A cost-sensitive edge-assisted live streaming method and system

Abstract

The application discloses a cost-sensitive edge-assisted live streaming method and system, and proposes a cost-sensitive edge-assisted live streaming method and system, which utilizes a Shapley value estimation method to model actual bandwidth costs of a content distribution network and an edge server in different time periods under 95 peak pricing; meanwhile, based on the estimated Shapley value, for a high-computational-complexity online request scheduling problem, a plurality of acceleration technologies are utilized to reduce computational overhead without sacrificing performance, and a request scheduling greedy algorithm with theoretical guarantee is used for solving, so that a coupling problem between cost modeling and request scheduling is solved, thereby effectively reducing bandwidth costs of the edge-assisted live streaming system. Compared with other methods, the system and method have significant advantages in improving transmission capacity of a current live streaming system and saving bandwidth costs.

Description

Technical Field

[0001] This invention relates to the field of video transmission technology, and more particularly to cost-sensitive edge-assisted live streaming methods and systems. Background Technology

[0002] Live streaming is becoming increasingly common in our daily lives. To ensure stability and video quality, live streaming systems widely utilize Content Delivery Network (CDN) services to accelerate video distribution. However, the unprecedented number of users, the massive amount of high-definition video content, and emerging video types are driving a rapid increase in the bandwidth demands of live streaming systems, resulting in extremely high bandwidth costs. On the other hand, huge traffic spikes pose challenges to the live streaming capabilities of CDNs. Therefore, improving the transmission capacity of live streaming systems while reducing the bandwidth costs of live streaming CDNs is of great significance.

[0003] As an emerging paradigm, edge computing provides a better way to deliver video to live streaming systems by providing edge node services (ENS) that are closer to end users. For example, DeepCast (F. Wang, C. Zhang, F. Wang, J. Liu, Y. Zhu, H. Pang, and L. Sun, “Intelligent edge-assisted crowdcast with deep reinforcement learning for personalized QoE,” in IEEE INFOCOM 2019-IEEE Conference on Computer Communications, 2019, pp. 910–918.) improves the Quality of Experience (QoE) and reduces computational and bandwidth costs. The AGM (X. Chen, C. Xu, M. Wang, Z. Wu, S. Yang, L. Zhong, and G.-M. Muntean, “A universal transcoding and transmission method for livecast with networked multi-agent reinforcement learning,” in IEEE INFOCOM 2021-IEEE Conference on Computer Communications, 2021, pp. 1–10.) algorithm optimizes the joint scheduling problem of transcoding and delivery.

[0004] However, these works employ a data volume pricing model, which limits their application in operational systems, as many large-scale live streaming systems use the 95% peak pricing model. The 95% peak pricing model specifically refers to sampling the total bandwidth usage for each time slot (typically 5 minutes) within a billing cycle (usually one month), and then using 95% of all sampled values ​​as the pricing model for billed bandwidth.

[0005] Regarding this pricing model, TrafficShaper (W. Li, X. Zhou, K. Li, H. Qi, and D. Guo, “Trafficshaper: Shaping inter-datacenter traffic to reduce the transmission cost,” IEEE / ACM Transactions on Networking, vol. 26, no. 3, pp. 1193–1206, 2018.) reduces bandwidth costs by delaying data transmission; CASCARA (R. Singh, S. Agarwal, M. Calder, and P. Bahl, “Cost-effective cloud edge traffic engineering with cascara,” in 18th USENIX Symposium on Networked Systems Design and Implementation (NSDI 21). USENIX Association, Apr. 2021, pp. 201–216.) and OnTPC (H. Chen, H. Zhan, H. Tan, H. Xu, W. Shan, S. Chen, and X.-Y. Li, “Online traffic allocation based on…”)… "On percentile charging for practical CDNs," in the 2022 IEEE / ACM 30th International Symposium on Quality of Service (IWQoS), 2022, pp. 1–10., demonstrates how to rationally distribute traffic across multiple links by accurately estimating the 95% peak of bandwidth usage.

[0006] However, these methods are primarily designed for data center networks or traditional CDNs. They are not suitable for real-time streaming systems because real-time streaming requires real-time transmission, and these methods lack effective long-term bandwidth prediction techniques.

[0007] Therefore, under the 95% peak pricing model, it is crucial to design an effective method to minimize the bandwidth cost of edge-assisted live streaming systems. Summary of the Invention

[0008] The purpose of this invention is to propose a cost-sensitive edge-assisted live streaming method and system under the 95-peak pricing model, which minimizes bandwidth costs while reducing computational overhead, improves the transmission capacity of current live streaming systems, and solves the problem of excessively high bandwidth costs.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows:

[0010] On the one hand, the present invention provides a cost-sensitive edge-assisted live streaming method, which uses the Shapley value to model the contribution of bandwidth usage of each time slot to the overall bandwidth cost, and uses a greedy algorithm to solve the scheduling request based on the estimated Shapley value, and outputs the scheduling decision;

[0011] Specifically, the Monte Carlo method is used to calculate the Shapley value of bandwidth usage by the content delivery network and edge servers in each time slot t, and the Shapley value of participant i. The specific definitions are as follows:

[0012]

[0013] Where S N Let S(π,i) represent all possible permutations of set N, S(π,i) represent the set of participants who arrive no later than i in order π, v(S(π,i)) represent the cost contribution of a participant, and v(S(π,i)\{i}) represent the cost contribution of participants other than i.

[0014] By arranging the broadband usage order Sampling is performed on k subsets, using Then, the marginal cost increment is averaged, and expressed as:

[0015]

[0016] in, This represents the bandwidth usage of the first t time slots e, where e represents the edge server and t represents the time slot;

[0017] The Monte Carlo method is accelerated using the following three methods:

[0018] S11. Use 95th percentile estimation to eliminate half of the sorting operations;

[0019] S12. Use interpolation methods to reuse the Shapley value estimation results;

[0020] S13. Use vectorized operations to complete all repetitive operations at once.

[0021] Furthermore, in S11, the specific process of using 95th percentile estimation to eliminate half of the sorting operation includes:

[0022] Sort the sampled subset S(π,i)\{i} to obtain its i-th and the The larger numbers are denoted as a and b respectively;

[0023] Suppose that the i-th largest number in π is x, then the 95-degree peak of the two sequences S(π,i)\{i} and S(π,i) is:

[0024] Q(S(π,i)\{i},0.95)=a

[0025]

[0026] Where S(π,i)\{i} represents the set after removing i from the set S(π,i), and S(π,i) represents the participants in the sequence π who are before i and are in the set S.

[0027] Furthermore, in S12, the specific process of using interpolation methods to reuse the Shapley value estimation results includes:

[0028] For content delivery networks, from 0 to... 1000 values ​​were sampled uniformly within the range;

[0029] For edge servers, the bandwidth range is from 0 to the edge server's maximum bandwidth C. e Sample 1000 values ​​within the range;

[0030] The Shapley value for the bandwidth usage u' to be estimated is:

[0031]

[0032] Among them, u i and p i These represent the sampling of bandwidth usage and the corresponding Shapley value, respectively.

[0033] Furthermore, in S13, the specific process of using vectorization to complete all repetitive operations at once includes:

[0034] (i) In sampling using the Monte Carlo method, Connection as Then, all Monte Carlo samples are generated at once using a vectorized permutation operation.

[0035] (ii) In the Shapley value estimation, the pre-calculated Shapley values ​​are concatenated and interpolated for all edge servers and all channels at once.

[0036] Furthermore, the process of the greedy algorithm solving for the scheduling request includes:

[0037] S21, Scheduling Decision Initialized as a zero vector, the Shapley value is estimated using the vectorization operation and the 95 peak estimation interpolation sample method. Before each iteration task, a combination of candidate edge servers and channels, P = {(e′, s′)}, is constructed, where e represents the edge server, s represents the live channel, and e′ and s′ represent the candidate edge node and the candidate live channel, respectively.

[0038] S22. Calculate all Shapley values ​​using vectorized interpolation. The value is assigned to 1, which means that the candidate live channel s′ in time slot t is distributed by the candidate edge server e′, and it is guaranteed that (e′,s′)∈P will not cause e to exceed the bandwidth limit;

[0039] S23. Select the combination (e′, s′) that most significantly reduces bandwidth costs, and... If the selected combination (e′, s′) cannot reduce the cost, the iteration process ends, and y is eventually assigned a value of 1. t Return it as a result.

[0040] Furthermore, in S23, The cost reduction after assigning a value of 1 is represented as:

[0041]

[0042] in, This refers to scheduling decisions, specifically whether channel s in time slot t is distributed by the edge server, and y t It is a binary vector, where 1 indicates that channel s in time slot t is distributed by edge server e, and 0 indicates that channel s in time slot t is not distributed by edge server e. Indicates in y t Lieutenant General 1 is assigned to The result is C(y) t ) indicates the use of scheduling decision y t The target cost is given by e′ and s′, which represent the candidate edge nodes and candidate live streaming channels, respectively.

[0043] On the other hand, the present invention also provides a cost-sensitive edge-assisted live streaming system, comprising the following modules for implementing the cost-sensitive edge-assisted live streaming method described in any of the above claims:

[0044] The Shapley value estimation module uses the Shapley value to model the contribution of bandwidth usage for each time slot to the overall bandwidth cost;

[0045] The scheduling module requests a scheduling request and, based on the estimated Shapley value, uses a greedy algorithm to solve the scheduling request and outputs a scheduling decision.

[0046] The Shapley value estimation module uses the Monte Carlo method to calculate the Shapley value of the bandwidth used by the content delivery network and edge servers in each time slot t, while employing the following three methods to accelerate the Monte Carlo method:

[0047] S11. Use 95th percentile estimation to eliminate half of the sorting operations;

[0048] S12. Use interpolation methods to reuse the Shapley value estimation results;

[0049] S13. Use vectorized operations to complete all repetitive operations at once.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention proposes a cost-sensitive edge-assisted live streaming method and system. Under peak pricing (95%), it models the actual bandwidth costs of the content delivery network and edge servers at different time periods using the Shapley value estimation method. Simultaneously, based on the estimated Shapley value, it addresses the computationally complex online request scheduling problem by employing various acceleration techniques to reduce computational overhead without sacrificing performance. A theoretically guaranteed greedy algorithm for request scheduling is used to solve the problem, resolving the coupling issue between cost modeling and request scheduling, thereby effectively reducing the bandwidth cost of the edge-assisted live streaming system. Compared with other methods, this system and method have significant advantages in improving the transmission capacity of current live streaming systems and saving bandwidth costs. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments recorded in this invention, and those skilled in the art can obtain other drawings based on these drawings.

[0053] Figure 1 This is a framework diagram of a cost-sensitive edge-assisted live streaming system provided in an embodiment of the present invention. Detailed Implementation

[0054] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0055] This invention designs an edge-assisted live streaming system that minimizes bandwidth costs under the 95% peak pricing model, i.e., a cost-sensitive edge-assisted live streaming system, referencing... Figure 1 As shown, the algorithm includes a Shapley value estimation module and a request scheduling module. For a given time slot t, the algorithm takes historical bandwidth usage as input and a scheduling decision as output. This invention proposes a series of methods to implement this system, a cost-sensitive edge-assisted live streaming method, including the following steps:

[0056] (1) Estimating the Shapley value

[0057] This invention models bandwidth costs and uses them as input to the request scheduling algorithm during the Shapley value estimation process. The Monte Carlo method is used to calculate the Shapley value of bandwidth usage for CDN and edge servers in each time slot t. This method can be extended to large-scale time slots.

[0058] The Shapley value is an intuitive concept from cooperative game theory that characterizes the fair cost-sharing among participants; it represents the Shapley value of participant i. The specific definitions are as follows:

[0059]

[0060] Where S N Let S(π,i) represent all possible permutations of set N, let S(π,i) represent the set of participants who arrive no later than i in order π, let v(S(π,i)) represent the cost contribution of a participant, and let v(S(π,i)\{i}) represent the cost contribution of participants other than i.

[0061] The time complexity of directly calculating the Shapley value using the above formula is O(N!), where N is the number of time slots. Once N reaches a very large value, such as 100, direct calculation using this formula becomes impractical. For billing cycles with hundreds (1-hour intervals) or thousands (5-minute intervals) of time slots, such precise calculation is impractical. This invention utilizes a simple stochastic method to approximate the Shapley value, which can be scaled to thousands of time slots.

[0062] To avoid an excessively large N!, we only consider random permutations of k. use This is represented by an average of the incremental edge costs:

[0063]

[0064] The parameter k determines the error; the larger k is, the smaller the error.

[0065] (2) Request scheduling

[0066] For demand scheduling, this invention proposes a request scheduling greedy algorithm, which makes online request scheduling decisions in each time slot t to minimize bandwidth costs. The algorithm takes the estimated number of viewing requests as input and outputs a binary vector representing the scheduling decision for each channel in each edge region.

[0067] The pseudocode for the greedy algorithm (Algorithm 1) is as follows:

[0068]

[0069] Where 's' represents the live streaming channel, 'c' represents the CDN server, 'e' represents the edge server, and 't' represents the time slot. This indicates that time slot t requests s within the area covered by e, f c This indicates the price of 95% peak bandwidth for a CDN node, b s C represents the average bitrate of channel s. e This indicates the maximum bandwidth of edge server e. The scheduling decision, i.e., whether channel s in time slot t is distributed by the edge server, is a binary vector. 1 indicates that channel s in time slot t is distributed by the edge server e, and 0 indicates that channel s in time slot t is not distributed by the edge server e. Indicates in y t Lieutenant General 1 is assigned to The result is C(y) t ) indicates the use of scheduling decision y t Target cost, ΔC (e′,s′) (y t ) refers to Cost reduction after setting it to 1:

[0070]

[0071] The algorithm described above greedily reduces the total bandwidth cost by delegating requests to edge servers as much as possible. First, y t Initialized as a zero vector, before each iteration of the task, a combination P = {(e′, s′)} of candidate edge servers and channels is constructed, and then... We assign a value of 1, meaning that in time slot t, channel s′ is distributed by edge server e′, and we guarantee that (e′, s′) ∈ P will not cause e to exceed the bandwidth limit. Then, we select the combination (e′, s′) that most significantly reduces bandwidth costs. Assign a value of 1. If the selected (e′, s′) does not lead to a cost reduction, the iteration process ends. Finally, y... t Return it as a result.

[0072] The greedy algorithm has a maximum of ES iterations. In each iteration: (1) the time complexity of finding candidate pairings is O(ES); (2) the time complexity of calculating the Shapley value of each candidate combination is O(ESk); (3) the time complexity of finding the pairing with the greatest cost reduction is O(ES); (4) the time complexity of variable assignment is O(1). Therefore, the total time complexity of the algorithm is O(k(ES)). 2 Of these, Shapley value estimation is the most time-consuming operation.

[0073] As can be seen from the above time complexity analysis, Shapley value estimation accounts for most of the computational cost of the algorithm. In order to reduce the computational cost, this invention uses three techniques to accelerate the Monte Carlo method: (1) 95th percentile estimation to eliminate half of the sorting operations, (2) using interpolation methods to reuse Shapley value estimation results, and (3) using vectorization operations to complete all repetitive operations at once.

[0074] (1) 95 Peak Estimation. As mentioned above, the Monte Carlo method samples by arranging k subsets of bandwidth usage and calculates the marginal cost after summing the bandwidth usage to be estimated, as calculated above. The 95-peak value is defined as such. To calculate the 95-peak value of the bandwidth, the conventional approach requires sorting the sampled subsets S(π,i) and S(π,i) and obtaining their respective 95-peak values. The optimization opportunity of this invention is that only the subset S(π,i) needs to be sorted; the 95-peak values ​​of S(π,i) and S(π,i) can be derived conditionally.

[0075] Specifically, we sort S(π,i)\{i} to obtain its i-th and the Let the larger numbers be denoted as a and b, respectively. Assuming the i-th largest number in π is x, then the 95% peak of these two sequences is:

[0076] Q(S(π,i)\{i},0.95)=a

[0077]

[0078] (2) Interpolation. For CDN, from 0 to... We sample 1000 values ​​uniformly within the range; for edge servers, we sample values ​​from 0 to C. e Sample 1000 values ​​within the (edge ​​server's bandwidth limit) range. For convenience, for CDN or edge servers, temporarily set u... i and p i Let i = 1, ..., 1000 represent the sample bandwidth usage and Shapley value, respectively; u' represents the bandwidth usage to be estimated. We estimate the Shapley value of u' as follows:

[0079]

[0080] (3) Vectorization. Each edge server needs to perform a large number of repetitive operations, such as permutation operations in Monte Carlo sampling and interpolation operations in Shapley value estimation. Therefore, this invention utilizes vectorization to complete these repetitive operations in one step: (i) In Monte Carlo sampling, we will Connection as And use a vectorized permutation operation to generate all Monte Carlo samples at once. (ii) Similarly, in the Shapley value estimation, the pre-computed Shapley values ​​are concatenated and interpolated for all edge servers and all channels at once.

[0081] The pseudocode for the request scheduling greedy algorithm (Algorithm 2) optimized using acceleration techniques is as follows:

[0082]

[0083]

[0084] This invention utilizes an optimized request scheduling greedy algorithm based on acceleration technology, which can reduce the time complexity from O(k(ES)). 2 The value is reduced to O((ES)). 2 ), where ES is the upper limit of the number of iterations. The reason for the reduction in complexity is that the original Shapley value estimation with a time complexity of O(kES) in each iteration is transformed into vectorized interpolation calculation, which has a time complexity of approximately O(1).

[0085] This invention conducted a large-scale experiment using data from an operational live streaming system over two months, involving nearly one million users and billions of viewing requests across 20 cities. The experimental results show that this invention has better timeliness, saving 17.7%, 15.51%, and 14.81% of bandwidth costs respectively compared to random methods (edge ​​servers randomly handle half of the requests, and CDN handles the remaining requests), dedicated CDN methods, and the most advanced methods currently available that do not use Shapley values.

[0086] In summary, the method and system provided by this invention have the following innovations: 1) This invention is the first edge-assisted live streaming system to use the 95 Peak pricing model; 2) This invention proposes a time-sensitive Shapley value estimation method and utilizes various acceleration technologies to reduce computational overhead without sacrificing performance; 3) This invention uses a request scheduling greedy algorithm to solve the online request scheduling problem.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A cost-sensitive edge-assisted live streaming method, characterized in that, The Shapley value is used to model the contribution of bandwidth usage in each time slot to the overall bandwidth cost, and a greedy algorithm is used to solve the scheduling request based on the estimated Shapley value, and the scheduling decision is output. Specifically, the Monte Carlo method is used to calculate the content delivery network and edge servers in each time slot. The Shapley value used for bandwidth in the participants Shapley value The specific definitions are as follows: ， in Represents a set All possible permutations Indicates no later than by The set of participants who arrive in sequence. Cost contribution representing participants Representative except Cost contributions from external participants; By arranging the broadband usage order of Sampling is performed on a subset of data, using Then, the marginal cost increment is averaged, and expressed as: ， in, Representative before Each time slot Bandwidth usage, Represents edge servers. Represents a time slot; The Monte Carlo method is accelerated using the following three methods: S11. Use 95th percentile estimation to eliminate half of the sorting operations; S12. Use interpolation methods to reuse the Shapley value estimation results; S13. Use vectorized operations to complete all repetitive operations at once; In S11, the specific process of using 95th percentile estimation to eliminate half of the sorting operations includes: For the sampled subset Sort, to get its first and the Large numbers are denoted as follows: and ; Assuming in The Middle Large numbers are ,but and The 95% peak values ​​of the two sequences are: ， ， in, Representative from Remove from set The set after, Representative sequence In China Previously and in the set Participants in; In S12, the specific process of using interpolation methods to reuse the Shapley value estimation results includes: For content delivery networks, from arrive 1000 values ​​were sampled uniformly within the range; For edge servers, the bandwidth limit is from 0 to the edge server's maximum bandwidth. Sample 1000 values ​​within the range; Bandwidth usage to be estimated The Shapley value is: ， in, and These represent the sampling of bandwidth usage and the corresponding Shapley value, respectively. In S13, the specific process of using vectorization to complete all repetitive operations at once includes: (i) In sampling using the Monte Carlo method, Connection as And all Monte Carlo samples are generated at once using vectorized permutation operations; (ii) In the Shapley value estimation, the pre-calculated Shapley values ​​are concatenated and interpolated for all edge servers and all channels at once; The greedy algorithm's process for solving scheduling requests includes: S21, Scheduling Decision Initialized as a zero vector, the Shapley value is estimated using vectorization operations and the 95-peak estimation interpolation sampling method. Before each iteration, a combination of candidate edge servers and channels is constructed. , Represents edge servers. Represents the live streaming channel. These represent candidate edge nodes and candidate live streaming channels, respectively. S22. Calculate all Shapley values ​​using vectorized interpolation. The value is assigned to 1, that is, in Live Channels for Time Slot Candidates By candidate edge servers Distribute and ensure It will not cause Exceeding bandwidth limit; S23. Select the combination that most significantly reduces bandwidth costs. ,Will Assigned value If the selected combination If it does not lead to cost reduction, the iteration process ends, and ultimately... Return it as a result.

2. The cost-sensitive edge-assisted live streaming method according to claim 1, characterized in that, In S23, Assigned value The cost reduction is expressed as follows: ， in, Refers to scheduling decisions, that is, in Time slot channel Whether it is distributed by edge servers It is a binary vector. Indicates in Time slot channel By edge server distribution, Indicates in Time slot channel Not by edge server distribution, Indicates in Lieutenant General Assign to As a result, Indicates the use of scheduling decisions The target cost These represent candidate edge nodes and candidate live streaming channels, respectively.

3. A cost-sensitive edge-assisted live streaming system, characterized in that, The following modules are included to implement the cost-sensitive edge-assisted live streaming method according to any one of claims 1-2: The Shapley value estimation module uses the Shapley value to model the contribution of bandwidth usage for each time slot to the overall bandwidth cost; The scheduling module requests a scheduling request and, based on the estimated Shapley value, uses a greedy algorithm to solve the scheduling request and outputs a scheduling decision. The Shapley value estimation module uses the Monte Carlo method to calculate the content delivery network and edge servers in each time slot. The bandwidth used in the calculation is the Shapley value, and the following three methods are employed to accelerate the Monte Carlo method: S11. Use 95th percentile estimation to eliminate half of the sorting operations; S12. Use interpolation methods to reuse the Shapley value estimation results; S13. Use vectorized operations to complete all repetitive operations at once.

Images