Adaptive content caching method

Abstract

The application discloses a self-adaptive content caching method, and the steps are as follows: 1) a request processor extracts characteristic parameters of the content access request information and writes the characteristic parameters into an original characteristic database of a local cache server; 2) content popularity in a next time window is predicted and written into a popularity database; and 3) a cache access interface of the local cache server checks whether the local cache server caches the requested content according to the characteristic parameters extracted by the request processor, if the requested content exists, the local cache server transmits the requested content to a user terminal, if the requested content does not exist, the requested content is acquired from an upstream server and transmitted to the user terminal. The application can fully mine and utilize the characteristics of content requests under real network environment and user access conditions, and dynamically adjust a caching strategy to improve caching efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content distribution network, in particular to an adaptive content caching method. BACKGROUND

[0002] The interconnection of traditional enterprises, the electronicization of business and transactions, the rapid growth of mobile Internet and video and game services have promoted the widespread deployment of content distribution services. The content distribution network (CDN) alleviates network congestion, improves service response speed and user experience by deploying cache servers closer to users at the network edge, and solves the dilemma of content "centralization and decentralization", which plays an important role in building a good Internet value chain. Among them, content caching attempts to learn the distribution of content requests in some way to ensure the efficient availability of content on cache servers. Content caching is usually driven by user requests. When the content requested by the user does not exist in the cache server, the cache server obtains the content from the source or other cache servers in real time. The existing relatively classic content caching methods include: Least Recently Used (LRU), Least Frequently Used (LFU), First In First Out (FIFO), Adaptive Replacement Cache (ARC), etc. The LRU algorithm stores the recently accessed content near the top of the cache, and when the cache is full, the least accessed content is removed from the bottom of the cache. The LFU algorithm uses a counter to record the frequency of content access, and when the cache is full, the content with the lowest access number will be removed first. The core principle of the FIFO algorithm is that if a content enters the cache first, it should be removed first. The ARC algorithm distinguishes between "recent" content that has only appeared once and "recent" content that has appeared at least twice, and adaptively divides the cache space between the two.

[0003] However, the existing content caching method only relies on time, frequency and other single content request characteristics, and lacks more sufficient feature mining and application. In addition, in the real network environment, content popularity, user access behavior and server service preference are constantly changing and even influencing each other, and the performance of the caching method based on the Independent Reference Model (IRM) in actual application still has a gap with the theoretical value. SUMMARY

[0004] The purpose of the present application is to solve the problems in the prior art.

[0005] The technical scheme adopted to achieve the object of the present application is as follows: an adaptive content caching method mainly comprises the following steps:

[0006] 1) The user end sends content access request information to the request processor.

[0007] 2) The request processor extracts the characteristic parameters of the content access request information and writes them into the original feature database of the local caching server.

[0008] The characteristic parameters of the content access request information include semantic features and context features.

[0009] 3) The learning processing module of the request processor periodically imports data from the original feature database and trains the deep neural network stored in the request processor. The request processor uses the trained deep neural network to predict the content popularity of the next time window and writes it into the popularity database.

[0010] Further, the main steps of the deep neural network to predict the content popularity of the next time window are as follows:

[0011] 3.1) Obtain the features x corresponding to the time window t t and the hidden state h t-1 corresponding to the previous time window t-1.

[0012] 3.2) Establish the prediction function of the deep neural network with L layers, i.e.:

[0013]

[0014] where f (l) (x) is the regression function of the lth layer of the deep neural network. α (l) > 0 is the weight of the regression function of the lth layer of the deep neural network. L is the total number of layers of the deep neural network.

[0015] where the regression function of the lth layer of the deep neural network is as follows:

[0016] f (l) (x) = Θ (l) h t (l) . (2)

[0017] where Θ (l) is the parameter learned by f (l) (x).

[0018] The reset gate r t , the update gate z t , the hidden state h t and the candidate hidden state of the deep neural network are as follows:

[0019]

[0020]

[0021]

[0022]

[0023] where {W, U, b} are GRU cell parameters. σ(·) is sigmoid activation function. ⊙ denotes element-wise multiplication. W denotes weights and b denotes biases. denote the connection matrices from the previous time step hidden layer to the reset gate r, update gate z and candidate hidden state respectively. t denotes the time step.

[0024] 3.3) Calculate error function i.e.

[0025]

[0026] where y t is the true content popularity of the next time window observed at time window t. x t is the content popularity of time window t.

[0027] 3.4) Update the deep neural network weights using online gradient descent algorithm based on minimizing the mean relative squared error and the GRU cell parameters of the deep neural network

[0028] GRU cell parameters are calculated using backpropagation algorithm.

[0029] The backpropagation algorithm is shown as follows:

[0030]

[0031] where η is the learning rate.

[0032] Deep neural network weights are updated as follows:

[0033]

[0034]

[0035]

[0036] where β ∈ (0, 1) is the discount factor, κ is the threshold parameter for smoothing noise data. ζ is the weight update parameter.

[0037] wherein the normalization factor Z at time window t+1 t+1 As shown below:

[0038]

[0039] 3.5) Utilizing the updated GRU cell parameters of the deep neural network and the deep neural network weights The deep neural network after the update predicts the content popularity of the next time window.

[0040] 4) The cache access interface of the local cache server checks whether the requested content is cached in the local cache server according to the feature parameters extracted by the request processor. If the requested content exists, the local cache server transmits the requested content to the user terminal. If the requested content does not exist, step 5 is entered.

[0041] 5) The local cache server obtains the requested content from the upstream server and transmits it to the user terminal. The decision interface of the local cache server removes the content with the lowest popularity that already exists in the local cache server according to the popularity database.

[0042] Further, the method for removing the content with the lowest popularity that already exists in the local cache server is to remove the top element of the content priority queue Q. The content priority queue Q is stored in the local cache server. The elements in the content priority queue Q are the requested contents arranged in ascending order of popularity.

[0043] It is worth noting that the present application maximizes the request hit rate of the local cache server by using a deep neural network to learn the dynamic evolution pattern of content popularity. At the same time, the present application can ensure the prediction ability of the model while reducing the consumption of computing resources as much as possible.

[0044] The technical effects of the present application are self-evident. The present application uses a deep neural network to represent the dynamic characteristics of dynamic requests and changing cache spaces, and integrates this representation into a deep reinforcement learning framework to help the cache server make content caching decisions online and adaptively. The present application can fully exploit and utilize the characteristics of content requests under real network environments and user access conditions, and dynamically adjust the caching strategy to improve cache efficiency. The present application can effectively learn the dynamic characteristics of content requests and cache spaces, and make content caching decisions accordingly to maximize cache efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0045] Fig. 1 Module and operation schematic diagram of the cache server;

[0046] Fig. 2An adaptive content caching model based on LSTM-DQN. DETAILED DESCRIPTION

[0047] The application will be further described below with reference to the embodiments, but should not be understood as limiting the above-mentioned subject matter of the application to the following embodiments. Various replacements and modifications can be made according to the ordinary technical knowledge and common practices in the art without departing from the above-mentioned technical idea of the application, and all of them should be included in the protection scope of the application.

[0048] Embodiment 1

[0049] Reference Figs. 1-2 , the adaptive content caching method mainly includes the following steps:

[0050] 1) The user terminal sends content access request information to the request processor.

[0051] 2) The request processor extracts the feature parameters of the content access request information and writes them into the original feature database of the local cache server.

[0052] The feature parameters of the content access request information include semantic features and context features.

[0053] 3) The learning processing module of the request processor periodically imports data from the original feature database and trains the deep neural network stored in the request processor. The request processor uses the trained deep neural network to predict the content popularity of the next time window and writes it into the popularity database.

[0054] The main steps of the deep neural network to predict the content popularity of the next time window are as follows:

[0055] 3.1) Obtain the features x t and the hidden state h t-1 corresponding to the time window t

[0056] 3.2) Establish the prediction function of the L-layer deep neural network, that is:

[0057]

[0058] In the formula, f (l) (x) is the regression function of the lth layer of the deep neural network. α (l) > 0 is the weight of the regression function of the lth layer of the deep neural network. L is the total number of layers of the deep neural network.

[0059] In which, the regression function of the lth layer of the deep neural network is as follows:

[0060] f (l) (x) = Θ (l) ht (l) (2)

[0061] where Θ (l) is the parameter of learning f (l) (x). h t (l) is the hidden state of the l-th layer deep neural network corresponding to time window t.

[0062] The reset gate r t , update gate z t , hidden state h t and candidate hidden state of the deep neural network are respectively as follows:

[0063]

[0064]

[0065]

[0066]

[0067] where {W, U, b} are the GRU unit parameters. σ(·) is the sigmoid activation function. ⊙ denotes element-wise multiplication. is the parameter of the reset gate, is the parameter of the update gate, is the parameter of the hidden state. The superscript l denotes the l-th layer. is the initial hidden state. W denotes the weight and b denotes the bias; denote the connection matrix of the hidden layer to the reset gate r, update gate z and candidate hidden state at the previous time, respectively; t denotes the time;

[0068] 3.3) Calculate the error function i.e.

[0069]

[0070] where y t is the true content popularity of the next time window observed by the time window t. x t is the content popularity of the time window t. denotes the error function of # and *.

[0071] 3.4) Update the deep neural network weights and the GRU unit parameters of the deep neural network based on minimizing the mean relative squared error using the online gradient descent algorithm

[0072] ​GRU unit parameters are calculated by using backpropagation algorithm.

[0073] The backpropagation algorithm is shown as follows:

[0074]

[0075] wherein, denotes gradient, and η is learning rate.

[0076] The reset gate is used to determine how much past information should be ignored. The update gate helps the model determine how much past information needs to be passed to the future. The candidate hidden state can facilitate the calculation of the subsequent hidden state.

[0077] Deep neural network weights The update is as follows:

[0078]

[0079]

[0080]

[0081] wherein, β∈(0, 1) is a discount factor, k is a threshold value parameter for smoothing noise data, and ζ is a weight update parameter related to the number of neural network layers;

[0082] wherein, the normalization factor Z of the deep neural network at the time window t+1 t+1 is shown as follows:

[0083]

[0084] 3.5) Using the updated GRU unit parameters of the deep neural network to predict the content popularity of the next time window. and the deep neural network weights

[0085] 4) The cache access interface of the local cache server checks whether the requested content is cached in the local cache server according to the feature parameters extracted by the request processor. If the requested content exists, the local cache server transmits the requested content to the user end. If the requested content does not exist, step 5 is entered.

[0086] 5) The local cache server obtains the requested content from the upstream server and transmits it to the user end. The decision interface of the local cache server removes the content with the lowest popularity that already exists in the local cache server according to the popularity database.

[0087] ​The method for removing the least popular content already existing in the local cache server is to remove the top element of the content priority queue Q. The content priority queue Q is stored in the local cache server. The elements in the content priority queue Q are the requested contents arranged in ascending order of popularity.

[0088] Embodiment 2

[0089] The adaptive content caching method mainly includes the following steps:

[0090] 1) The user end sends content access request information to the request processor.

[0091] 2) The request processor extracts the characteristic parameters of the content access request information and writes them into the original feature database of the local cache server.

[0092] 3) The learning processing module of the request processor, i.e., the learning processing program, periodically imports data from the original feature database and trains the deep neural network stored in the request processor. The request processor uses the trained deep neural network to predict the content popularity of the next time window and writes it into the popularity database.

[0093] 4) The cache access interface of the local cache server checks whether the requested content is cached in the local cache server according to the characteristic parameters extracted by the request processor. If the requested content exists, the local cache server transmits the requested content to the user end. If the requested content does not exist, go to step 5.

[0094] 5) The local cache server obtains the requested content from the upstream server and transmits it to the user end. The decision interface of the local cache server removes the least popular content already existing in the local cache server according to the popularity database.

[0095] Embodiment 3

[0096] The adaptive content caching method mainly includes the steps of Embodiment 2, wherein the main steps of the request processor using the trained deep neural network to predict the content popularity of the next time window are as follows:

[0097] 1) Obtain the features x corresponding to the time window t t and the hidden state h t-1 corresponding to the previous time window t-1.

[0098] 2) Establish the prediction function of the L-layer deep neural network, i.e.,

[0099]

[0100] where f (l) (x) is the regression function of the l-th layer deep neural network. a (l)> 0 is the weight of the l-th layer deep neural network regression function.

[0101] The regression function of the l-th layer deep neural network is as follows:

[0102] f (l) (x) = Θ (l) t (l) (2)

[0103] In the formula, Θ (l) is the parameter of learning f (l) (x).

[0104] The reset gate r t , the update gate z t , the hidden state h t and the candidate hidden state of the deep neural network are as follows:

[0105]

[0106]

[0107]

[0108]

[0109] In the formula, {W, U, b} is the GRU unit parameter. σ(·) is the sigmoid activation function. ⊙ represents multiplication by element.

[0110] 3) Calculate the error function That is:

[0111]

[0112] In the formula, y t is the true content popularity of the next time window observed by the time window t. x t is the content popularity of the time window t.

[0113] 4) Based on minimizing the mean relative square error, the deep neural network weights and the GRU unit parameters of the deep neural network

[0114] The GRU unit parameters are calculated by using the back propagation algorithm.

[0115] The back propagation algorithm is as follows:

[0116] ​

[0117] where η is the learning rate.

[0118] Deep neural network weights is updated as follows:

[0119]

[0120]

[0121]

[0122] where β∈(0, 1) is the discount factor and κ is the threshold parameter for smoothing the noisy data.

[0123] where the normalization factor Zt+1at time window t+1 t+1 as follows:

[0124]

[0125] 5) Update the GRU cell parameters of the deep neural network and the deep neural network weights the deep neural network after the update to predict the content popularity of the next time window.

[0126] Embodiment 4:

[0127] The system based on the adaptive content caching method mainly comprises a user end, a request processor, an original feature database, a local cache server and an upstream server.

[0128] The content access request information is sent to the request processor;

[0129] The request processor extracts the feature parameters of the content access request information and writes them into the original feature database of the local cache server through a feature updater; the request processor has a trigger for mobilizing the cache access interface.

[0130] The learning processing module of the request processor regularly imports data from the original feature database and trains the deep neural network stored in the request processor; the request processor uses the trained deep neural network to predict the content popularity of the next time window and writes it into the popularity database;

[0131] The local cache server has a cache access interface and a decision interface. The local cache server has a database in which a plurality of requested contents are stored.

[0132] The cache access interface of the local cache server checks whether the request content is cached in the local cache server according to the feature parameters extracted by the request processor, if the request content exists, the local cache server transmits the request content to the user terminal, if the request content does not exist, the request content is obtained from the upstream server and transmitted to the user terminal, and the decision interface of the local cache server removes the content with the lowest popularity existing in the local cache server according to the popularity database.

Claims

1. A method of adaptive content caching, characterized by, The method comprises the following steps: 1) the user terminal sends content access request information to the request processor; 2) the request processor extracts characteristic parameters of the content access request information and writes them into an original characteristic database of the local cache server; 3) a learning processing module of the request processor periodically imports data from the original characteristic database and trains a deep neural network stored in the request processor; the request processor uses the trained deep neural network to predict content popularity in a next time window and writes the prediction into a popularity database; 4) a cache access interface of the local cache server checks whether the local cache server caches the requested content according to the characteristic parameters extracted by the request processor; if the requested content exists, the local cache server transmits the requested content to the user terminal; if the requested content does not exist, step 5 is performed; 5) the local cache server obtains the requested content from an upstream server and transmits the content to the user terminal; a decision interface of the local cache server removes the content with the lowest popularity that already exists in the local cache server according to the popularity database; The deep neural network predicts content popularity in a next time window in the following steps: 3.1) Obtain the feature x corresponding to the time window t t and the hidden state h corresponding to the previous time window t-1 t-1 ; 3.2) a prediction function of the deep neural network of the Lth layer is established, i.e. wherein f (l) (x) is a regression function of the lth layer of the deep neural network; a (l) > 0 is a weight of the regression function of the lth layer of the deep neural network; and L is the total number of layers of the deep neural network. wherein a regression function of the deep neural network of the lth layer is as follows: f (l) (x) = Θ (l) h t (l) (2) where Θ (l) is the parameter of learning f (l) (x); h t (l) is the hidden state of the l-th layer deep neural network corresponding to the time window t. reset gate r of a deep neural network t update gate z t hidden state h t and candidate hidden state are as follows, respectively: where {W, U, b} are GRU unit parameters; σ(·) is a sigmoid activation function; and denotes element-wise multiplication. is a parameter of the reset gate, is a parameter of the update gate, is a parameter of the hidden state; the superscript l denotes the l-th layer; is an initial hidden state; 3.3) Calculating the error function That is: In the formula, y t is the content true popularity of the next time window observed by the time window t; x t is the content popularity of the time window t; denotes the error function of # and * 3.4) Update the GRU cell parameters {W, U, b} and the deep neural network weights using an online gradient descent algorithm based on minimizing the mean relative squared error GRU unit parameters {W, U, b} are updated as follows: are computed via backpropagation from wherein is the gradient computed by backpropagation through formula (8); η is the learning rate; the reset gate is used to determine how much past information should be ignored; the update gate helps the model determine how much past information needs to be passed to the future; the candidate hidden state facilitates the computation of the subsequent hidden state; Deep neural network weights Updated as follows: wherein β ∈ (0, 1) is a discount factor and κ is a threshold value parameter for smoothing noise data; wherein the normalization factor Z of the deep neural network at time window t+1 t+1 As follows: 3.5) Utilize the updated GRU cell parameters {W, U, b} of the deep neural network and the deep neural network weights to predict the content popularity of the next time window.

2. The adaptive content caching method of claim 1, wherein, The characteristic parameters of the content access request information include semantic features and context features.

3. The adaptive content caching method of claim 1 or 2, characterized in that, The method for removing the content with the lowest popularity that already exists in the local cache server is removing a top element of a content priority queue Q; the content priority queue Q is stored in the local cache server; Elements in the content priority queue Q are arranged in ascending order of popularity.

Images