Patent Application
20240126612
The present invention relates to a resource allocation device, a resource allocation method, and a resource allocation program.
The distributed cloud environment is an architecture in which DC (data center) is distributed and arranged in an NW (Network) (NPL 1, 2). The DC provides computer resources (resources) for offloading (acting over) processing conventionally carried out by a server or a user. Hereinafter, a state management unit will be exemplified as an object to be allocated to the DC. The “state” is data used for a service to be exchanged in real time among a plurality of users. The “state management unit” is a processing unit that updates a state managed by the state management unit itself on the basis of access contents received from each user and shares the updated state with each user.
In the online system 9z1, a service providing device 1z provides a service for data-sharing the state in real time among a plurality of user terminals 4z. The state managed by the state management unit 5z is, for example, as described below.
In the online system 9z2, a device for operating the state management unit 5z is offloaded from the service providing device 1z to the DC 3z close to a user terminal 4z. By this off-load, the following effects can be obtained.
That is, in order to improve the service quality, it is important to select, from among a plurality of DC 3z candidates, which DC 3z the state management unit 5z should be arranged in. A resource allocation method for obtaining a delay between each user terminal 4z and reducing the maximum value (maximum E2E delay) of the delay is described in NPL 3.
Even a user who uses the environment of the same online system allowed to execute various applications for each group of users. In this case, for example, the following problems are defined.
It is assumed that the following prerequisites have already been set for the above problems.
In the distributed cloud environment 8z, two DCs (DC1, DC2) exist in the NW indicated by the wavy line, and five users (UA1, UA2, UA, UB1, UB2) are accommodated in the DCs. In the example shown in
Here, as described in the above problem definition, it is assumed that the first group (UA1, UA2, UA3: the second character represents group “A”) and a second group (UB1, UB2: the second character represents group “B”) is formed between users are formed. In order for members of the first group to play the same competitive game, a state of UA1 and UA2 (ST1) and a state of UA3 (ST2) need to be shared (synchronized) between DCs. The same applies to the second group.
In addition, service requirements may be different between and application of the first group and an application of the second group.
For example, in a survival game played by the first group, emphasis is placed on having a large number of participants, such as 100 people, and information about opponents is not as strictly required as in fighting games.
On the other hand, in a competitive fighting game played by the second group, the screen information seen by the members and the operation information input by the members need to be reflected quickly (frame by frame) to the opponent side, although the number of players is small, such as one against one.
A method for efficiently performing resource allocation suitable for each group while taking such service requirements into consideration is required. However, the conventional resource allocation technique such as the one in NPL 3 has a plurality of groups, and each group cannot cope with such a complicated situation that requires various service requirements.
Therefore, the present invention mainly aims to perform resource allocation according to requirements for each group formed by a plurality of users.
In order to solve the problems described above, a resource allocation device of the present invention has the following characteristics.
The present invention includes: a request receiving unit that receives a request to allocate a state management unit for sharing a state for each group composed of a plurality of user terminals within a group, to any of a plurality of data centers deployed in a distributed cloud environment; and an allocation calculation unit that calculates an allocation cost of allocating the state management unit to each of the data centers by using an allocation index corresponding to a requirement requested by the group, and determines a data center to which the state management unit is allocated, in accordance with the allocation cost.
According to the present invention, resource allocation according to requirements for each group formed by a plurality of users can be performed.
Embodiments of the present invention will be described below with reference to the drawings.
Therefore, state management units 5 manage the states of the respective groups. Each state management unit 5 is arranged in one of the DCs 3.
First, the resource allocation device 2 performs group matching processing such as (procedure 1A) to (procedure 5A).
As a detail of (procedure 4A), the DC 3 performs processing for updating the state managed by the state management unit 5 itself of each group. For example, in an online competitive game, each user terminal 4 sends a command to the DC 3 as shown in the following (procedure 1B) to (procedure 4B), and at the same time a virtual space (state) which is already synchronized and generated is received from the DC 3. Thus, the user terminal 4 can acquire the command of the other user terminal 4 only by performing one-to-one communication with the DC 3.
Although an example of an online competitive game is described above, in the above-described manner, the distributed cloud environment 8 may be applied to a case in which a plurality of specific modules are combined to obtain a final processing result in an IoT (Internet of Things) environment or the like. For example, a user terminal 4 for a weather prediction module, a user terminal 4 for a soil analysis module, and a user terminal 4 for a crop analysis module are individually prepared. Then, the three modules are grouped, and a processing unit (processing unit corresponding to the state management unit 5) for receiving the output result of the module group belonging to the group, processing them and managing a field is arranged in the DC 3.
In (procedure 3A) the resource allocation device 2 selects a DC 3 in which the state management unit 5 offloaded from the service providing device 1 is to be arranged (hereinafter referred to as “DC arrangement”), in accordance with a requirement for each group formed by a plurality of users (e.g., a requirement for each application used by the group or a requirement for each service used by the group). Therefore, the resource allocation device 2 includes a request receiving unit 21, an NW data collection unit 22, an allocation calculation unit 23, And a control unit 24.
The request receiving unit 21 receives a request for each group including a group configuration and a user delay requirement from the service providing device 1. This request a content of allocating the state management unit 5 for exchanging data for each group composed of a plurality of user terminals 4, to any of a plurality of DCs 3 arranged in the distributed cloud environment 8.
The NW data collection unit 22 measures and collects NW data including delay information between each user terminal 4 and each DC 3 and free capacity (the number of allocatable users) information of each DC 3 from the distributed cloud environment 8.
The allocation calculation unit 23 calculates the DC arrangement for each group according to a resource allocation scheme (hereinafter referred to as “scheme”) on the basis of a request from the request receiving unit 21 and NW data from the allocation calculation unit 23. The scheme information for determining to which DC the state management unit 5 is to be allocated (how much and what policy is to be allocated (which allocation indicator is primarily focus on), and the performance of the distributed cloud environment 8 strong depends on the scheme.
In the scheme, the resource allocation device 2 may receive the request from the service providing device 1.
Alternatively, the allocation calculation unit 23 may obtain the scheme by referring to the database on the basis of the application type received as a request from the service providing apparatus 1 by the resource allocation device 2. A scheme corresponding to each application type is registered in advance in the database.
Alternatively, the allocation calculation unit 23 may receive the designation of the scheme from members of the group. The control unit 24 allocates each of the state management units 5 to each of the DCs 3 according to the DC arrangement from the allocation calculation unit 23.
The resource allocation device 2 is configured as a computer 900 that includes a CPU 901, a RAM 902, a ROM 903, an HDD 904, a communication I/F 905, an input-output I/F 906, and a media I/F 907.
The communication I/F 905 is connected to an external communication device 915. The input-output I/F 906 is connected to an input-output device 916. The media I/F 907 reads data from a recording medium 917 and writes data into the recording medium 917. Further, the CPU 901 executes a program loaded into the RAM 902 (the request receiving unit 21, the NW data collection unit 22, the allocation calculation unit 23, a resource allocation program provided in the control unit 24), thereby controlling each processing unit. The program can be distributed via a communication line or can be recorded on the recording medium 917 such as a CD-ROM to be distributed.
On the other hand, in the distributed cloud environment 8 shown in
Also, the allocation calculation unit 23 can make DC arrangement suitable for requirements for each group, by referring to a scheme suitable for an application type for each group from the request.
That is, the allocation calculation unit 23 calculates an allocation cost of allocating the state management unit 5 to each data center, by using, an allocation index of a scheme corresponding to requirements for each group, and determines a data center to which the state management unit 5 is allocated according to the allocation cost. Examples of combinations of application types and schemes will be described below.
“Case 1: Complete synchronization type” is an application for allowing all users belonging to the group to browse the same screen by synchronizing states after waiting for communication of all users belonging to the group one by one. Hereinafter, a complete synchronous application will be described.
A complete synchronous scheme (allocation index) minimizes “maximum delay in a group.” This is because QoS (Quality of Service) and QoE (Quality of Experience) of the entire group depend on the user with the highest delay.
“Case 2: Semi-synchronization type” is an application that does not (cannot) take strict synchronization between users, and is, for example, a survival game of about 100 FPS (First-Person Shooter). By synchronizing the states without waiting for communication of a user with high delay who belongs to the group, a screen viewed by the user with high delay becomes choppy, and the latest position of the character of the user with high delay is not visible to other users.
The semi-synchronous scheme (allocation index) minimizes the “average delay within a group.” For example, if only about three members of a group of 100 experience large delays, it is sufficient if only three members feel the inconvenience and the remaining 97 are provided with a comfortable environment with a small average delay.
“Case 3: Fair environment type” is a real-time application which requires fairness in a play environment between users, and is exemplified below.
The fair environment scheme (allocation index) minimizes “variance in delay (variance or standard deviation) in a group.”
Generally, the larger the delay, the lower the QoS/QoE, and the more it affects the advantage (e.g., game score) in the virtual world. Therefore, it is desirable to reduce variations in delay in the group, not only in the far environment type application but also in the application of other cases.
Although three types of schemes have been described above, a plurality of allocation indexes may be combined, for example, by simultaneously minimizing “average delay in a group” and “maximum delay in a group” as the semi-synchronous scheme. Simdlarly, as the fair environment scheme, three allocation indexes, “average delay in the group,” “maximum delay in the group,” and “variation in delay in the group,” may be minimized in a balanced manner.
In the following description, a formulation for achieving both minimization and equalization of delay in a group is performed. More specifically, a model capable of evaluating the three assignment indexes in a comprehensive manner will be described. First, parameters used in the model are defined.
These parameters are given to the allocation calculation unit 23 as a request from the request receiving unit 21 and data collected by the NW data collection unit 22.
The allocation calculation unit 23 calculates (equation 1) and (equation 2) as allocation indexes. The calculation formula of the maximum delay will be described later in (equation 4). The left side “ajk” of (equation 1) indicates the average delay of a user i∈Ij when the group j is accommodated in the DCk. The left side “v2jk” of (equation 2) indicates the delay dispersion of the i∈Ij when the group j is accommodated in the DCk.
(Equation 3) represents an objective function of DC allocation. This objective function results in a problem of minimizing the sum of the allocation costs of the group j. The symbol “cjk” is an allocation cost when the group j is accommodated in the DCk.
The symbol “xjk” is the calculation result of the allocation calculation unit 23. “Xjk” is a decision variable and takes a value “1” when the group j is accommodated in the DCk, and takes a value “0” otherwise.
As a concrete calculation formula for obtaining the allocation cost Cjk by the allocation calculation unit 23, (equation 4) or (equation 5) will be described. As a numerical value indicating the variation in delay, dispersion is adopted in (equation 4), and standard deviation is adopted in (equation 5). By using the standard deviation in place of the dispersion, the term including the square is eliminated, so that the scale of each term can be adjusted to some extent.
α, β (α, β≥0, α+β≤1) are parameters for adjusting trade-off between the three allocation indexes, and according to service requirements or the like, a service provider, a user, or an NW provider or the like arbitrarily determine the result.
The “1−α−β” of the first term on the right side of (equation 4) is a hyper parameter that emphasizes “average delay within a group” as its value increases.
The “α” of the second term on the right side of (equation 4) is a hyper parameter that emphasizes “variation in delay within a group” as its value increases.
The “β” of the third term on the right side of (equation 4) is a hyper parameter that emphasizes “maximum delay within a group” as its value increases. The calculation formula after of the third term on the right side third is the calculation formula of the maximum delay in the group.
In other words, in (equation 4) or (equation 5), the priority of the three kinds of indexes can be adjusted according to the distribution of the three hyper parameters as follows.
The allocation calculation unit 23 may use (equation 6) instead of (equation 4) using the variance, as a calculation equation for obtaining the allocation, cost Cjkor may use (equation 7) instead (equation 5) using the standard deviation.
In (equation) 6 and (equation 7) three hyper parameters (α, β, γ) are used. Thus, the degree of freedom of parameter setting is high, but handling (optimum parameter setting) becomes difficult.
The “α” of the first term on the right side is a hyper parameter that emphasizes “average delay within a group” as its value increases.
The “β” of the second term on the right side is a hyper parameter that emphasizes “variation of delay within a group” as its value increases.
The “γ” of the third term on the right side is a hyper parameter that emphasizes “maximum delay within a group” as its value increases.
On the other hand, in (equation 4) and (equation 5), the number of parameters is reduced by one, and the setting range is narrow. However, However, since the weight can be set by ratio, it is likely to be easier to handle than using three types of hyper parameters (α, β, γ).
(Equation 8) represents constraint conditions (subject to) corresponding to the objective function of (equation 3). The constraint conditions as follows.
The problem formulated as (equation 3) to (equation 8) is a combination optimization problem, and in order to obtain a global optimal solution, the allocation calculation unit 23 needs to investigate all combinations (brute force calculation). The computational complexity of this problem in this case is huge, being the computational complexity order (n-th power of m), where n is the number of groups and m is the number of DCs.
Therefore, as an approximation algorithm used by the allocation calculation unit 23 in place of the brute force calculation, greedy algorithm and a local search method will be described in order. The allocation calculation unit 23 uses the greedy method alone or uses the greedy method and the local search method in combination. Thus, an optimum solution (or a semi-optimum solution having a score close to the optimum solution) obtained by the brute force calculation can be obtained with a smaller computational complexity than the total brute force calculation.
The allocation calculation unit 23 creates a cost table (to be described later in
The allocation calculation unit 23 sorts the minimum value of the cost in ascending order for each group of the cost table (S103).
The allocation calculation unit 23 substitutes an initial value 1 in a variable j of the group (S104).
Here, the allocation calculation unit 23 executes a loop for sequentially selecting a group j up to the number of groups of the cost table one by one (S105 to S107), and calls a subroutine (
The allocation calculation unit 23 reads a j-th group line of the cost table (S111), and tries to allocate a DC whose cost is minimum (S112).
The allocation calculation unit 23 allocates the group j to a DC when the user delay requirement is satisfied (S113, Yes) and the DC has a sufficient capacity (S114, Yes) (S115). On the other hand, in case of (S113, No) or (S114, No), the allocation calculation unit 23 deletes the DC from the cost table of the j-th group (S116).
A cost table 201 is a two-dimensional table in which the group j is a row and the DCk is a column with respect to the allocation cost Cjk when the group is accommodated in the DCk. The allocation calculation unit 23 creates the cost table 201 in S101 in
DC allocation tables 211 to 217 are obtained by making the calculation result “xjk” of the allocation calculation unit 23 based on the cost table 202 into a table form. For example, in the DC allocation table 214, two groups of “G4, G2” are allocated to the DC 1, no group is allocated to the DC 2 (symbol “−”), and one group “G5” is allocated to the DC 3.
According to the following procedure, the allocation calculation unit 23 allocates each group to each DC by calling a subroutine of S106 in order from the group positioned at the upper rank of the cost table 202.
The greedy method has been described above with reference to
A pseudo code for explaining the algorithm of the greedy method in detail is exemplified hereinafter. This pseudo code is a procedural language for performing substitution “A←1” (substitution of a value 1 to a variable A) repetitive control “for to end for, while to end while,” and branch control “if to end if.” Further, line numbers (L01, L02, ) for explanation are added to the head of the pseudo code. Each function (function) performs a predetermined calculation based on the input variable given by “Input:˜” and responds (returns) the result of the calculation as an output variable indicated by “Output:˜.”
In the Greedy_Allocation function (α, β, wij, dik, Di, Ck), a cost table is created using α and β (L05) and preprocessing, for performing (L11) DC allocation is performed preferentially starting from the group with the lowest cost in the table. Actual DC allocation is executed by calling the GroupDC_Mapping function with arguments wij, dik, Di, Ckthe cost table c created here, and the permutation J referring to the cost table as arguments (L12).
In the flow chart shown in
In the GroupDC_Mapping function (J, c, w, d, D, C), the given cost table c is referenced in order (L17) according to the permutation J (i.e., a certain row c[j] in the cost table is extracted) and the DC_Selection function is called sequentially to assign a DC to each group to obtain an allocation (L20, L23). If the DC_Selection function indicates that there is no DC capable of satisfying the user delay requirement D, the delay requirement D is ignored and the allocation is performed.
Note that X[j]←k in L20 and L23 is equivalent to the value 1 (xjk′=1 if k′=k) if there is allocation, or 0 (xjk′=0 otherwise) if there is no allocation.
In the flow chart shown in
In addition, the DC_Selection function is shown.
In the DC_Selection function (c[j], Ij, d, D, C), for group member Ijamong the DCs that satisfy the delay requirement D for each user and whose remaining capacity C is greater than the group size |Ij| (L36), the DCk with the lowest delay cost c[j] [k] is selected for allocation (L37, L38).
In the flowchart shown in
As described above, when the greedy method is used alone, the allocation calculation unit 23 can call the Greedy_Allocation function. Thus, by sorting the groups and preferentially allocating the groups, starting from the groups whose costs are low, highly accurate optimization can be done with a small calculation time. On the other hand, the allocation calculation unit 23 may obtain a better solution by combining a proposal method and a local search method.
The following is a pseudo-code of an algorithm in which a proposal method and a local search method are combined.
In the pseudo codes (M01 to M11), first, the allocation calculation unit 23 executes the Greedy_Allocation function (M01). Next, the alllocation calculation unit 23 executes the GroupDC_Mapping function a fixed number of times (N times of M04) (M06). At this time, the allocation calculation unit 23 changes the permutation of the groups randomly in the part of “sorting of the cost table” (M05). Then, the allocation calculation unit 23 selects the allocation whose entire cost. is the lowest among N times of repetition (M08 to M10). This pseudo code (M01˜M11) may make the allocation cost even smaller than the greedy method, although the additional number of searches to the effect that the GroupDC_Mapping function is repeated N times increases the computational cost. The number N of searches is arbitrarily set by the service provider or the distributed environment operator.
Here, the number of groups given to the request receiving unit 21 is focused and classified as follows.
The Greedy_Allocation function described so far assumes that a coherent number of groups are given to the request receiving unit 21 (offline allocation or batch processing). On the other hand, if the groups to be allocated to the request receiver 21 come sequentially, the following pseudo-code for online allocation may be executed.
In the pseudo codes (N01 to N12) first, upon reception of an allocation request for a group j (N02), the allocation calculation unit 23 calculates all allocation costs cjk between the group j and each DC(N03), and create a cost table only for the groupj (N04).
Then, the allocation calculation unit 23 executes the DC_Selection function for the group j (N05, N08). In so doing, the allocation calculation unit 23 always monitors the capacity Ck=C[k] of each DC and increases the capacity by that amount when the group j leaves the DC.
Below is an example of the number of function calls for each classification to show the computational complexity for each of (Classification 1) to (Classification 3).
When only the greedy method is executed in (classification 1), the Greedy_Allocation function is executed once, the GroupDC_Mapping function is executed once, and the DC_Selection function is executed 10 times or more. The number of times the DC_Selection function is called more than 10 times is when there is no DC that can satisfy the user delay requirement (when the return value of line L20 is ∞), and the DC_Selection function is moved to line L23 and called again. When the greedy method and the local search are executed in combination in classification 1, the Greedy_Allocation function is executed once, and then the GroupDC_Mapping function and the DC_Selection function are repeated N times (the Greedy_Allocation function is executed once, the GroupDC_Mapping function is executed 1+N times, and the DC_Selection function is executed (1+N) times or more). N is a parameter which can be arbitrarily set by an operator or the like.
When only the greedy method is executed in (Classification 2), the following is performed.
When the greedy method and local search are executed together in (Classification 2) the Greedy_Allocation function is performed once each, and then the GroupDC_Mapping and DC_Selection functions are repeated N times during each batch processing as one set, which is as follows.
It is not possible to execute both the greedy method and local search in (Classification 3). If there is only one group, there is no room for local search because the order of allocation is irrelevant.
The resource allocation device 2 of the present invention includes the request receiving unit 21 for receiving a request to allocate the state management unit 5 for sharing a state for each group composed of a plurality of user terminals 4 within a group, to any of a plurality of DCs 3 arranged in the distributed cloud environment 8, and the allocation calculation unit 23 for calculating an allocation cost of allocating the state management unit 5 to each DC 3 by using an allocation index corresponding to requirements requested by the group, and determining the DC 3 as the allocation destination of the state management unit 5 according to the allocation cost.
This allows a user terminal 4, a member of the group, to assign the state management unit 5 to the allocation destination suitable for the application to be executed by the group sharing the state management unit 5. Thus, resource allocation according to requirements for each group formed by a plurality of users can be performed.
The allocation calculation unit 23 obtains delay time between each of the user terminals 4 constituting a group and the DC 3, and the average delay of the obtained delay time is used as an allocation index for the state management unit 5.
Thus, in an application such as a survival game in which many people participate, an allocation destination satisfying the many people can be found.
The present invention is characterized in that the allocation calculation unit 23 determines the delay time between each user terminal 4 and the DC 3 that comprises the group, and the maximum delay of those determined delay times is used as the allocation index for the state management unit 5.
This allows the discovery of a suitable allocation destination for an application that requires strict synchronization in the communication of all users belonging to a group, such as fighting games.
The present invention is characterized in that the allocation calculation unit 23 determines the delay time between each user terminal 4 and the DC 3 that comprises the group and uses the variance or standard deviation of those determined delay times as the allocation index for the state management unit 5.
This allows the discovery of a suitable allocation destination for a real-time application, such as a racing game, where fairness of the playing environment among users is particularly important.
The present invention is characterized that the allocation calculation unit 23 calculates an allocation destination for each group by an approximation algorithm using the greedy method, for a combination optimization problem of minimizing the sum of allocation costs of each group for a request received by the request receiving unit 21.
This allows for the calculation of a reasonable allocation with less computation than a brute force calculation for combination optimization problems.
The present invention is characterized in that the allocation calculation unit 23 calculates the allocation destination for each group by an approximation algorithm that combines the greedy method and the local search method for the combination optimization problem of minimizing the sum of allocation costs of each group for a request received by the request receiving unit 21.
This increases the likelihood of finding even better allocation destinations than if the greedy method were used alone.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2021/005175 | 2/12/2021 | WO |
