QoS (Quality of Service) measurement and uncertainty evaluation method for edge cloud scene
By using Bernoulli sampling and GNN-variant Bayesian encoder in edge cloud, the problem of network service quality fluctuations in edge cloud is solved, efficient QoS measurement and uncertainty evaluation are achieved, and sampling is reduced Cost, improve prediction accuracy and data recovery performance.
Patent Information
- Application Number
- CN202410886683.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-03
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Device establishment based on heterogeneous resources in edge clouds may cause significant fluctuations in network service quality, and an effective QoS measurement and uncertainty evaluation method is needed to dynamically select reliable edge nodes to realize real-time network resource scheduling.
Using Bernoulli sampling and GNN-variant Bayesian encoder methods, sparse measurement and completion of real-world edge service data sets, using multi-layer perceptron decoder to obtain real QoS, and evaluate through relative RMSE. Restore accuracy.
It improves prediction accuracy, achieves superior data recovery performance at the same sampling rate, reduces the overall sampling cost, and is reasonable and universal applicable.
Smart Images

Figure CN119094385A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of network service technology, and in particular to a QoS measurement and uncertainty assessment method for edge cloud scenarios. Background Art
[0002] The rise of artificial intelligence applications (such as ChatGPT) has triggered a huge demand for computing resources to train large-scale models, which has brought serious workload pressure to data centers. As an alternative solution, emerging edge computing (such as distributed edge cloud) makes full use of decentralized computing resources to provide low-latency and high-quality network services.
[0003] The network service quality of edge cloud relies on densely deployed edge nodes, which physically bring edge servers closer to users. Existing studies have found that network service latency is only slightly affected by network fluctuations when the distance between the server and the user is within 30 kilometers. Compared with centralized clouds, edge clouds can reduce network latency by 84.1% and latency fluctuations by 0.5 milliseconds. However, achieving this latency advantage requires a lot of infrastructure investment for dense deployment, which poses a huge budget challenge to network service providers.
[0004] Using existing resources to build a densely distributed edge cloud can reduce the cost of node construction. For example, network service providers can use crowdsourcing technology to use personal computers or mobile phones that are willing to provide network services to nearby users as edge nodes. However, edge clouds built on existing devices with heterogeneous resources (such as network bandwidth and computing power) may cause significant fluctuations in the quality of network services. Therefore, it is necessary to measure QoS (quality of service) and dynamically select reliable edge nodes to provide network services, thereby achieving real-time network resource scheduling. Summary of the invention
[0005] The purpose of the present invention is to provide a QoS measurement and uncertainty assessment method for edge cloud scenarios, which effectively utilizes the spatiotemporal correlation in the data set for learning and inference, improves the prediction accuracy, and achieves superior data recovery performance at the same sampling rate; based on the sparse measurement method, it reduces the overall sampling cost and is reasonable and generally applicable in practical application scenarios.
[0006] To achieve the above object, the present invention provides a QoS measurement and uncertainty assessment method for edge cloud scenarios, comprising the following steps:
[0007] S1. Collection of real-world edge service datasets;
[0008] S2, perform QoS sampling on the above data set based on the Bernoulli sampling method;
[0009] S3, use GNN-Variational Bayes encoder to complete QoS;
[0010] S4, a decoder based on a multi-layer perceptron obtains the real QoS;
[0011] S5. Evaluate the uncertainty of the obtained QoS.
[0012] Preferably, in step S1, the collected data sets include an EEL data set and a FCTE data set.
[0013] Preferably, in step S2, for a given sampling rate α, the Bernoulli sampling method is used for sampling, which is an approximate uniform distribution, ensuring the expected number of samples of n×n×α, which meets the required sampling size, and the sampling mask is represented as follows:
[0014]
[0015] in, is uniformly distributed.
[0016] Preferably, in step S3, a GNN-variational Bayes encoder is used to complete the QoS, and the specific process is as follows:
[0017] S31. First, a Bayesian edge-enhanced graph convolutional network is designed to extract the characteristics of network service quality, so that the potential representation distribution of each edge cloud node is generated as follows:
[0018]
[0019] S32. Based on the Bayesian neural network, a known distribution is found for the prior distribution, and the prior distribution of the node feature representation is set to the standard normal distribution, as shown below:
[0020]
[0021] S33. To find the posterior distribution of the potential representation of the node Then use the variational distribution to approximate the posterior distribution as follows
[0022]
[0023] Among them, There are two parameters μ t ={μ i} and σ t ={σ i}, use stacked graph convolutional networks to generate μ and σ; expand GraphSAGE to Edge-enhancedGraphSAG, and redefine the aggregation function AGGv in GraphSAGE as follows:
[0024]
[0025] Among them, l represents the current layer index of the EGS model, represents the neighbor node of node v, σ represents the sigmoid function, represents the learnable parameters of the current layer, C is the connection function used to combine two vectors, represents the feature representation of node j at layer l, e ij represents the network service quality value from node i to node j;
[0026] S34, the hidden state of the lth layer is as follows:
[0027]
[0028] Among them, W2 is another trainable parameter;
[0029] S35. Define κ(·) as the stacking operator, and the latent representation vectors of μ and σ as follows:
[0030]
[0031] S36, the original potential representation is as follows:
[0032] Z=μ+σ⊙∈
[0033] Among them, ∈~N(0,1).
[0034] Preferably, in step S4, according to the obtained representation distribution of each edge cloud node Z, a decoder based on a multi-layer perceptron is used to decode the real QoS from the distribution. The specific process is as follows:
[0035] S41, for ∈ i ~N(0,1) to obtain the representation The specific value of
[0036] S42. Design an inner product model with RELU activation function to convert the high-dimensional representation matrix into the original QoS matrix as shown below:
[0037]
[0038] Preferably, in step S5, relative-RMSE is used to evaluate the recovery accuracy based on the data set collected in step S1, as shown below:
[0039]
[0040] Therefore, the present invention adopts the above-mentioned QoS measurement and uncertainty assessment method for edge cloud scenarios, effectively utilizes the spatiotemporal correlation in the data set for learning and inference, improves the prediction accuracy, and achieves superior data recovery performance at the same sampling rate; based on the sparse measurement method, it reduces the overall sampling cost, and is reasonable and generally applicable in actual application scenarios.
[0041] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a flow chart of a QoS measurement and uncertainty assessment method for edge cloud scenarios of the present invention;
[0043] Figure 2 It is a framework diagram of network service quality sampling and collection of the present invention;
[0044] Figure 3 is the relative RMSE value of each baseline model after the recovery error converges. DETAILED DESCRIPTION
[0045] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0046] As shown in the figure, the present invention provides a QoS measurement and uncertainty assessment method for edge cloud scenarios, comprising the following steps:
[0047] S1. Collection of real-world edge service datasets;
[0048] S2, perform QoS sampling on the above data set based on the Bernoulli sampling method;
[0049] S3, use GNN-Variational Bayes encoder to complete QoS;
[0050] S4, a decoder based on a multi-layer perceptron obtains the real QoS;
[0051] S5. Evaluate the uncertainty of the obtained QoS.
[0052] Example
[0053] S1. Collection of real-world edge service datasets.
[0054] The data sets collected in this embodiment include the EEL data set and the FCTE data set. The EEL data set was collected in cooperation with PPIO, a national edge cloud service provider in China; the data set contains 5174 end-to-end latency measurements performed on a crowdsourced edge cloud, with a total of 943,191,155 data rows. The FCTE data set was measured on a leading public edge platform that has been densely deployed nationwide, and contains 331,098,145 data rows.
[0055] S2. Perform QoS sampling on the above data set based on the Bernoulli sampling method.
[0056] In order to achieve low-cost service quality measurement, the network service quality is measured by sampling and prediction. In order to ensure that the model can better capture the characteristics of the entire QoS matrix, the entire QoS matrix is uniformly sampled. Based on the fact that the factors affecting the network service quality are unified in the entire edge cloud, the QoS sampling matrix is assumed to be independent and identically distributed. In order to facilitate feasibility analysis and ensure consistency, the consistency of sampling is also crucial.
[0057] For a given sampling rate α, the Bernoulli sampling method is used for sampling as an approximate uniform distribution. This ensures the expected number of samples of n×n×α, which meets the required sampling size. Therefore, the sampling mask is represented as follows:
[0058]
[0059] in, is uniformly distributed.
[0060] S3. Use the GNN-Variational Bayes encoder to complete the QoS operation.
[0061] like Figure 2 As shown in the figure, the sampling of network service quality is realized. Next, the GNN-Variational Bayes encoder is used to complete the QoS implementation. The specific process is as follows:
[0062] S31. First, a Bayesian edge-enhanced graph convolutional network is designed to extract the characteristics of network service quality, so that the potential representation distribution of each edge cloud node is generated as follows:
[0063]
[0064] S32. Based on the Bayesian neural network, find a known distribution for the prior distribution. Since the node feature is a continuous value and the QoS inference problem is a regression problem, the prior distribution represented by the node feature is set to a standard normal distribution, as shown below:
[0065]
[0066] S33. To find the posterior distribution of the potential representation of the node Then use the variational distribution to approximate the posterior distribution as follows
[0067]
[0068] Among them, There are two parameters μ t ={μ i} and σ t ={σ i}, use stacked graph convolutional networks to generate μ and σ. GraphSAGE is extended to Edge-enhanced GraphSAG (EGS), and the aggregation function AGGv in GraphSAGE is redefined as follows:
[0069]
[0070] Among them, l represents the current layer index of the EGS model, represents the neighbor node of node v, σ represents the sigmoid function, represents the learnable parameters of the current layer, C is the connection function used to combine two vectors, represents the feature representation of node j at layer l, e ij Represents the network service quality value from node i to node j.
[0071] S34, the hidden state of the lth layer is as follows:
[0072]
[0073] Among them, W2 is another trainable parameter.
[0074] S35. Define κ(·) as the stacking operator, and the latent representation vectors of μ and σ as follows:
[0075]
[0076] S36, the original potential representation is as follows:
[0077] Z=μ+σ⊙∈
[0078] Among them, ∈~N(0,1).
[0079] S4. The decoder based on multi-layer perceptron obtains the real QoS.
[0080] According to the obtained representation distribution of each edge cloud node Z, the decoder based on the multi-layer perceptron decodes the real QoS from the distribution. The specific process is as follows:
[0081] S41, for ∈ i ~N(0,1) to obtain the representation The specific value of .
[0082] S42. Design an inner product model with RELU activation function to convert the high-dimensional representation matrix into the original QoS matrix as shown below:
[0083]
[0084] S5. Evaluate the uncertainty of the obtained QoS.
[0085] Based on the data set collected in step S1, relative-RMSE is used to evaluate the recovery accuracy as follows:
[0086]
[0087] As shown in Table 1, by comparing with each baseline model, it can be observed that the present invention achieves the best sampling performance at the same sampling rate. This shows that the present invention effectively utilizes the spatiotemporal correlation within the data set for learning and inference, thereby achieving superior data recovery performance compared to other models at the same sampling rate.
[0088] We further found that the error rate was reduced by at least 80% on the EEL dataset and 17% on the FCTE dataset. The slight difference in the recovery performance between the two datasets is attributed to the partial data loss in the FCTE dataset, which causes the actual sampling rate to be lower than the estimated sampling rate.
[0089] By comparing the columns in Table 1, we can observe different degrees of sampling error convergence. Regardless of the filling method used, including mean filling, mathematical filling methods (KNN, SVD, MICE, Spectral) or machine learning-based methods (GCN, GAT, Ours), the recovery error tends to converge to a fixed value as the sampling rate increases. This shows that for each sampling rate, there is an optimal sampling upper limit, beyond which increasing the sampling rate does not significantly improve the final prediction results. This also emphasizes the rationality behind sparse sampling and shows the effectiveness of the present invention in practical applications.
[0090] Table 1 Performance comparison
[0091]
[0092] Secondly, the present invention performs well in adapting to low sampling rates, resulting in significant cost reduction. Based on this, the relative RMSE value of each baseline model after the recovery error converges is measured, such as Figure 3As shown. Compared with other models, the present invention only needs about 20% data sampling on the EEL dataset to achieve higher recovery accuracy, and about 50% data sampling on the FCTE dataset to achieve the best recovery accuracy. The convergence sampling rate of the present invention is at least 20% lower than that of other models, indicating that the sampling cost is reduced by at least 20%.
[0093] Therefore, the present invention adopts the above-mentioned QoS measurement and uncertainty assessment method for edge cloud scenarios, effectively utilizes the spatiotemporal correlation in the data set for learning and inference, improves the prediction accuracy, and achieves superior data recovery performance at the same sampling rate; based on the sparse measurement method, it reduces the overall sampling cost, and is reasonable and generally applicable in actual application scenarios.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A QoS measurement and uncertainty assessment method for edge cloud scenarios, characterized in that: The following steps are involved: S1. Collection of real-world edge service datasets; S2, perform QoS sampling on the above data set based on the Bernoulli sampling method; S3, use GNN-Variational Bayes encoder to complete QoS; S4, a decoder based on a multi-layer perceptron obtains the real QoS; S5. Evaluate the uncertainty of the obtained QoS.
2. The QoS measurement and uncertainty assessment method for edge cloud scenarios according to claim 1, characterized in that: In step S1 , the collected data sets include an EEL data set and a FCTE data set.
3. According to claim 1, a QoS measurement and uncertainty assessment method for edge cloud scenarios is characterized in that: In step S2, for a given sampling rate α, the Bernoulli sampling method is used for sampling. As an approximate uniform distribution, the expected number of samples of n×n×α is ensured, which meets the required sampling size. The sampling mask is expressed as follows: in, is uniformly distributed.
4. According to claim 1, a QoS measurement and uncertainty assessment method for edge cloud scenarios is characterized in that: In step S3, the GNN-Variational Bayes encoder is used to complete the QoS. The specific process is as follows: S31. First, a Bayesian edge-enhanced graph convolutional network is designed to extract the characteristics of network service quality, so that the potential representation distribution of each edge cloud node is generated as follows: S32. Based on the Bayesian neural network, a known distribution is found for the prior distribution, and the prior distribution of the node feature representation is set to the standard normal distribution, as shown below: S33. To find the posterior distribution of the potential representation of the node Then use the variational distribution to approximate the posterior distribution as follows Among them, There are two parameters μ t ={μ i } and σ t ={σ i }, use stacked graph convolutional networks to generate μ and σ; expand GraphSAGE to Edge-enhanced GraphSAG, and redefine the aggregation function AGGv in GraphSAGE as follows: Among them, l represents the current layer index of the EGS model, represents the neighbor node of node v, σ represents the sigmoid function, represents the learnable parameters of the current layer, C is the connection function used to combine two vectors, x j l represents the feature representation of node j at layer l, e ij represents the network service quality value from node i to node j; S34, the hidden state of the lth layer is as follows: Among them, W2 is another trainable parameter; S35. Define κ(·) as the stacking operator, and the latent representation vectors of μ and σ as follows: S36, the original potential representation is as follows: Z=μ+σ☉∈ Among them, ∈~N(0,1).
5. The QoS measurement and uncertainty assessment method for edge cloud scenarios according to claim 1, characterized in that: In step S4, according to the obtained representation distribution of each edge cloud node Z, the real QoS is decoded from the distribution based on the decoder of the multi-layer perceptron. The specific process is as follows: S41, for ∈ i ~N(0,1) to obtain the representation The specific value of S42. Design an inner product model with RELU activation function to convert the high-dimensional representation matrix into the original QoS matrix as shown below:
6. The QoS measurement and uncertainty assessment method for edge cloud scenarios according to claim 1, characterized in that: In step S5, based on the data set collected in step S1, relative-RMSE is used to evaluate the recovery accuracy as follows:
