JP6944155B2 - Orchestrator equipment, programs, information processing systems, and control methods - Google Patents

Description

The present invention relates to an orchestrator device, a program, an information processing system, and a control method.

Conventionally, there is known a technique of developing an application for providing a network service by dividing it into a plurality of components (microservices) for each function and linking them on a network (see, for example, Non-Patent Document 1). The function that realizes each microservice is called a "virtual network function (VNF)", and the entire service composed of a plurality of VNFs in cooperation is called a "service chain". For example, an online shopping service chain may be configured to include a plurality of VNFs such as a customer management function, an order management function, an inventory management function, and a shipping management function. The VNF is located on a physical platform (eg, hardware such as a physical machine). This technology is expected to continue to be a promising technology because of the efficiency of design and testing in system construction, the ease of maintenance and cost reduction in operation, and so on.

Lianping Chen, “Microservices: Architecting for Continuous Delivery and DevOps”, in IEEE International Conference on Software Architecture, March 2018, Seattle, USA: IEEE

With the development of networks, lower cost of hardware, higher performance, etc., the degree of freedom in arranging VNFs that make up the service chain will increase. In addition, various VNFs will be arranged on the physical platform according to their respective characteristics. However, with multiple VNFs on a single physical platform, when a particular service chain's utilization or network bandwidth peaks, that service chain is affected by other co-existing services and services. There is a problem that the quality deteriorates. If the service provider and the physical platform provider are different, it is difficult to avoid the problem.

An object of the present invention made in view of such circumstances is to reduce the possibility that the future service quality of the service chain will deteriorate.

The orchestrator device according to the embodiment of the present invention is
An orchestrator device that manages a service chain composed of multiple virtual network functions (VNFs) arranged on multiple physical platforms in cooperation with each other.
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Department and
Based on the estimated value of the future service quality of the service chain output from the neural network, it is predicted whether or not the future service quality of the service chain will deteriorate.
If it is predicted that the future service quality of the service chain will deteriorate, the estimated value of the future service quality of the service chain corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. One of the plurality of arrangement patterns is determined based on the above.
A control unit for rearranging the plurality of VNFs on the plurality of physical platforms so as to match the determined arrangement pattern is provided.

The program according to the embodiment of the present invention
Information processing device,
It is an orchestrator device that manages a service chain composed of a plurality of virtual network functions (VNFs) arranged on a plurality of physical platforms in cooperation with each other, and is a performance measurement value of each said physical platform and the plurality of physical platforms. A storage unit that stores one or more neural networks that input the position information of each VNF in the above and output an estimated value of the service quality of the service chain, and the future of the service chain that is output from the neural network. Based on the estimated value of the service quality of the above, it is predicted whether or not the future service quality of the service chain will deteriorate, and if it is predicted that the future service quality of the service chain will deteriorate, it is acquired by using the neural network. One of the plurality of placement patterns is determined and determined based on the estimated value of the future service quality of the service chain corresponding to each of the plurality of placement patterns of the plurality of VNFs. It functions as an orchestrator device including a control unit for rearranging the plurality of VNFs on the plurality of physical platforms so as to match the arrangement pattern.

The information processing system according to the embodiment of the present invention is
An information processing system including a plurality of physical platforms and an orchestrator device for managing a service chain configured by linking a plurality of virtual network functions (VNFs) arranged on the plurality of physical platforms.
The orchestrator device is
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Department and
Based on the estimated value of the future service quality of the service chain output from the neural network, it is predicted whether or not the future service quality of the service chain will deteriorate.
If it is predicted that the future service quality of the service chain will deteriorate, the estimated value of the future service quality of the service chain corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. One of the plurality of arrangement patterns is determined based on the above.
It has a control unit that rearranges the plurality of VNFs on the plurality of physical platforms so as to match the determined arrangement pattern.

The control method according to the embodiment of the present invention is
It is a control method of an orchestrator device that manages a service chain composed of multiple virtual network functions (VNFs) arranged on multiple physical platforms in cooperation with each other.
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Steps and
A prediction step for predicting whether or not the future quality of service will deteriorate based on the estimated value of the future quality of service of the service chain output from the neural network.
If it is predicted that the quality of service will deteriorate in the future, the plurality of arrangements will be based on the estimated value of the quality of service corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. A step of determining one of the above-mentioned arrangement patterns, and
Includes a step of rearranging the plurality of VNFs on the plurality of physical platforms to match the determined placement pattern.

According to the orchestrator device, the program, the information processing system, and the control method according to the embodiment of the present invention, it is possible to reduce the possibility that the future service quality of the service chain will deteriorate.

It is a figure which shows the schematic structure of the information processing system which concerns on one Embodiment of this invention. It is a figure which shows the example of the neural network. It is a flowchart which shows the 1st operation of an orchestrator apparatus. It is a flowchart which shows the 2nd operation of an orchestrator apparatus. It is a flowchart which shows the 3rd operation of an orchestrator apparatus.

Hereinafter, embodiments of the present invention will be described. As a general rule, the invention according to the present embodiment improves reliability in a microservice environment in which a plurality of VNFs constituting a service chain are distributed and arranged on a network. Specifically, in the present embodiment, the performance of the VNFs distributed on the network is measured, the result is learned by the neural network, the VNF is moved, the performance is measured again, and the change of the VNF arrangement pattern is performed. We aim to automatically perform optimal VNF placement by letting the neural network continuously learn the changes in performance due to the above.

For example, assuming the presence of an amount Q representing the service quality of service chain, Q is assumed to be determined by the function Q (h _i) by all the variables h _i which affects the quality of service. In reality, can not enumerate all such variables h _i, probe associated with each VNF constituting the service chain (locator) collects information h _i as possible, it to orchestrator Send it and use it as an argument of the pseudo function Q'of Q. If if quality Q is has a primary linearity, and h _i is assumed to be position independent, but Q'by multiple regression analysis is required, since VNF is movable between physical platforms, Q'is Not simply asked. Therefore, in the present embodiment, an estimated value of service quality of the service chain is obtained by using a neural network.

(Configuration of information processing system)
An outline of the information processing system 1 according to the embodiment of the present invention will be described with reference to FIG. The information processing system 1 includes a plurality of physical platforms b (b _{1 to} b _n , where n is an integer of 2 or more) and an orchestrator device 10 as hardware components. Physical The plurality of physical platforms _b 1 ~b _n, a plurality of VNFa for configuring the service chain _(a 1 ~a _m. However, m is an integer of 2 or more.) Is disposed. One or more VNFa may be arranged on one physical platform b. The dotted frame in FIG. 1 indicates that VNFa is not arranged. Two or more physical platforms b may be located in the same geographical area or may be located in different geographical areas.

The physical platform b may be, for example, one information processing device, or may be a plurality of information processing devices installed in the same geographical area and capable of communicating with each other. The physical platform b includes, for example, a communication interface, a memory, and a processor. The physical platform b is connected to a network 30 including, for example, the Internet, etc. via a communication interface. The physical platform b stores arbitrary software resources used to realize the deployed VNFa in memory. The physical platform b realizes VNFa by executing information processing using the software resource stored in the memory in the processor.

Further, the physical platform b includes a probe (exploration device) 20 which is a very small instance for measuring the performance of the physical platform b. The performance of the physical platform b is, for example, information indicating an observable background load of the physical platform b, but is not limited to this and may be arbitrarily determined. The smaller the background load, the higher the performance of the physical platform b. The probe 20 is composed of hardware, software, or a combination thereof.

The orchestrator device 10 is an information processing device such as a computer, and is connected to the network 30. Orchestrator apparatus 10 includes a plurality of VNFa ₁ ~a _m manages the service chain constructed by linking. Specifically, the orchestrator 10 arranges the plurality of VNFa ₁ ~a _m multiple physical platforms _b 1 ~b _n, by linking them to form a service chain. In the present embodiment, "arranging the VNFa on the physical platform b" means that the execution subject of the VNFa is determined to be the physical platform b, and the software resources for realizing the VNFa are stored in the physical platform b. May include letting.

As a general rule, in the information processing system 1 according to the present embodiment, each physical platform b transmits a measured value of the performance of the physical platform b measured by the probe 20 to the orchestrator device 10. The orchestrator device 10 acquires an estimated value of future service quality of the service chain by using the measured value of the performance of each physical platform b and the like. The service quality of a service chain is information indicating the observable quality of the service chain, for example, information obtained by quantifying the quality of experience (QoE) of end users who use the service chain by an arbitrary method. However, the present invention is not limited to this, and may be arbitrarily determined. The orchestrator device 10 predicts whether or not the future quality of service will deteriorate based on the estimated value of the future quality of service. For example, if the estimated value of future service quality satisfies a predetermined deterioration prediction condition, it is predicted that future service quality will deteriorate. If future service quality is predicted to decrease, the information processing system 1, the orchestrator 10, to determine one of the arrangement patterns from among a plurality of arrangement patterns of a plurality of VNFa ₁ ~a _m, in the arrangement pattern Consistently relocating VNFa ₁ ~a _m.

Thus, according to this embodiment, when the future quality of service of the service chain is predicted to be lowered, since a plurality of VNFa ₁ ~a _m is relocated, future service quality of the service chain is reduced The probability can be reduced.

Next, the configuration of the orchestrator device 10 will be described in detail. The orchestrator device 10 includes a communication unit 11, a storage unit 12, and a control unit 13 as hardware components.

The communication unit 11 includes a communication module connected to the network 30. For example, the communication unit 11 may include a communication module corresponding to a wired LAN (Local Area Network) standard. In the present embodiment, the orchestrator device 10 is connected to the network 30 via the communication unit 11.

The storage unit 12 includes one or more memories. In the present embodiment, the "memory" is, for example, a semiconductor memory, a magnetic memory, an optical memory, or the like, but is not limited thereto. Each memory included in the storage unit 12 may function as, for example, a main storage device, an auxiliary storage device, or a cache memory. The storage unit 12 stores arbitrary information used for the operation of the orchestrator device 10. For example, the storage unit 12 may store a system program, an application program, and the like.

The control unit 13 includes one or more processors. In the present embodiment, the "processor" is a general-purpose processor or a dedicated processor specialized for a specific process, but is not limited thereto. The control unit 13 controls the operation of the entire orchestra device 10. Hereinafter, the first to third operations will be described as examples of the main operations of the orchestra device 10 controlled by the control unit 13, but the feasible operations of the orchestrator device 10 are not limited thereto.

(First operation)
The first operation of the orchestrator device 10 will be described. As a general rule, the first operation is an operation of generating and learning a neural network or a similar supervised learning machine F, which will be described later. Control unit 13, a plurality of physical platforms _b 1 ~b _n measurements of performance _d 1 to d _n and a plurality of physical platforms _b 1 of the plurality of ~b _n VNFa ₁ ~a position information of the _m L _(a 1) ~L _{(a m)} as input, multiple neural networks to output an estimate q quality of service of the service chain _{F (F} 1 _{~F j.} However, j is an integer of 2 or more.) to generate, store Store in part 12. Hereinafter, the performance measurements _d 1 to d _n and the position information _L (a 1) ~L a vector comprising a vector component _{(a m),} also referred to as input data X. Further, the estimated value q of the service quality is also referred to as output data y. Each neural network F behaves as a function y = F (X) that inputs input data X that is a vector and outputs output data y that is a scalar.

FIG. 2 shows a configuration example of the neural network F. Typically, the neural network F has one input layer (l = 1), one or more arbitrary number of intermediate layers (l = 2, ..., L-1), and one output layer (l = 1). It is composed of an L layer including L) (where L is an integer of 3 or more). The input layer contains the same number of neurons as the number of components of the input data X of the neural network F (the number of dimensions of the vector X). In the present embodiment, the number of components of the input data X is n + n × m as described later, and the number of neurons included in the input layer shown in FIG. 2 is also n + n × m. The middle layer contains any number of neurons of two or more. The number of intermediate layers may be arbitrarily determined. In the example shown in FIG. 2, the number of intermediate layers is 3 (l = 2, 3, 4), and the number of neurons contained in each intermediate layer is r. The number of neurons may be different for each middle layer. The output layer contains one neuron. However, the configuration of the neural network F is not limited to this example, and may be arbitrarily determined. In FIG. 2, x _i indicates the output of the i-th neuron in the input layer, z _i ^(j) indicates the output of the i-th neuron in the middle layer of l = j, and y indicates the output of the neuron in the output layer. Is shown.

The neural network F is created based on a parameter vector including hyperparameters such as the number of neurons in each layer and the acquisition time interval of learning data in online learning, but is not limited to this. You may. The parameter vector may be determined based on, for example, user input, or may be automatically determined by the control unit 13. Each neural network F may have the same number of intermediate layers and / or the number of neurons contained in each intermediate layer, or may be different from each other. Further, when creating the neural network F, the type of neuron (for example, linear, RSTM (Long Short-Term Memory), or GRU (Gated Recurrent Unit), etc.), and the type of function used as an activation function (for example, step). A type of non-linear function such as a function, a sigmoid function, or a softmax function) may be specifiable. Each neural network F may have the same network architecture (for example, forward propagation type, CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), etc.) or may be different from each other.

The control unit 13 sequentially acquires learning data including input data X (k-1) and teacher data q ₀ (k), and modifies each neural network F by online learning using the learning data. Here, k is a discrete time, the input data X (k-1) indicates the input data at the time k-1, and the teacher data q ₀ (k) indicates the teacher data at the time k.

Input data X, as described above, a plurality of physical platforms _b 1 ~b _n measured values of the performance (background _load) d 1 to d _n, a plurality of VNFa in a plurality of physical platforms _b 1 ~b _{n 1} ~a _m respective position information _L (a 1) ~L vector _X = _[d 1 includes the _{(a m),} as vector _{components, ..., d n, L (} a 1), ..., L (a m) ] ^T. Measurements _d 1 to d _n of performance is measured by a plurality of physical platforms _b 1 ~b _n probe 20, it is received from a plurality of physical platforms _b 1 ~b _n through the communication unit 11. Meanwhile, VNFa ₁ ~a position information _L (a 1) of the _m ~L _{(a m)} is determined as follows. The probe 20 of each physical platform b detects a VNF located on the physical platform b. Control unit 13, the via the communication unit 11 each physical platform b, the physical platform b in receiving information indicating each VNFa disposed, position information of each VNFa ₁ ~a _m based on the information L _(a 1) determining ~L _{(a m).} Here, the position information of VNFa L (a) is, for example, a component _g 1 to g _n a plurality of (n) respectively corresponding to a plurality of physical platforms _b 1 ~b _n, and the VNFa is arranged 1 component g corresponding to have physical platform b, vector with zero component g corresponding to the physical platform b in which the VNFa is not placed _{(i.e., L (a) = [g} 1, ..., g n]) Is. Accordingly, the input data _{X = [d 1, ...,} d n, L (a 1), ..., L (a m)] T frequency components (the number of dimensions of the vector X) is a n + n × m pieces. For example, n = 4, and four physical platform _b 1 ~b _4, respectively Sapporo, Tokyo, Osaka, when located in Fukuoka, if VNFa is in Sapporo L (a) = [1, It can be expressed as 0,0,0], and when it is located in Tokyo and Osaka, it can be expressed as L (a) = [0,1,1,0]. However, the position information L (a) of the VNFa is not limited to the above-mentioned format, and each physical platform b on which the VNFa is located may be indicated in any format.

The teacher data q ₀ is the actual value (scalar) of the service quality of the service chain. The probe 20 of each physical platform b detects arbitrary information used to calculate the actual value of service quality of the service chain. The control unit 13 calculates the actual value of service quality based on the information received from each physical platform b via the communication unit 11.

The control unit 13 inputs the input data X (k-1) to each neural network F, and acquires the estimated value q (k) of the service quality output from each neural network F. The control unit 13 modifies (learns) each neural network F so as to reduce the difference between the estimated service quality value q (k) and the actual value q _{0 (k).} The modification of each neural network F is carried out by, for example, an error backpropagation method, but the modification is not limited to this, and may be carried out by any method. The control unit 13 modifies each neural network F each time the learning data is sequentially acquired at a predetermined time interval. In this way, online learning of each neural network F is carried out. By inputting the current input data X (k) into the trained neural network F, for example, it is possible to estimate the future service quality of the service chain (that is, output the estimated value q (k + 1)). be.

(Second operation)
The second operation of the orchestrator device 10 will be described. Second operation as schematically is an operation of rearranging the VNFa ₁ ~a _m. Control unit 13 sets the accuracy information indicating the accuracy of the plurality of neural networks _{F 1 ~F j s (F 1} ) ~s the _{(F j)} to the initial value. The initial value may be determined by any method, but the accuracy information s (F) of each neural network F can be sequentially updated by the third operation described later. The accuracy information s (F) of the neural network F is information indicating the estimation accuracy of the future service quality of the service chain estimated using the neural network F. By using the neural network F with high estimation accuracy, the future service quality of the service chain can be estimated with high accuracy. Hereinafter, among the plurality of neural networks F _{1 to} F _j , the neural network F having the highest estimation accuracy (that is, the highest accuracy) indicated by the accuracy information s (F) is also referred _{to as F high, and the accuracy information s (F).} the lowest estimation accuracy as indicated by (i.e., the minimum accuracy) referred to the neural network F and F _low.

The control unit 13 acquires the current input data X (k). The control unit 13 acquires the estimated value q (k + 1) of the future service quality of the service chain by inputting the input data X (k) into the highest-precision neural network _High. The control unit 13 predicts whether or not the future service quality will deteriorate based on the acquired estimated value q (k + 1). For example, if the estimated value q (k + 1) satisfies a predetermined reduction prediction condition, it is predicted that the future service quality will deteriorate. In the present embodiment, the decrease prediction condition is, for example, a condition that the estimated value q (k + 1) is lower than _{the actual value q 0 (k) of the service quality at the time k of the service chain by a predetermined threshold value or more.} The threshold can be determined, for example, by experiment or simulation. However, the reduction prediction condition is not limited to the above example, and may be arbitrarily determined. For example, the decline prediction condition may be a condition that the estimated value q (k + 1) of the future service quality is less than a predetermined threshold value.

If future service quality is predicted to decrease, the control unit 13, so as to conform to a plurality of VNFa ₁ determines one arrangement pattern from a plurality of arrangement patterns of ~a _m, it determined the one arrangement patterns , the rearranging of VNFa ₁ ~a _m multiple physical platforms _b 1 ~b _n. Here, the "arrangement pattern" means a combination of physical platforms b of each of the plurality of VNFa ₁ ~a _m are arranged.

Here, a method for determining the arrangement pattern will be described. Control unit 13, using a neural network _{F high} up accuracy, to obtain an estimate q'the corresponding service quality to each of a plurality of arrangement patterns of a plurality of VNFa ₁ ~a _m. Specifically, the control unit 13, for example, the input at time k data _{X (k) = [d 1} , ..., d n, L (a 1), ..., L (a m)] of the ^T, 1 or more changing the position information of VNFa L (a) is output from the neural network F _high by input to the neural network F _high, the estimated value of the service quality in the case of changing the current arrangement pattern q'( Get k + 1). For example, the control unit 13, the estimated value of the service quality for all the arrangement patterns capable of multiple _{VNFa 1 ~a m q'(k +} 1) may be acquired, or for a portion of the arrangement pattern of the total arrangement pattern The estimated value of service quality q'(k + 1) may be obtained. For example, consider a case where the physical platform _{b 5} the destination VNFa _1. The position information after the movement of VNFa ₁ is represented _{as L (a 1} , b _5). The VNFa ₁ is moved to the physical platform _{b 5,} the estimate of the quality of service remain fixed without moving for the remaining _{VNFa 2 ~a m q'(k +} 1) is, _d 1 to d of the current (k = 1) with _{n, q'(k + 1)} = F high (d 1, ..., d n, L (a 1, b 5), L (a 2), ..., L (a m)) is.

The control unit 13 may determine one arrangement pattern having the highest estimated service quality q'(k + 1) from the plurality of arrangement patterns. The control unit 13, so as to meet the determined the arrangement pattern, rearranging of VNFa ₁ ~a _m. According to this method, since a plurality of VNFa ₁ ~a _m are rearranged so that the highest placement pattern probabilities improve future service quality of the service chain, future service quality of the service chain is reduced The probability can be reduced.

Alternatively, the control unit 13, based on an estimate of the quality of service corresponding to each of the plurality of arrangement patterns q'(k + 1), the probability distribution for the arrangement pattern of the plurality of VNFa ₁ ~a _m as a random variable, service It may be determined that the higher the estimated quality q'(k + 1) is the arrangement pattern, the higher the probability. Any method can be adopted to determine the probability distribution. For example, the probability pi of the _i- th arrangement pattern among the plurality of arrangement patterns is calculated by the following equation (1), _{where q ′ h} is the estimated value of the service quality corresponding to the h-th arrangement pattern among the plurality of arrangement patterns. ) Or equation (2).

The control unit 13 determines one arrangement pattern from the plurality of arrangement patterns by probability according to the determined probability distribution.

According to the above method using a probability distribution, with relatively high probability, since the plurality of VNFa ₁ ~a _m are rearranged such that the probable arrangement pattern to improve the future quality of service of the service chain, It is possible to reduce the probability that the future quality of service of the service chain will deteriorate. According to this approach, for example, a plurality of other VNFa' 1 ~ for realizing other services chain, _{a'2, if ...} are also arranged in a plurality of physical platforms b ₁ ~b _n, e.g. Both VNFa and VNFa'are relocated to the high-performance physical platform b, or the arrangement pattern after the relocation of the plurality of VNFa constituting the service chain and the relocation of the other VNFa' constituting the other service chain. The probability that the placement pattern will be the same after placement is reduced. That is, if VNFa and VNFa'are rearranged on the same physical platform b, or if the arrangement patterns after the rearrangement of both are the same, the background load of the physical platform b may increase and the performance may decrease. Poor performance of the physical platform b on which the VNF is located can lead to poor quality of service in the service chain. Therefore, according to the above method using a probability distribution, when a plurality of physical platforms b ₁ ~b _n is used to provide multiple services chain, future service quality of the service chain under the influence of another service chain It is possible to reduce the probability of decrease.

(Third operation)
The third operation of the orchestrator device 10 will be described. As a general rule, the third operation updates the accuracy information s (F ₁ ) to _{s (F j} ) of a plurality of neural networks F _{1 to} F _j , and changes the minimum accuracy neural network F _low to the highest accuracy neural network F. _This is an operation of replacing _{with a new neural network F new} having neighborhood parameters of the parameters constituting high. Control unit 13, after rearranging VNFa ₁ ~a _m by the second operation described above, it acquires the input data X (k-1) and service quality of the actual value _q 0 (k).

The control unit 13 acquires service quality estimates q ₁ (k) to q _j (k) _{output from a plurality of neural networks F 1 to} F _j by inputting input data X (k-1). do.

The control unit 13 is based on a comparison between each of the estimated service quality values q ₁ (k) to q _j (k) and the actual service quality value q ₀ (k), and the control unit 13 has a plurality of neural networks F ₁ to 1. F _j of accuracy information _s (F 1) to update the ~s _{(F j).} Specifically, the control unit 13 _{determines the absolute value or the square value of the difference between the actual value of service quality q 0} (k) and the estimated value q (k) of service quality acquired by using the neural network F. It is determined that the larger the value is, the lower the estimation accuracy of the neural network F is, and the accuracy information s (k) of the neural network F is updated. However, the update of the accuracy information s (F) based on the comparison between the actual value q ₀ (k) and the estimated value q (k) of the service quality is not limited to the above example, and may be performed by any method. For example, along with the difference between the actual value q ₀ (k) of the service quality and the estimated value q (k), from the input of the input data X (k-1) to the neural network F to the output of the output data y. The accuracy information s (k) of the neural network F may be updated in consideration of time. For example, the control unit 13 may determine that the longer the time, the lower the estimation accuracy of the neural network F, and may update the accuracy information s (k) of the neural network F.

The control unit 13 _{generates a new} neural network F new and stores it in the storage unit 12. The generation of the neural network F _new may be carried out by any method. For example, the control unit 13 generates a _new neural network F new by, for example, simulated annealing, based on the parameter vector (hyperparameter) of the most accurate neural network F _{high among the plurality of neural networks F 1 to} F _j. .. Specifically, the control unit 13, the parameter vector of the highest accuracy of the neural network F _high, the parameter vector of the next highest estimation accuracy neural network F, moved from the intermediate point in each direction only small value randomly A new parameter vector is determined, and a new neural network F _new is generated using the new parameter vector.

Control unit 13, a minimum accuracy of the neural network _{F low} among a plurality of neural networks _F 1 to F _j, replaced with a new neural network _{F new new.} Here, the control unit 13 may discard (that is, erase from the storage unit 12) _{the neural network Flow with the lowest accuracy.}

After replacing the lowest precision neural network _Flow with a new neural network F _new by the above-mentioned third operation, when the above-mentioned second operation and the third operation are executed again, the control unit 13 receives the neural network F _new . Until the learning progresses to a certain extent, the second operation and the third operation are performed _{without using the neural network F new} (that is, assuming that the neural network F _new is not included in "multiple neural networks F _{1 to} F _j"). You may do it. Further, the past learning data may be stored in the storage unit 12, and the new neural network F _new may be batch-learned using the past learning data.

(Flow of the first operation of the orchestrator device)
The flow of the first operation described above of the orchestra device 10 will be described with reference to FIG.

Step S100: The control unit 13 includes a plurality of physical platforms _b 1 ~b _n measurements of performance _d 1 to d _n and a plurality of physical platforms _b 1 of the plurality of ~b _n VNFa ₁ ~a position information of the _m L ( a ₁₎ as input ~L _{(a m),} and generates a plurality of neural networks _F 1 to F _j to output the estimated value q of the service quality of the service chain in the storage unit 12.

Step S101: The control unit 13 acquires learning data including input data X (k-1) and teacher data q _{0 (k).} Wherein the input data X (k-1) is a vector _{X (k-1) =} a _{[d 1, ..., d n} , L (a 1), ..., L (a m)] T, the teacher data q ₀ (k) is the actual value (scalar) of the service quality of the service chain.

Step S102: The control unit 13 modifies each neural network F by online learning using the training data.

Step S103: The control unit 13 increments the discrete time k. The process then returns to step S101.

(Flow of the second operation of the orchestrator device)
The flow of the above-described second operation of the orchestra device 10 will be described with reference to FIG.

Step S200: The control unit 13 sets the accuracy information indicating the accuracy of the plurality of neural networks _{F 1 ~F j s (F 1} ) ~s the _{(F j)} to the initial value.

Step S201: The control unit 13 acquires the current input data X (k).

Step S202: The control unit 13 inputs the input data X (k) to the highest-precision neural network F _high, and the estimated value q (k + 1) of the future service quality of the service chain is output from the neural network F _high. ) To get.

Step S203: The control unit 13 predicts whether or not the future service quality will deteriorate based on the acquired estimated value q (k + 1). If it is predicted that the quality of service will deteriorate in the future (step S203-Yes), the process proceeds to step S204. On the other hand, if it is predicted that the quality of service will not deteriorate in the future (step S203-No), the process proceeds to step S206.

Step S204: The control unit 13 determines one of the arrangement patterns from among a plurality of arrangement patterns of a plurality of VNFa ₁ ~a _m.

Step S205: The control unit 13, so as to match the arrangement pattern determined in step S204, rearranging of VNFa ₁ ~a _m multiple physical platforms _b 1 ~b _n.

Step S206: The control unit 13 increments the discrete time k. The process then returns to step S201.

(Flow of the third operation of the orchestrator device)
The flow of the above-described third operation of the orchestra device 10 will be described with reference to FIG. The third operation, for example, a plurality of VNFa ₁ ~a _m is executed each time it is re-arranged by the second operation described above.

Step S300: The control unit 13, after rearranging VNFa ₁ ~a _m by the second operation described above, it acquires the input data X (k-1) and service quality of the actual value _q 0 (k).

Step S301: The control unit 13 inputs the input data X (k-1) and outputs the estimated service quality q ₁ (k) to q _j (k _{) output from the plurality of neural networks F 1 to} F _j. ) To get.

Step S302: The control unit 13 uses a plurality of neural networks based on a comparison between _{each of the estimated service quality values q 1} (k) to q _j (k) and the actual service quality value q _{0 (k).} F ₁ to F _j of accuracy information _s (F 1) to update the ~s _{(F j).}

Step S303: The control unit 13 _{generates a new} neural network F new and stores it in the storage unit 12.

Step S304: The control unit 13, a minimum accuracy of the neural network _{F low} among a plurality of neural networks _F 1 to F _j, replaced with a new neural network _{F new new.} Then the process ends.

Above As mentioned, the orchestrator apparatus 10 according to the present embodiment, a plurality of physical platforms _b 1 ~b _n measurements _d 1 to d _n of performance, and a plurality of physical platforms _b 1 of the plurality of ~b _n VNFa ₁ ~a _m position information _L (a 1) of the type a ~L _{(a m),} stores a plurality of neural networks _F 1 to F _j to output the estimated value q of the service quality of the service chain. The orchestrator device 10 predicts whether or not the future service quality will deteriorate based on the estimated value q of the service quality output from the neural network F. If future service quality is predicted to decrease, the orchestrator 10 is obtained by using the neural network F, a plurality of VNFa ₁ ~a estimate of service quality corresponding to each of the plurality of arrangement patterns of _m Based on q', one arrangement pattern is determined from the plurality of arrangement patterns. The orchestrator 10 is to match the determined the arrangement pattern, rearranging of VNFa ₁ ~a _m multiple physical platforms _b 1 ~b _n.

According to such a configuration, a future quality of service when it is predicted to decrease, more VNFa ₁ ~a _m are rearranged. Thus, for example because before the service quality actually decreases more VNFa ₁ ~a _m may be rearranged, future service quality of the service chain is possible to reduce the probability to be lowered.

In addition, in the conventional auto scaling and disaster recovery, since the service chain is moved or the routing is changed after the failure or performance deterioration occurs, the VNF communication packet of the chain in the middle may be lost from the queue. was there. Therefore, in order to improve the reliability of the service, the service provider has to pay a high fee to the platform service provider to secure abundant resources. As a compensation, there was a problem that the reserved resources were idle when the service request was small. In this embodiment, in order to improve reliability, rather than reserving a certain amount of resources at an expensive price, a large number of inexpensive services are reserved, a copy of the service chain connecting them is prepared, and performance is provided at regular intervals. by the measures provided to the machine learning, to estimate the future quality of service of the service chain, by relocating VNFa ₁ ~a _m when future service quality is predicted to decrease, reliability at low cost It makes it possible to provide highly reliable services.

Although the present invention has been described with reference to the drawings and examples, it should be noted that those skilled in the art can easily make various modifications and modifications based on the present disclosure. Therefore, it should be noted that these modifications and modifications are within the scope of the present invention. For example, the functions and the like included in each means or each step and the like can be rearranged so as not to be logically inconsistent, and a plurality of means or steps and the like can be combined or divided into one. ..

For example, in the above-described embodiment, the orchestrator device 10 may be divided into a plurality of information processing devices capable of communicating with each other.

Further, in the embodiment described above, the input data X, a plurality of physical platforms _b 1 ~b _n and the measured value _d 1 to d _n of each performance, a plurality of VNFa ₁ ~ in a plurality of physical platforms _b 1 ~b _n and a _m respective position information _{L (a 1) ~L (a} m), a configuration was described containing as a vector component. However, the content of the input data X is not limited to the above configuration, and any observable data that may affect the service quality of the service chain may be included in the input data X. When the number of components of the input data X is different from the above-described embodiment, it is necessary to specify the number of neurons included in the input layer of the neural network F according to the number of components of the input data X.

Further, in the above-described embodiment, a method of acquiring an estimated value q of the future service quality of the service chain using the neural network F and predicting whether or not the future service quality will deteriorate based on the estimated value q. Was explained. However, the method of predicting whether or not the quality of service will deteriorate in the future is not limited to this example. For example, it has been conventionally observed that small-scale failures occur periodically or aperiodically before the service quality deteriorates due to the occurrence of a large-scale failure in a network service system. Therefore, by learning the occurrence pattern of such a small-scale failure as a sign of the occurrence of a large-scale failure, a binary signal indicating whether or not the future service quality will deteriorate, or the future service quality It is also possible to create a neural network that outputs a signal indicating the probability that the quality of service will decrease, and to predict whether or not the quality of service will decrease in the future based on the output of the neural network. With this configuration, it is possible to learn service chain failures and performance due to time-series factors with a neural network, and if necessary, bypass service chains (placement patterns) where failures are foreseen in the near future. This will improve the reliability of service chains that use low-cost VNFs.

Further, in the above-described embodiment, the configuration in which the orchestrator device 10 uses a plurality of neural networks F _{1 to} F _j has been described. However, for example, the orchestrator 10 generates and stores a one neural network F, the configuration may be rearranging VNFa ₁ ~a _m by using the one neural network F. In such a case, it is not necessary to determine the accuracy information s (F) of the one neural network F.

Further, in the above-described embodiment, an example in which the position information L (a) of VNFa is represented by a vector has been described. However, if, for example, the locations of n physical platforms are arranged in a straight line and the locations can be linearly affected by the quality even if the locations are represented by a scalar, the position information L (a) of VNF may be represented by a scalar. good. Alternatively, it may be represented by a scalar (binary data) in which _{each component g 1 to g n} of the position information L (a) represented by a vector in the above-described embodiment is arranged.

Further, in order to perform online learning of the neural network F, the information processing system 1 may have a function of intentionally adding obstacles and performance deterioration. According to such a function, learning data when a failure or performance deterioration occurs can be easily obtained.

Further, for example, an information processing device such as a smartphone or a computer can be configured to function as the orchestrator device 10 according to the above-described embodiment. Specifically, a program describing processing contents for realizing each function of the orchestrator device 10 according to the embodiment is stored in the memory of the information processing device, and the processor of the information processing device reads and executes the program. Therefore, the invention according to the present embodiment can also be realized as a program that can be executed by the processor. The program can be recorded on a recording medium and provided, or can be provided through a network.

1 Information processing system 10 Orchestrator device 11 Communication unit 12 Storage unit 13 Control unit 20 Probe 30 Network a VNF
b Physical platform

Claims (8)

An orchestrator device that manages a service chain composed of multiple virtual network functions (VNFs) arranged on multiple physical platforms in cooperation with each other.
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Department and
Based on the estimated value of the future service quality of the service chain output from the neural network, it is predicted whether or not the future service quality of the service chain will deteriorate.
If it is predicted that the future service quality of the service chain will deteriorate, the estimated value of the future service quality of the service chain corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. One of the plurality of arrangement patterns is determined based on the above.
A control unit that rearranges the plurality of VNFs on the plurality of physical platforms so as to match the determined arrangement pattern.
An orchestrator device equipped with. The orchestrator device according to claim 1.
The storage unit stores the plurality of neural networks and the accuracy information of each of the neural networks.
The control unit
Based on the estimated value of the future service quality of the service chain output from one of the most accurate neural networks among the plurality of neural networks, it is predicted whether or not the future service quality of the service chain will deteriorate. ,
After rearranging the plurality of VNFs, the said value of each of the neural networks is based on a comparison between the estimated value of the service quality of the service chain output from each of the neural networks and the actual value of the service quality of the service chain. An orchestrator device that updates accuracy information. The orchestrator device according to claim 1 or 2.
Based on the estimated value of the future service quality of the service chain corresponding to each of the plurality of arrangement patterns of the plurality of VNFs, the control unit sets a probability distribution in which each of the arrangement patterns is a random variable. An orchestrator device that determines that a higher arrangement pattern has a higher probability, and determines one arrangement pattern from the plurality of arrangement patterns according to the probability distribution. A program that causes an information processing device to function as the orchestrator device according to any one of claims 1 to 3. An information processing system including a plurality of physical platforms and an orchestrator device for managing a service chain configured by linking a plurality of virtual network functions (VNFs) arranged on the plurality of physical platforms.
The orchestrator device is
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Department and
Based on the estimated value of the future service quality of the service chain output from the neural network, it is predicted whether or not the future service quality of the service chain will deteriorate.
If it is predicted that the future service quality of the service chain will deteriorate, the estimated value of the future service quality of the service chain corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. One of the plurality of arrangement patterns is determined based on the above.
A control unit that rearranges the plurality of VNFs on the plurality of physical platforms so as to match the determined arrangement pattern.
Information processing system. The information processing system according to claim 5.
The storage unit stores the plurality of neural networks and the accuracy information of each of the neural networks.
The control unit
Based on the estimated value of the future service quality of the service chain output from one of the most accurate neural networks among the plurality of neural networks, it is predicted whether or not the future service quality of the service chain will deteriorate. ,
After the rearrangement of the plurality of VNFs, the said value of each of the neural networks is based on a comparison between the estimated value of the service quality of the service chain output from each of the neural networks and the actual value of the service quality of the service chain. An information processing system that updates accuracy information. It is a control method of an orchestrator device that manages a service chain composed of multiple virtual network functions (VNFs) arranged on multiple physical platforms in cooperation with each other.
Storage that stores one or more neural networks that input the measured value of the performance of each physical platform and the position information of each VNF on the plurality of physical platforms and output the estimated value of the service quality of the service chain. Steps and
A prediction step for predicting whether or not the future quality of service will deteriorate based on the estimated value of the future quality of service of the service chain output from the neural network.
If it is predicted that the quality of service will deteriorate in the future, the plurality of arrangements will be based on the estimated value of the quality of service corresponding to each of the plurality of arrangement patterns of the plurality of VNFs acquired by using the neural network. A step of determining one of the above-mentioned arrangement patterns, and
A step of rearranging the plurality of VNFs on the plurality of physical platforms so as to match the determined arrangement pattern, and
Control methods, including. The control method according to claim 7.
The storage step stores the plurality of neural networks and the accuracy information of each of the neural networks.
The prediction step predicts whether or not the quality of service will deteriorate in the future based on the estimated value of the quality of service output from the neural network having the highest accuracy among the plurality of neural networks.
After rearranging the plurality of VNFs, a step of updating the accuracy information of each of the neural networks based on the comparison between the estimated value of the service quality output from each of the neural networks and the actual value of the service quality is performed. Further including control methods.

Images