption are obtained.

[0066] The first calculation module is used to obtain the overall system delay according to the local calculation delay, the edge calculation delay and the user transmission delay, and obtain the overall system energy consumption according to the local calculation energy consumption, the edge calculation energy consumption and the user transmission energy consumption;

[0067] The second calculation module is used to calculate the system comprehensive cost according to the system overall delay and the system overall energy consumption, and after building a computing task offloading model according to the system comprehensive cost, solve the computing task offloading model to obtain the optimal offloading strategy;

[0068] The third computing module is used to obtain the final bandwidth and example resource amount according to the optimal offloading strategy, and optimize the semantic coding strategy using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode the large model computing tasks according to the final bandwidth, example resource amount and optimal semantic rate.

[0069] In a possible implementation manner of the second aspect, if the local computing mode is selected to offload the large model computing task, the local computing delay and the local computing energy consumption are obtained, specifically:

[0070] If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are determined. The calculation formula for the local computing delay is:

[0071]

[0072] In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing latency;

[0073] The local computing energy consumption is:

[0074]

[0075] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 : A flow chart of an embodiment of a method for offloading large model computing tasks and optimizing semantic coding provided by the present invention;

[0077] Figure 2 : A schematic diagram of a BMC network model system oriented to semantic authorization for an embodiment of the large model computing task offloading and semantic coding optimization method provided by the present invention;

[0078] Figure 3 : A schematic diagram of performance convergence of combining reinforcement learning and a dichotomy algorithm for an embodiment of a large model computing task offloading and semantic coding optimization method provided by the present invention;

[0079] Figure 4 : A schematic diagram showing the system cost comparison between the optimization strategy algorithm and other strategies when the weighting factors are different and other parameter conditions are consistent, according to an embodiment of the large model computing task offloading and semantic coding optimization method provided by the present invention;

[0080] Figure 5 : A schematic diagram showing a comparison of system costs between the optimization strategy algorithm and other strategies when the semantic coding is different and other parameter conditions are consistent, according to an embodiment of the large model computing task offloading and semantic coding optimization method provided by the present invention;

[0081] Figure 6 : A schematic diagram of an embodiment of the large model computing task offloading and semantic coding optimization method provided by the present invention when the SNR of the transmitting end is different and other parameter conditions are consistent. A schematic diagram comparing the system cost of this optimization strategy algorithm with other strategies;

[0082] Figure 7 : A system structure diagram of another embodiment of the large model computing task offloading and semantic coding optimization method provided by the present invention. DETAILED DESCRIPTION

[0083] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] Embodiment 1

[0085] Please refer to Figure 1 , which is a flow chart of an embodiment of a method for offloading large model computing tasks and optimizing semantic coding provided by an embodiment of the present invention, including steps S11 to S15, each of which is specifically as follows:

[0086] S11. Receive large model computing tasks from the user.

[0087] In this embodiment, a semantically authorized large model computing network (BMC) consisting of M users and a computing node (CAP) is established. The semantic network deploys a well-known semantic communication framework (DeepSC), in which the encoder and decoder of DeepSC are integrated on the user side and the CAP side respectively to achieve semantic communication. By transmitting a pilot signal, the system estimates the wireless channel parameters. At the same time, the system collects the task size of each user, and then uses deep reinforcement learning to derive the unloading strategy of each user's large model based on the obtained channel parameters and task size. At the same time, based on the unloading strategy of each user's large model, a currently optimal semantic coding strategy is obtained through dichotomy.

[0088] A well-known semantic communication framework for transmission text (DeepSC) is deployed, in which the encoder and decoder of DeepSC are integrated on the user side and CAP side respectively to achieve data compression and efficient task transmission, thereby further reducing transmission delay and energy consumption, and strongly supporting the communication and computing of large models.

[0089] The described acquisition of BMC network parameters. In the present invention, the present invention aims to cope with dynamic BMC requirements. In order to make a dynamic large model unloading strategy, it is necessary to acquire BMC network parameters in real time, such as wireless channel parameters and the task size of the user large model.

[0090] S12. Select local computing mode or edge computing mode to unload large model computing tasks. If local computing mode is selected to unload large model computing tasks, local computing delay and local computing energy consumption will be obtained. If edge computing mode is selected to unload large model computing tasks, edge computing delay and edge computing energy consumption will be obtained.

[0091] In a preferred embodiment, if the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are obtained, specifically:

[0092] If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are determined. The calculation formula for the local computing delay is:

[0093]

[0094] In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing latency;

[0095] The local computing energy consumption is:

[0096]

[0097] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.

[0098] In a preferred embodiment, if the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are obtained, specifically:

[0099] If the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are determined. The calculation formula for the edge computing delay is:

[0100]

[0101] In the formula, is the computing power allocated to user m by the CAP side, Indicates edge computing latency;

[0102] Edge computing consumes zero energy.

[0103] In this embodiment, a semantically authorized large model computing network (BMC) consisting of M users and a computing access node (CAP) is established. The user set can be represented by At the same time, the semantic encoder and decoder are deployed on the user side and CAP side respectively to achieve semantic communication. Among them, user m will generate a large model computing task, represented as l m . Large model computing tasks m Contains mEach neural network contains an average of L parameters. User m can use the semantic encoder to extract K parameters of each parameter from the large model calculation task. m Symbols are then uploaded to the CAP end through the wireless network. Once the CAP end receives these semantic symbols, the CAP uses the corresponding semantic decoder to reconstruct the original large model, then performs calculations and provides the calculation results to the user.

[0104] S13. According to the local computing delay, edge computing delay and user transmission delay, the overall system delay is obtained, and according to the local computing energy consumption, edge computing energy consumption and user transmission energy consumption, the overall system energy consumption is obtained.

[0105] In a preferred embodiment, the overall system delay is obtained based on the local computing delay, edge computing delay and user transmission delay, specifically:

[0106] Obtain the user transmission delay, where the calculation formula for the user transmission delay is:

[0107]

[0108] In the formula, represents the user transmission delay, r m represents the semantic rate, n m represents the number of neural networks, and I represents the average amount of semantic information contained in each neural network sent to the CAP end;

[0109] According to the local computing delay, edge computing delay and user transmission delay, the overall delay of each network is obtained. The calculation formula of the overall delay of each network is:

[0110]

[0111] In the formula, represents the user transmission delay, T m represents the overall delay of each network, Indicates edge computing latency, represents the local computing delay, α m represents the binary unloading factor;

[0112] The overall system delay is obtained based on the overall delay of each network, where the overall system delay is:

[0113]

[0114] In the formula, T represents the overall system delay, m represents the Mth user, Represents a collection of users.

[0115] In a preferred embodiment, the overall system energy consumption is obtained based on local computing energy consumption, edge computing energy consumption and user transmission energy consumption, specifically:

[0116] Obtain the user transmission energy consumption, where the calculation formula of the user transmission energy consumption is:

[0117]

[0118] In the formula, Indicates the user transmission delay, represents the transmission power of user m, Indicates the user transmission energy consumption;

[0119] The overall delay of each network is obtained based on the local computing energy consumption, edge computing energy consumption and user transmission energy consumption. The calculation formula for the overall delay of each network is:

[0120]

[0121] In the formula, represents the user transmission energy consumption, E m represents the overall energy consumption of each network, represents the local computing energy consumption, α m represents the binary unloading factor;

[0122] The overall energy consumption of the system is obtained according to the overall energy consumption of each network, where the overall energy consumption of the system is:

[0123]

[0124] In the formula, E represents the overall energy consumption of the system, m represents the Mth user, Represents a collection of users.

[0125] In this embodiment, the transmission delay and energy consumption of user m are defined as:

[0126]

[0127]

[0128] In the formula, represents the user transmission delay, r m represents the semantic rate, n m represents the number of neural networks, I represents the average amount of semantic information contained in each neural network sent to the CAP end, represents the user transmission delay, represents the transmission power of user m, Indicates the user transmission energy consumption;

[0129] User m can adopt a binary offloading method and can choose between local computing mode and edge computing mode. When the user chooses local computing, the local computing latency is:

[0130]

[0131] In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing latency;

[0132] Similarly, its local computing energy consumption is:

[0133]

[0134] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.

[0135] When the user selects the edge computing mode, the task of the large model is first transmitted to the CAP end and then calculated. Then the task l from user m calculated by the CAP end m The delay consumed is:

[0136]

[0137] In the formula, is the computing power allocated to user m by the CAP side, Indicates edge computing latency;

[0138] Considering that CAP has a stable power supply, its computing energy consumption is considered negligible.

[0139] By summarizing the above two calculation situations, we can obtain the task of executing large model calculation l m The overall delay is:

[0140]

[0141] In the formula, represents the user transmission delay, T m represents the overall delay of each network, Indicates edge computing latency, represents the local computing delay, α m represents the binary unloading factor;

[0142] Similarly, we can conclude that the task of executing large model calculations is m The total energy consumption is:

[0143]

[0144] In the formula, represents the user transmission energy consumption, E m represents the overall energy consumption of each network, represents the local computing energy consumption, α m represents the binary unloading factor;

[0145] In addition, for the semantically aware BMC network, the tasks of each device are independent and can be calculated in parallel on the CAP. Therefore, the overall system latency and energy consumption of the computing tasks are:

[0146]

[0147]

[0148] In the formula, T represents the overall system delay, m represents the Mth user, represents the user set, E represents the overall energy consumption of the system, m represents the Mth user, Represents a collection of users.

[0149] S14. The system comprehensive cost is calculated based on the system overall delay and the system overall energy consumption. After a computing task offloading model is constructed based on the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy.

[0150] In a preferred embodiment, the system comprehensive cost is calculated based on the system overall delay and the system overall energy consumption. After the computing task offloading model is constructed based on the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy, which is specifically:

[0151] The system comprehensive cost is calculated based on the system overall delay and system overall energy consumption, where the calculation formula for the system comprehensive cost is:

[0152] Φ=ηT+(1-η)E

[0153] In the formula, Φ represents the comprehensive cost of the system, T represents the overall system delay, E represents the overall system energy consumption, and η∈[0,1] is the weighting factor for evaluating system delay and energy consumption;

[0154] The computing task offloading model is constructed according to the comprehensive cost of the system, where the objective function and constraints of the computing task offloading model are:

[0155]

[0156] In the formula, Φ represents the comprehensive cost of the system, w CAP is the total system bandwidth, F CAP is the total computing power of the CAP side, is the minimum local computing power, is the maximum local computing power, m represents the Mth user, and M represents the user set;

[0157] The computing task offloading model is solved through the Markov decision process to obtain the optimal offloading strategy.

[0158] In this embodiment, the system delay T and energy consumption E are key parameters for evaluating the BMC network performance. However, considering only one of the indicators cannot fully show the overall picture of the system performance. Therefore, considering the two factors of delay and energy consumption together can more comprehensively evaluate the system performance. The comprehensive cost of the system can be described as:

[0159] Φ=ηT+(1-η)E

[0160] In the formula, Φ represents the comprehensive cost of the system, T represents the overall system delay, E represents the overall system energy consumption, and η∈[0,1] is the weighting factor for evaluating system delay and energy consumption;

[0161] At the same time, for the semantic authorization BMC system, we introduce a semantic similarity threshold ψ th To ensure the quality of network transmission. Next, we study the impact of offloading strategy, encoding strategy, joint bandwidth and computing resource allocation on the system, and formulate a system cost minimization problem:

[0162]

[0163] In the formula, Φ represents the comprehensive cost of the system, w CAP is the total system bandwidth, F CAP is the total computing power of the CAP side, is the minimum local computing power, is the maximum local computing power, m represents the Mth user, Represents a collection of users;

[0164] Since P1 is a typical mixed integer nonlinear programming (MINLP), traditional methods are difficult to effectively solve this MINLP problem. Therefore, the invention adopts a strategy that combines deep reinforcement learning and binary search algorithm to optimize task offloading and encoding strategies, while evenly distributing computing power and bandwidth resources. To this end, the invention first formulates a Markov decision process (MDP) to obtain an offloading strategy. In the current BMC environment, the agent can take actions by observing the environment, and the environment will give the agent a reward and enter a new state at the same time. The agent repeatedly interacts with the environment until an optimal offloading strategy is obtained. Specifically, the states that the agent can observe are:

[0165] s t ={ht , α t-1 , n t}

[0166] In the formula, h t is the channel parameter vector at time t, α t-1 is the unloading strategy vector at time t-1, n t is the task sentence vector at time t. When the agent observes state s t , will be based on ε - Greedy criterion selects action α t , expressed as:

[0167]

[0168] Where ε is the exploration factor, φ is the neural parameter, and Q(s, a) is a network approximation of the Q function, which describes the cumulative reward after performing action a. t , the system will give a reward:

[0169]

[0170] Among them, τ1, τ2, τ3 are three positive numbers. After several interactions, the agent uses the time difference method (TD) as the loss function, which is expressed as:

[0171]

[0172] Where ξ is the discount factor and is the target neural network.

[0173] S15. Obtain the final bandwidth and example resource amount according to the optimal offloading strategy, and optimize the semantic encoding strategy using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode the large model computing tasks according to the final bandwidth, example resource amount and optimal semantic rate.

[0174] In a preferred embodiment, the final bandwidth and instance resource amount are obtained according to the optimal offloading strategy, specifically:

[0175] The optimal offloading strategy is calculated using the preset formula to obtain the final bandwidth and instance resource quantity. The preset formula is:

[0176]

[0177]

[0178]

[0179] In the formula, S o Indicates that it has been unloaded to the CAP end user set, Sl represents the user set of local calculation, |·| is the set trace operation, Indicates local computing power, is the computing power allocated to user m by the CAP side, is the minimum local computing power, The maximum value of local computing power.

[0180] In a preferred embodiment, the semantic coding strategy is optimized using the optimal unloading strategy to obtain the optimal semantic rate, specifically:

[0181] According to the optimal offloading strategy and semantic similarity, the semantic rate is calculated using a binary algorithm to obtain the optimal semantic rate.

[0182] In this embodiment, after obtaining the offloading decision, we continue to optimize the semantic encoding strategy. The semantic authorization BMC network uses the semantic rate to evaluate the transmission capacity. Considering the semantic similarity, the semantic rate can be expressed as:

[0183]

[0184] where I is the average amount of semantic information contained in each neural network sent to the CAP, represents the bandwidth allocated to user m, and represents ψ(γ m , k m )When SNR is and γ m The number of symbols for each parameter is K m The semantic similarity of .

[0185] In addition m It can be described in detail as:

[0186]

[0187] in, is the transmit power of user m, δ 2 is the variance of additive white Gaussian noise (AWGN), h m ~CN(0,β) represents the channel parameter. To simplify the notation, the symbol ψ m Denoted as ψ(γ m , k m ).

[0188] By analyzing, bandwidth and SNRγ are maintained m Under the condition of no change, increase K m The value of will result in the semantic rate r m Gradually decreases, while the semantic similarity ψ m This observation is consistent with our expectation that higher K mThe value allows more semantic symbols to be transmitted in the network, thus increasing the semantic similarity ψ m However, using more semantic symbols to maintain the same amount of semantic information I requires a lower semantic rate. m The size of the value, such as K m The value decreases, which will cause the semantic rate r m Monotonically decreasing, while semantic similarity ψ m Monotonically decreases. Then, while ensuring the semantic similarity ψ m ≥ψ th Based on this, an optimal K can be obtained through a simple binary algorithm. m , the best semantic rate has been achieved.

[0189] The binary search method is to set the search interval of psi_m, divide this interval into two, and see which of the two sub-intervals meets the requirements of semantic rate and psi_{th}. The sub-interval that meets the requirements replaces the original curve to achieve the function of halving the search range. This process continues, and the search range is halved each time until the search results no longer change or the search space is small enough.

[0190] As an example of this embodiment, this embodiment records Figure 2 The system model diagram recommended for the invention, in which 1 CAP terminal and 10 users are configured in the process of the invention. Each user and CAP has integrated semantic encoder and decoder to achieve semantic communication for large model training. The specific steps are:

[0191] Step 1: Establish a semantically authorized big model computing network (BMC) consisting of M users and a computing node (CAP). The semantic network deploys a well-known semantic communication framework (DeepSC), in which the encoder and decoder of DeepSC are integrated on the user side and CAP side respectively to achieve semantic communication;

[0192] Step 2: By transmitting a pilot signal, the system estimates the wireless channel parameters. At the same time, the system collects the task size of each user;

[0193] Step 3: Based on the channel parameters and task size in step 2, deep reinforcement learning is used to obtain the offloading strategy of each user's large model. At the same time, based on the offloading strategy of each user's large model, a current optimal semantic encoding strategy is obtained through dichotomy;

[0194] Step 4: Repeat step 2 and input the wireless channel parameters and model size output as parameters into step 3. When the system delay and energy consumption weighted converge, stop the algorithm and output the optimal strategy.

[0195] Then, a variety of different task offloading strategies are used to verify and evaluate the algorithm of the present invention. In addition, in the simulation environment, all channels experience Rayleigh flat fading, the average gain of each user's channel is 1 and the AWGN noise power is 0.01W. If not specified, the user end transmit power is 0.1W. The number of task deep neural networks n for user m m Follows a uniform distribution Each neural network contains an average of 20 parameters, each of which can be encoded into 40 bits for calculation, or encoded into k m ~[1, 20] semantic symbols for semantic transmission. The semantic transmission threshold is 0.85. The total bandwidth of the system is 1MHz, and the total computing power of the CAP end is 2.5×10 9 cycle / s. At the same time, the computing power of each user is

[0196] Figure 3 The performance convergence diagram of the proposed combined reinforcement learning and bisection algorithm is shown. Figure 2 It can be seen that the proposed algorithm can achieve good convergence performance after only 50 training rounds.

[0197] Figure 4 The system cost comparison diagram of the proposed strategy algorithm and other strategies is described under different weighting factors η and the same other parameters. Figure 4 It can be seen that no matter how much the weighting factor η is, the proposed strategy algorithm is better than other strategy algorithms. This is because the proposed algorithm can dynamically adapt to the training and calculation of the BMC network, thereby making an accurate decision to improve the overall performance of the BMC system.

[0198] Figure 5 In the semantic encoding k m Different, other parameters are the same, the system cost comparison diagram of the proposed strategy algorithm and other strategies. Figure 5 , we discovered an interesting phenomenon: when K=5, the performance of the algorithm we proposed is worse than that of the "traditional optimization" algorithm. The explanation for this phenomenon is that a smaller K value may lead to lower semantic similarity, which in turn aggravates semantic distortion and leads to the failure of the upload task. In this context, compared with the task uploading of the semantic-aware network, the effect of transmitting tasks through the bitstream network is superior. However, when K gradually increases to 20, the algorithm we proposed guides the system cost to gradually decrease, while the "traditional optimization" strategy gradually stabilizes. The explanation for this trend of change is that a larger K value helps to achieve a higher degree of semantic similarity, further improves the quality of semantic transmission, and reduces the possibility of task upload failure. In this scenario, uploading tasks through the semantic-aware network is obviously better than the bitstream network.

[0199] Figure 6 The system cost comparison diagram of the proposed strategy algorithm and other strategies is described under different transmitter SNRs and the same other parameter conditions. Figure 6 It can be seen that as the signal-to-noise ratio (SNR) increases, the overall cost of various task offloading strategies decreases. This is because a higher SNR helps to increase the transmission rate, allowing devices to upload more tasks to the cloud-edge collaboration platform (CAP) to fully utilize its powerful computing resources. In addition, the proposed strategy algorithms are all better than other strategy algorithms, which verifies the effectiveness of the proposed algorithm.

[0200] Embodiment 2

[0201] Accordingly, see Figure 7 , Figure 7 The present invention provides a large model computing task unloading and semantic coding optimization system. As shown in the figure, the large model computing task unloading and semantic coding optimization system includes:

[0202] Receiving module 701, used for receiving large model calculation tasks from the user end;

[0203] An unloading module 702 is used to select a local computing mode or an edge computing mode to unload the large model computing task. If the local computing mode is selected to unload the large model computing task, the local computing delay and the local computing energy consumption are obtained. If the edge computing mode is selected to unload the large model computing task, the edge computing delay and the edge computing energy consumption are obtained.

[0204] The first calculation module 703 is used to obtain the overall system delay according to the local calculation delay, the edge calculation delay and the user transmission delay, and obtain the overall system energy consumption according to the local calculation energy consumption, the edge calculation energy consumption and the user transmission energy consumption;

[0205] The second calculation module 704 is used to calculate the system comprehensive cost according to the system overall delay and the system overall energy consumption, and after building a computing task offloading model according to the system comprehensive cost, solve the computing task offloading model to obtain the optimal offloading strategy;

[0206] The third computing module 705 is used to obtain the final bandwidth and example resource amount according to the optimal offloading strategy, and optimize the semantic coding strategy using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode the large model computing tasks according to the final bandwidth, example resource amount and optimal semantic rate.

[0207] In a preferred embodiment, if the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are obtained, specifically:

[0208] If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are determined. The calculation formula for the local computing delay is:

[0209]

[0210] In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing latency;

[0211] The local computing energy consumption is:

[0212]

[0213] In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.

[0214] The more detailed working principle and step flow of this embodiment can refer to, but are not limited to, the relevant records of Embodiment 1.

[0215] In summary, the implementation of the embodiments of the present invention has the following beneficial effects:

[0216] Receive the large model computing task from the user end, select the local computing mode or the edge computing mode to offload the large model computing task. If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are obtained. If the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are obtained. The overall system delay is obtained based on the local computing delay, edge computing delay and user transmission delay. The overall system energy consumption is obtained based on the local computing energy consumption, edge computing energy consumption and user transmission energy consumption. The overall system delay and overall system energy consumption are calculated to obtain the system comprehensive cost. The overall system cost is calculated based on the overall system delay and overall system energy consumption. After constructing the computing task offloading model, the computing task offloading model is solved to obtain the optimal offloading strategy. The final bandwidth and example resource amount are obtained according to the optimal offloading strategy, and the semantic encoding strategy is optimized using the optimal offloading strategy to obtain the optimal semantic rate, so that the system can offload and encode the large model computing tasks according to the final bandwidth, example resource amount and optimal semantic rate. This method optimizes the large model computing task offloading strategy and semantic encoding strategy, so that the system can extract low-dimensional semantic information from the high-dimensional tasks of the large model for transmission, reducing the communication delay and energy consumption overhead, thereby improving the communication performance and supporting the development and application of large models.

[0217] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for offloading large model computing tasks and optimizing semantic coding, characterized in that: include: Receive large model computing tasks from the user; Select local computing mode or edge computing mode to unload the large model computing task. If the local computing mode is selected to unload the large model computing task, the local computing delay and local computing energy consumption are obtained. If the edge computing mode is selected to unload the large model computing task, the edge computing delay and edge computing energy consumption are obtained. According to the local computing delay, the edge computing delay and the user transmission delay, the overall system delay is obtained, and according to the local computing energy consumption, the edge computing energy consumption and the user transmission energy consumption, the overall system energy consumption is obtained; The system comprehensive cost is calculated according to the overall system delay and the overall system energy consumption, and after a computing task offloading model is constructed according to the system comprehensive cost, the computing task offloading model is solved to obtain an optimal offloading strategy; Obtaining a final bandwidth and a calculation example resource amount according to the optimal offloading strategy, and optimizing a semantic encoding strategy using the optimal offloading strategy to obtain an optimal semantic rate, so that the system can offload and encode the large model computing task according to the final bandwidth, the calculation example resource amount, and the optimal semantic rate; The system comprehensive cost is calculated based on the overall system delay and the overall system energy consumption. After the computing task offloading model is constructed based on the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy, which is specifically: The system comprehensive cost is calculated based on the overall system delay and the overall system energy consumption, wherein the calculation formula of the system comprehensive cost is: Φ=ηT+(1-η)E In the formula, Φ represents the comprehensive cost of the system, T represents the overall system delay, E represents the overall system energy consumption, and η∈[0,1] is the weighting factor for evaluating system delay and energy consumption; A computing task offloading model is constructed according to the system comprehensive cost, wherein the objective function and constraint conditions of the computing task offloading model are: Where Φ represents the system comprehensive cost, W CAP is the total system bandwidth, F CAP is the total computing power of the CAP side, is the minimum local computing power, is the maximum local computing power, m represents the Mth user, represents the user set, α m represents the binary unloading factor, k m Represents a parameter symbol in a neural network, represents the bandwidth allocated to user m, Indicates local computing power, is the computing power allocated to user m by the CAP side, ψ th Semantic similarity threshold, ψ m Semantic similarity; The final bandwidth and instance resource amount are obtained according to the optimal offloading strategy, specifically: Solving the computing task offloading model through a Markov decision process to obtain an optimal offloading strategy; The optimal offloading strategy is calculated using a preset formula to obtain the final bandwidth and the amount of resources for the example, wherein the preset formula is: In the formula, S o Indicates that it has been unloaded to the CAP end user set, S l represents the user set of local calculation, |·| is the set trace operation, Indicates local computing power, is the computing power allocated to user m by the CAP side, is the minimum local computing power, is the maximum local computing power, F CAP is the total computing power of the CAP side, W CAP is the total system bandwidth, represents the bandwidth allocated to user m; The optimal unloading strategy is used to optimize the semantic coding strategy to obtain the optimal semantic rate, specifically: According to the optimal offloading strategy and semantic similarity, the semantic rate is calculated using a binary algorithm to obtain the optimal semantic rate.

2. The large model computing task offloading and semantic coding optimization method according to claim 1, characterized in that: If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are obtained, which are specifically: If the local computing mode is selected to offload the large model computing task, the local computing delay and the local computing energy consumption are determined. The calculation formula of the local computing delay is: In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing delay, n m represents the number of neural networks, and L represents the number of parameters in a neural network; The local computing energy consumption is: In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.

3. The large model computing task offloading and semantic coding optimization method according to claim 1, characterized in that: If the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are obtained, specifically: If the edge computing mode is selected to offload the large model computing task, the edge computing delay and edge computing energy consumption are determined, wherein the calculation formula for the edge computing delay is: In the formula, is the computing power allocated to user m by the CAP side, represents edge computing latency, n m represents the number of neural networks, and L represents the number of parameters in a neural network; The edge computing energy consumption is 0.

4. The large model computing task offloading and semantic coding optimization method according to claim 1, characterized in that: The overall system delay is obtained according to the local computing delay, the edge computing delay and the user transmission delay, specifically: Obtain the user transmission delay, wherein the calculation formula of the user transmission delay is: In the formula, represents the user transmission delay, r m represents the semantic rate, n m represents the number of neural networks, and I represents the average amount of semantic information contained in each neural network sent to the CAP end; According to the local computing delay, the edge computing delay and the user transmission delay, the overall delay of each network is obtained, wherein the calculation formula of the overall delay of each network is: In the formula, represents the user transmission delay, T m represents the overall delay of each network, Indicates edge computing latency, represents the local computing delay, α m represents the binary unloading factor; The overall system delay is obtained according to the overall delay of each network, wherein the overall system delay is: In the formula, T represents the overall system delay, m represents the Mth user, Represents a collection of users.

5. The large model computing task offloading and semantic coding optimization method according to claim 1, characterized in that: The overall energy consumption of the system is obtained according to the local computing energy consumption, the edge computing energy consumption and the user transmission energy consumption, specifically: Obtain user transmission energy consumption, wherein the calculation formula of the user transmission energy consumption is: In the formula, represents the user transmission delay, represents the transmission power of user m, represents the transmission energy consumption of the user; The overall delay of each network is obtained according to the local computing energy consumption, the edge computing energy consumption and the user transmission energy consumption, wherein the calculation formula of the overall delay of each network is: In the formula, represents the user transmission energy consumption, E m represents the overall energy consumption of each network, represents the local computing energy consumption, α m represents the binary unloading factor; The overall energy consumption of the system is obtained according to the overall energy consumption of each network, wherein the overall energy consumption of the system is: In the formula, E represents the overall energy consumption of the system, m represents the Mth user, Represents a collection of users.

6. A large model computing task offloading and semantic coding optimization system, characterized in that: include: A receiving module, used for receiving large model computing tasks from the user end; An unloading module, used to select a local computing mode or an edge computing mode to unload the large model computing task. If the local computing mode is selected to unload the large model computing task, the local computing delay and the local computing energy consumption are obtained. If the edge computing mode is selected to unload the large model computing task, the edge computing delay and the edge computing energy consumption are obtained. A first calculation module, configured to obtain a system overall delay according to the local calculation delay, the edge calculation delay, and the user transmission delay, and obtain a system overall energy consumption according to the local calculation energy consumption, the edge calculation energy consumption, and the user transmission energy consumption; A second calculation module is used to calculate the system comprehensive cost according to the overall system delay and the overall system energy consumption, and after building a computing task offloading model according to the system comprehensive cost, solve the computing task offloading model to obtain an optimal offloading strategy; A third computing module is used to obtain a final bandwidth and a calculation example resource amount according to the optimal offloading strategy, and optimize the semantic encoding strategy using the optimal offloading strategy to obtain an optimal semantic rate, so that the system can offload and encode the large model computing task according to the final bandwidth, the calculation example resource amount and the optimal semantic rate; The system comprehensive cost is calculated based on the overall system delay and the overall system energy consumption. After the computing task offloading model is constructed based on the system comprehensive cost, the computing task offloading model is solved to obtain the optimal offloading strategy, which is specifically: The system comprehensive cost is calculated based on the overall system delay and the overall system energy consumption, wherein the calculation formula of the system comprehensive cost is: Φ=ηT+(1-η)E In the formula, Φ represents the comprehensive cost of the system, T represents the overall system delay, E represents the overall system energy consumption, and η∈[0,1] is the weighting factor for evaluating system delay and energy consumption; A computing task offloading model is constructed according to the system comprehensive cost, wherein the objective function and constraint conditions of the computing task offloading model are: Where Φ represents the system comprehensive cost, W CAP is the total system bandwidth, F CAP is the total computing power of the CAP side, is the minimum local computing power, is the maximum local computing power, m represents the Mth user, represents the user set, α m represents the binary unloading factor, k m Represents a parameter symbol in a neural network, represents the bandwidth allocated to user m, Indicates local computing power, is the computing power allocated to user m by the CAP side, ψ th Semantic similarity threshold, ψ m Semantic similarity; The final bandwidth and instance resource amount are obtained according to the optimal offloading strategy, specifically: Solving the computing task offloading model through a Markov decision process to obtain an optimal offloading strategy; The optimal offloading strategy is calculated using a preset formula to obtain the final bandwidth and the amount of resources for the example, wherein the preset formula is: In the formula, S o Indicates that it has been unloaded to the CAP end user set, S l represents the user set of local calculation, |·| is the set trace operation, Indicates local computing power, is the computing power allocated to user m by the CAP side, is the minimum local computing power, is the maximum local computing power, F CAP is the total computing power of the CAP side, W CAP is the total system bandwidth, represents the bandwidth allocated to user m; The optimal unloading strategy is used to optimize the semantic coding strategy to obtain the optimal semantic rate, specifically: According to the optimal offloading strategy and semantic similarity, the semantic rate is calculated using a binary algorithm to obtain the optimal semantic rate.

7. The large model computing task offloading and semantic coding optimization system according to claim 6, characterized in that: If the local computing mode is selected to offload the large model computing task, the local computing delay and local computing energy consumption are obtained, which are specifically: If the local computing mode is selected to offload the large model computing task, the local computing delay and the local computing energy consumption are determined. The calculation formula of the local computing delay is: In the formula, ω represents the number of bits corresponding to the parameters of a neural network, ∈ represents the CPU frequency required to calculate each bit, Indicates local computing power, Indicates the local computing delay, n m represents the number of neural networks, and L represents the number of parameters in a neural network; The local computing energy consumption is: In the formula, represents the local computing energy consumption, K represents the computing energy consumption coefficient, Indicates the local computing latency.