JP4286703B2 - Resource planning program - Google Patents

Description

  The present invention relates to a resource plan creation program for creating an arrangement plan for a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters. The present invention relates to a resource plan creation program capable of preventing each computer cluster from being overloaded and efficiently using the computing resources of the computer cluster.

  In recent years, data centers have been operating a plurality of network services entrusted by a plurality of service providers within a data center at the same time. For services with a high load, the allocation of computational resources in the data center is dynamically increased, and for services with a low load, the allocation of computational resources in the data center is changed. Resource allocation control is performed based on a utility method that reduces the amount of data (see, for example, Patent Document 1).

  However, when some services operated in the data center are temporarily overloaded and there is a need to significantly increase the resource allocation for this service, there are additional idle resources to be allocated in the data center. There may be a situation where there is a shortage and resources cannot be added.

  In order to prevent such a situation from occurring, the required resource amount when the total load of all operation services in the data center reaches a peak is predicted, and the peak required resource amount is constantly prepared in the data center. However, in this case, during the period other than the peak time, most of the amount of resources remains unused as idle resources, and the cost performance and computational resource utilization efficiency of the entire data center is reduced.

  In view of this, International Patent Application JP 03/03273 proposes a method using a logistic support data center. In this method, each service is normally operated at a specific data center (hereinafter referred to as “original center”) designated by the service provider that requested the service operation (hereinafter referred to as such service). ("Primary service").

  Then, if the original service is temporarily overloaded due to load fluctuations and the original center itself is overloaded as a whole, one of the original services that have been overloaded on the original center. Are processed in a different data center called a logistic support center, an auxiliary service (hereinafter referred to as “derivative service”) corresponding to the original service is generated on the logistic support center and Start operation.

  As a result, a part of the load borne by the overloaded original service can be taken over by the derived service on the logistic support center, and the overload state of the original center as a whole can be eliminated.

  Here, the distribution ratio of the user request amount arriving at each service to the original service and the derived service is set in the load balancer or redirector in the network, and the load balancer etc. is based on the set ratio. The amount of user requests arriving at is distributed between the primitive service and the derived service.

  In such a conventional data center operation mode, there are generally a plurality of original centers and one logistic support center in many cases. Also, the division of roles between the original center and the logistic support center was clear, and the logistic support center was used exclusively for the operation of derived services.

JP 2002-24192 A

  However, in the conventional technology in which a specific data center is defined as a logistic support center and all derived services are operated exclusively in the data center, a number of original services on a plurality of original centers can be simultaneously used as one logistic support center. On the other hand, there is a problem that the logistic support center becomes overloaded as a whole when the requests for generation of derived services are concentrated.

  In addition, if the amount of resources required at the logistical support center is predicted when the concentration of generation requests for derived services reaches a peak, and the required amount of resources at the peak is constantly prepared in the logistic support center, Most of the computing resources of the support center will not be used, and resource utilization efficiency will decrease.

  The present invention has been made to solve the above-described problems caused by the prior art, and prevents each computer cluster, which is a collection of computer resources such as a data center as a typical example, from being overloaded. Another object is to provide a resource plan creation program that can efficiently use the computing resources of a computer cluster.

In order to solve the above-described problems and achieve the object, the present invention provides a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters including a plurality of computers to the plurality of computer clusters. A resource plan creation program for creating a resource plan to be deployed, wherein an operation of a computer cluster to be operated from each computer cluster to another computer cluster based on a load prediction of the source service from a service in which the computer cluster to be operated is determined Derived service generation procedure for generating a part of the derived service for each time slot obtained by dividing the prediction period of load prediction into a plurality of time slots at a predetermined time interval, and the derived service generated by the derived service generating procedure as resource planning for optimal arrangement of the group to the plurality of computer cluster, each time slot An arrangement plan sequence in which time slot arrangement plans, which are derived service arrangement plans, are arranged in the order of the timeslots, is similar to the load equality between computer clusters and the time slot arrangement plan between adjacent time slots on the time axis. And a plan creation procedure for creating a plan using the degree as an evaluation index .

According to this invention, a derived service in which a part of the operation is transferred from each computer cluster to another computer cluster based on the load prediction of the source service from the source service, which is a service for which the computer cluster to be operated is determined , is loaded. Arrangement of derived services for each time slot as a resource plan that generates a forecast period for each time slot divided into a plurality of time slots at a predetermined time interval and optimally allocates the derived service group to a plurality of computer clusters An arrangement index series that arranges time slot allocation plans, which are plans, in the order of time slot time, and evaluates the load equality between computer clusters and the similarity of time slot allocation plans between adjacent time slots on the time axis since it is configured to create used as, for service computer cluster overload operated one It can take over the entire computer cluster system.

  According to the present invention, a part of services operated by an overloaded computer cluster is taken over by the entire computer cluster system, so that each computer cluster is prevented from being overloaded and the computing resources of the computer cluster are efficiently allocated. There is an effect that it can be used.

  Exemplary embodiments of a resource plan creation program according to the present invention will be described below in detail with reference to the accompanying drawings. Here, the case where the present invention is applied to the creation of a resource plan in which computing resources of a plurality of data centers are assigned to a plurality of services will be mainly described.

  First, the concept of the resource plan according to the present embodiment will be described. FIG. 1 is an explanatory diagram for explaining the concept of the resource plan according to the present embodiment. The graph on the left side of FIG. 1 plots the annual load fluctuation prediction regarding the WEB service provided by the XX ticket sales company.

  The height indicated by the horizontal dotted line in this graph represents the amount of load that can be processed by the amount of resources owned by the own center of the XX ticket sales company. The load fluctuation in this graph has two peaks, and at any peak, the load amount (the height of the dotted line) that can be processed by the company's center resource has been exceeded. Therefore, in the resource plan according to the present embodiment, a derived service is generated in a data center other than its own center at the peak time, and the load exceeding the dotted line is taken over.

  Here, in the peak generation period on the left side, almost no resources are consumed in the data center X, whereas most resources are consumed in the data center Y by other services. Accordingly, the derivative service is arranged in the data center X at the left peak.

  On the other hand, during the peak occurrence period on the right side, most of the resources in the data center X are consumed by other services and the resources in the data center Y are hardly consumed. To do.

  That is, in the resource plan according to the present embodiment, among the plurality of original centers, an original center that temporarily has a large amount of idle resources due to a low load on the operating primitive service is selected. Treat it as a temporary logistic support center and operate the derivative service at the temporary logistic support center.

  Based on such an operation mode, the division of roles between the original center and the logistic support center is not unique to each data center, the data center X constantly operates the primitive service A, and the data center Y When the source service B is constantly operated, the data center X may always have the role of the original center for the source service A and the role of the logistic support center for the source service B. . Hereinafter, it is assumed here that the roles of the original center and the logistic support center are determined by the relative relationship with the service as described above.

  Here, when trying to operate a large number of primitive services in an environment where a large number of data centers are connected by a network, when a certain primitive service is overloaded in the future, which data center can be The question is whether to select a logistical support center. That is, there is a problem of when and at what time each derived service expected to occur in the future should be placed in which data center.

  However, it is necessary to consider the calculation amount and the calculation time when searching for the optimum solution for how to arrange each derived service occurring at a certain future time point in each computer cluster. For example, if there are three data centers and 30 derived services are generated simultaneously at a certain future time point, one derived service has two data centers other than the original center of the originating service. Therefore, 30 derived services can be arranged in 3 data centers in 2 30 powers (1 billion or more).

  Therefore, it is wasted a lot of time to evaluate all of these by a simple search, and in finding the optimal solution for the arrangement of the derived service set in the data center set, an optimal solution that can suppress the number of search steps as little as possible. A search method is required.

  Specifically, when a set of data centers and primitive services is given, and the original center that operates each primitive service is specified, the future occurrence time of the derived service is predicted in advance, and the calculation resources of all data centers In order to improve the use efficiency, it is necessary to optimize in which data center each derived service generated at various times is operated (center arrangement of the derived service) (FIG. 2).

  In addition, the configuration of the derived service set, the load amount processed by each derived service, the resource consumption status of each data center, etc. change from time to time, so the optimal arrangement of the derived service set to the data center group over time It is necessary to redo periodically. When the above optimization of arrangement is periodically performed again, the time axis is divided into periods called time slots, and an optimum derived service arrangement is obtained for each time slot (FIG. 3).

  Therefore, finally, the optimum candidate of the arrangement pattern of the derived service set in the data center group is obtained for every fixed time period (each time slot), and the information in which the time series of the time slots are arranged in time order is obtained. It is necessary to output. This output information is hereinafter referred to as a data center group resource plan.

  In addition, when searching for the time series of the optimum derived service arrangement, the search space becomes huge depending on the number of derived services that occur at the same time, and a simple search takes a lot of time. In such resource planning, the following measures are taken.

  The solution to be searched is generally evaluated by a plurality of indices, and a solution that optimizes all the indices as much as possible must be finally selected, but at least one of the indices in the search space seems to be optimal. Narrow down a set composed of various solutions (Pareto solution set), select an arbitrary solution from the Pareto solution set according to a specific criterion, and set the solution as a provisional solution.

  Then, the provisional solution is asymptotically brought close to the true optimum solution by the NDC replacement operation described later, and when the evaluation value of the provisional solution enters a certain convergence radius, the provisional solution is output as the final optimum solution.

  The services operated by the data center include a transaction-type service that requires a response within a few seconds after receiving a user request, such as a WEB service, and a daily and monthly service. There is a batch-type service in which it is necessary to complete processing of a plurality of job groups input at a predetermined time in a predetermined cycle by a predetermined deadline time.

  A batch-type service is composed of a plurality of jobs, and is generally divided into a plurality of times, and a plurality of jobs are collectively put into a data center at a predetermined time. This job submission schedule is repeated at regular intervals such as a daily cycle or a one-month cycle. The job submission schedule is specified in advance by a consignor who has entrusted the operation of the batch service.

  As for the batch type service, as long as all of the input jobs are processed by the designated deadline time, each job may be processed anytime, so that it does not require real time.

  Therefore, when the primitive service set operated in the data center group is composed of two types of services, a transaction type service and a batch type service, it is considered that the transaction type service is a service that requires real-time property. However, the computing resources of the data center group are preferentially allocated to the transaction type service.

  After that, it is necessary to generate a resource plan that allocates surplus computing resources that have not been allocated to the transaction type service to processing of the batch type service. Therefore, to generate a resource plan for a data center group, first create a resource plan on the assumption that only the transaction-type primitive service is operated in the data center group, and assign it to the transaction-type service in the resource plan. A resource plan for a batch service is generated for surplus computing resources that have not been assigned.

  Also, regarding the load fluctuation prediction of batch-type primitive services, since the job input schedule of batch-type services is known in advance, it is not possible to predict the future load fluctuation of services using a mathematical prediction algorithm such as the ARIMA model method. In the future, the transition of future load fluctuations can be planned by a specific job scheduling algorithm.

  Therefore, in the generation of a resource plan for a set of batch-type primitive services, the future load fluctuations of the batch-type primitive services are not predicted, but those that are planned in advance by a job scheduling algorithm are used.

  The specific job scheduling algorithm will be described later. The characteristic of the scheduling algorithm is that the jobs that make up the batch-type primitive service are scheduled in the order of late submission time, and the jobs to be scheduled are scheduled. The job is scheduled at the time when the surplus computing resources of the data center group are the largest in the time zone between the time when the job is submitted and the deadline time.

  The reason for adopting such a job scheduling algorithm is that the processing of the load of the batch type service is selected as much as possible when the load of the transaction type service is free.

  Next, the configuration of the resource plan creation device according to the present embodiment will be described. FIG. 4 is a functional block diagram illustrating the configuration of the resource plan creation device according to the present embodiment. As shown in the figure, the resource plan creation apparatus 100 includes a transaction plan unit 110, a batch plan unit 120, and a resource plan transmission unit 130.

  The transaction planning unit 110 is a processing unit that creates a resource plan for a transaction-type service, and includes a prediction unit 111, a derived service generation unit 112, a DC generation unit 113, an NDC generation unit 114, and an optimal solution search unit. 115.

  In the following description of the transaction planning unit 110, a service, a primitive service, and a derived service mean a transaction-type service unless otherwise specified. In the generation of a resource plan for a transaction type service, the time slot length is manually determined by the operator of the data center group as a unique value for the entire data center group.

  The prediction unit 111 is a processing unit that performs future prediction of load fluctuations of all primitive services to be operated. Here, the period to be predicted is hereinafter referred to as a prediction period, and its length is one day, one month, one year, or the like.

  The derived service generation unit 112 obtains the generation time of the derived service, the number of generations, and the load distribution ratio between each derived service and the original service based on the load fluctuation prediction by the prediction unit 111 and the amount of resources held by the original center. It is a processing unit.

  Examples of the conditions for generating a derived service include the following. FIGS. 5A and 5B are explanatory diagrams (1) and (2) for explaining the generation conditions of the derived service. FIG. 5A shows a case where a derived service is generated when the resource usage rate of the primitive service arranged in the data center exceeds 90%.

  Also, FIG. 5-2 calculates the average of the resource usage rates of the three data centers X, Y, and Z, and when the resource usage rate of each data center becomes 1.2 times or more of the average, the derived service is displayed. The case where it generates is shown.

  Actually, the derived service generation unit 112 calculates the standard deviation of the resource usage rate by the primitive service for each data center, and adopts the generation condition that the derived service is generated when the value exceeds a certain threshold. ing.

  The DC generation unit 113 divides the prediction period into time slots having a predetermined period, and generates all the arrangement combination pattern candidates that determine in which data center the derived service generated in the time slot is arranged for each time slot. It is a processing unit.

  Here, each of the generated arrangement combination patterns is called a DC (Deployment Candidate), and a DC belonging to the set of DCs generated for the time slot (t) is a DC belonging to the time slot (t). Call.

The NDC generation unit 114 is a processing unit that creates an NDC (Normal DC) by excluding a VDC (Violating DC) from a DC belonging to each time slot. Here, the VDC is a DC in which the evaluation results e ₁ ,..., E _k based on a plurality of evaluation indices m ₁ ,..., M _k violate a predetermined condition, and the NDC is a DC other than that.

  The evaluation index is such that the resource utilization efficiency of the entire data center group is improved by optimizing the arrangement combination of the derived service group to the data center group using the evaluation index. There are equality (FIG. 6-1), NDC similarity (FIG. 6-2), and the like. Here, the smaller the evaluation value, the better the evaluation.

  FIG. 6A illustrates an arrangement in which the load arrangement is not uniform in the three data centers X, Y, and Z, and an arrangement in which the load arrangement is uniform and good. Shall be small.

  FIG. 6-2 shows the movement of the derived service and the time slot (K) when moving from the time slot (K-1) to the time slot (K) between the four data centers W, X, Y and Z. The movement of the derived service at the time of shifting to the time slot (K + 1) is shown. Here, it is assumed that the value of the NDC similarity is larger as the number of service movements at the time slot transition with time elapses.

  The optimal solution search unit 115 is a processing unit that creates an optimal NDC sequence based on the Pareto optimality theory. Here, the NDC sequence is a sequence in which NDCs are selected one by one from each time slot and are arranged in time series in the time slot order. That is, creating an optimum NDC sequence creates an optimum resource plan.

Specifically, the optimal solution search unit 115 sums each of the _N evaluation values e ₁ ,..., E _k of each NDC constituting the NDC sequence over the entire NDC sequence, and sets the sequence evaluation value. Let E ₁ ,…, E _k .

Then, out of the k series evaluation values E ₁ ,..., E _k , all the ones having one or more series evaluation values having the minimum value are selected and set as the initial series set. This initial sequence set corresponds to a Pareto solution set in a multi-objective optimization problem when a set of NDC sequences is a solution space and each of k sequence evaluation values E ₁ ,..., E _k is an objective function.

  Then, an NDC sequence that minimizes the value of the sequence evaluation function ξ is selected from the obtained initial sequence set and set as a sub-optimal NDC sequence. Then, for each time slot, each NDC constituting the sub-optimal NDC sequence is compared with all other NDCs belonging to the same time slot as the NDC, and when the following replacement condition is satisfied, the former NDC Is replaced with the latter NDC.

Here, e _i (T _p , c) is the value of e _i for the NDC belonging to the time slot T _p in the suboptimal NDC sequence, and e _i (T _p , o) is the NDC belonging to the time slot T _p. And defined as the value of e _i for NDC not included in the suboptimal NDC sequence.

For all k evaluation values of the NDC in the suboptimal NDC sequence belonging to the time slot T _p , e _i (T _p , c) ≧ e _i (T _p , o) (1 ≦ i ≦ k) If so, replace NDC. That is, as a result of the NDC replacement, if the evaluation values are improved for all evaluation values constituting the set of k evaluation values, the NDC replacement is performed.

  Also, NDC replacement is performed in the following cases. When some evaluation values are improved by replacement and other evaluation values are deteriorated among the evaluation values constituting the set of k evaluation values, the replacement control function is replaced with the improvement amount of the evaluation value to be improved. If the weighted average of the applied values is higher than the weighted average of the values obtained by applying the replacement control function to the amount of deterioration of the evaluation value to be deteriorated, the replacement is performed.

が成り立つならば、ＮＤＣ入れ替え操作を実行する。ここで、φ_iは重み係数、Δe_i(T
_p) = e_i(T_p,c)−e_i(T_p,o)、δe_i(T_p) = e_i(T_p,o)−e_i(T_p,c)、Λ_iおよびλ_iは評価値e_iに関する入れ替え制御関数である。 In other words, the set of integers I = [i | 1 ≦ i ≦ k] can be divided into two subsets I1 and I2, and when i ∈ I1, e _i (T _p , c) ≦ e _i (T _p , o ) And when i ∈ I2 and e _i (T _p , c)> e _i (T _p , o), the following equation (1)

If the above holds, the NDC replacement operation is executed. Where φ _i is a weighting factor, Δe _i (T
_p ) = e _i (T _p , c) −e _i (T _p , o), δe _i (T _p ) = e _i (T _p , o) −e _i (T _p , c), Λ _i and λ _i is a replacement control function for the evaluation value e _i .

  Then, the sub-optimal NDC sequence modified by the NDC replacement is evaluated by the sequence evaluation function ξ, and compared with the value evaluated by the ξ of the sub-optimal NDC sequence before the NDC replacement operation, a threshold with a constant improvement rate is set. If it falls below, the sub-optimal NDC sequence is used as the resource plan for the data center group. If not, NDC replacement is repeated with the NDC sequence modified by NDC replacement as a new sub-optimal NDC sequence.

  Here, a specific example of processing of the optimum solution searching unit 115 will be described with reference to FIGS. FIG. 7 is an explanatory diagram for explaining an outline of the optimization procedure by the optimum solution search unit 115. Here, load equality and NDC similarity are used as evaluation indexes.

  As shown in the figure, the optimum solution searching unit 115 generates an initial candidate group of provisional solutions based on the similarity of NDC as the optimization step (1), and the provisional solution as the optimization step (2). The initial candidate group is evaluated using the load uniformity and NDC similarity, and the one with the smallest evaluation value is set as a provisional solution.

  Then, in the optimization step (3), a better provisional solution is generated by replacing the NDC with the provisional solution. In the optimization step (4), it is determined whether or not the convergence condition is satisfied. Outputs the provisional solution as the final solution, and when the convergence condition is not satisfied, returns to the optimization step (3) and repeats the NDC replacement until the convergence condition is satisfied.

  FIG. 8 is an explanatory diagram for explaining the concept of the NDC table used for explaining the optimization procedure. As shown in the figure, the NDC table is a table in which NDCs belonging to each time slot are arranged in order of good load uniformity. Hereinafter, each optimization step will be described using this NDC table.

  FIG. 9 is an explanatory diagram for explaining the optimization step (1). As shown in the figure, in this optimization step, similar NDC sequences are generated by connecting NDCs having “NDC similarity = 0” between adjacent time slots. For example, in FIG. 9, a plurality of similar NDC sequences such as a similar NDC sequence starting from NDC (1) in time slot (0) and a similar NDC sequence starting from NDC (2) in time slot (0) are obtained. It is done.

  When NDCs are connected to each other, if there is no NDC having “NDC similarity = 0”, the NDC having the lowest NDC similarity is selected. As a result, the generated similar NDC sequence is an NDC sequence having the smallest value of the evaluation index as the evaluation index, and uses these as the initial candidate group of the provisional solution. An NDC sequence composed of NDCs having the minimum load uniformity in each time slot is an NDC sequence having the minimum total load uniformity as a whole sequence, and this is also included in the initial candidate group of the provisional solution.

  FIG. 10 is an explanatory diagram for explaining the optimization step (2). As shown in the figure, in this optimization step, all similar NDC sequences are evaluated using the load uniformity and NDC similarity as evaluation indexes, and the best similar NDC sequence having the smallest sequence evaluation value Ec is used as a provisional solution. .

  FIG. 11A is an explanatory diagram for explaining the optimization step (3). As shown in the figure, in this optimization step, in each time slot of the provisional solution, NDC other than NDC included in the provisional solution is replaced when NDC satisfies a predetermined condition.

Here, as shown in FIG. 11-2, the predetermined condition is that when there is NDC that simultaneously improves the load uniformity and the similarity of NDC, the value of the evaluation value E _i in the time slot is minimum. Select one and replace it.

  If only one of them improves, do the following: When only the load uniformity improves, ΔL among the NDCs in which the improvement amount ΔL of the load uniformity is larger than the value BL (ΔS) obtained by evaluating the deterioration amount ΔS of the NDC analogy by a predetermined function. When the value of −BL (ΔS) is the largest and only the similarity is improved, the amount of deterioration of the load uniformity is worse than the value BL (ΔS) obtained by evaluating the improvement ΔS of the similarity with a predetermined function. Among NDCs having a larger ΔL, the one having the largest BL (ΔS) −ΔL is selected and replaced.

  FIG. 12 is an explanatory diagram for explaining the optimization step (4). As shown in the figure, in this optimization step, the improvement rate of the sequence evaluation value Ec of the best similar NDC sequence subjected to the NDC replacement process is checked. If the improvement rate is within 3%, for example, the convergence condition is satisfied. When the final solution is output and the convergence condition is not satisfied, the process returns to the optimization step (3) and the replacement process is repeated.

  In this way, an initial candidate group of NDC sequences is generated based on the similarity of NDC, the best similar NDC sequence having the smallest sequence evaluation value Ec is set as a provisional solution from the initial candidate group, and the provisional solution is improved by replacing NDC. Thus, an optimal NDC sequence can be obtained. Here, although two evaluation indexes of load uniformity and NDC similarity are used, in general, an arbitrary number of evaluation indexes can be used. Also, the improvement rate used for the convergence condition can be a value other than 3%.

  The batch planning unit 120 is a processing unit that generates a resource plan for a batch type primitive service set, and includes a load amount calculation unit 121, a batch type derived service generation unit 122, a batch type DC generation unit 123, and a batch type. An NDC generation unit 124 and a batch-type optimal solution search unit 125 are included.

  In the following description of the batch planning unit 120, it is assumed that the job submission schedule is repeated in a daily cycle and a resource plan for a transaction-type service with a predicted period length of one day is obtained. To do.

  Also, the time required to process one job of a batch type service with one standard performance server is called “turnaround time”, and the length of this time is unique for each batch type service. Determined.

  Also, the load that one server with standard performance can run is called “unit server performance”. The deadline time and turnaround time are unique for each batch-type primitive service. Further, the time difference between the input time of each job of the batch type service and the start time of the prediction period is an integral multiple of the turnaround time of the batch type service.

  The load amount calculation unit 121 is a processing unit that calculates a future load amount. Specifically, the load amount calculation unit 121 calculates a future load amount according to the following procedure. First, the least common multiple of the turnaround time of all batch-type primitive services is obtained, and this value is set as the length of a new time slot.

  Then, for each data center and each measurement cycle, the load amount that can be distributed by the surplus resource amount (not allocated to the transaction type service) that exists in the data center in the measurement cycle is divided by the unit server performance. The surplus resource amount is converted into a representation of the standard performance server number (hereinafter referred to as the surplus server number).

Then, the future load fluctuation of each batch-type primitive service within the prediction period is initialized with zero. Then, for each batch type primitive service, job submission schedule for each batch type service, "at a time t ₁ is turned on the job of ₁ n, charged with ₂ n to t _2, ..., n _k matter to t _k The following (S1) to (S5) are executed for each integer i satisfying 1 ≦ i ≦ k. However, t ₁ > t ₂ >...> t _k .

(S1) The period from the time t _i to the deadline time of the batch type primitive service (this is called a schedule period) is divided into time quantums having the length of the turnaround time of the batch type primitive service.
(S2) A measurement cycle that minimizes the number of surplus servers for each time quantum of each data center is obtained, and the surplus server number at that time is set as the number of surplus servers in the time quantum.

(S3) In each time quantum, the data center having the maximum number of surplus servers is searched, the data center is set as the maximum center in the time quantum, and the number of surplus servers is totaled over all data centers for each time quantum. The total value is set as the R value of the time quantum.

(S4) The time quantum having the maximum R value is searched, the value of the surplus server number of the maximum center in the time quantum is subtracted by 1, and the maximum center and the R value are calculated again for the time quantum. A load amount equal to the unit server performance is added to the load amount corresponding to the time quantum of the future load fluctuation of the batch type primitive service.
(S5) Subtract 1 from n _i, and if n _i > 0, return to (S3).

  The batch type derivation service generation unit 122 is a processing unit that generates a batch type derivation service based on the load amount representing the future load fluctuation calculated by the load amount calculation unit 121. Specifically, the batch-type derived service generation unit 122 determines the time at which the batch-type derived service is generated from the batch-type original service based on the load amount indicating the future load fluctuation for each batch-type original service.

  At this time, an appropriate number of batch-type derived services are generated from the batch-type original service, and the load amount of the batch-type original service is appropriately distributed among the batch-type derived services. Then, for each time slot, a set of batch-type derived services that occur in the time slot is generated.

  The batch type DC generation unit 123 is a processing unit that generates, for each time slot, all candidates (DC) of arrangement combination patterns in the data center group of the batch type derived service set generated in the time slot.

The batch type NDC generation unit 124 is a processing unit that creates an NDC obtained by removing VDC from DC. That is, the batch type NDC generation unit 124 evaluates each DC belonging to the time slot based on the evaluation index m ₁ ,..., M _k for each time slot, and evaluates the evaluation value e ₁ , ..., e _k is calculated.

Then, for each time slot, each DC belonging to the time slot is classified into NDC and VDC based on the evaluation values e ₁ ,..., E _k , and the VDC is excluded from the set of DCs belonging to the time slot.

  The batch type optimal solution search unit 125 is a processing unit that creates a batch type resource plan. That is, the batch type optimal solution search unit 125 generates a Pareto solution set (initial series set) from the set of NDC sequences by the same method as that for the transaction type service. The NDC sequence having the smallest value is selected and set as a sub-optimal NDC sequence.

  Then, the NDC corresponding to each time slot in the sub-optimal NDC sequence is replaced by an NDC replacement operation (similar to the transaction type service), and the values of the sequence evaluation functions of the sub-optimal NDC sequences before and after the NDC replacement operation are compared. If the improvement rate is 3% or less, the sub-optimal NDC sequence is set as a batch type resource plan. Otherwise, the NDC replacement is repeated.

The resource plan transmission unit 130 is a processing unit that transmits the created resource plan to n data centers 10 ₁ to 10 _n connected via the Internet 20. Each data center operates the service according to the resource plan transmitted by the resource plan transmission unit 130.

  Next, the processing procedure of the transaction planning unit 110 shown in FIG. 3 will be described. In the following, an XML-based WEB service is assumed as an example of the transaction-type service, and all the original services are assumed to be WEB services.

In addition, the time axis representing the prediction period that is the target of load fluctuation prediction of the original service is divided into time slots having a fixed period, and each time slot is divided into measurement periods that are periods for measuring the load of the original service. The total number of time slots included in the prediction period is represented as N _t , and the i-th time slot within the prediction period (counted from the earliest time on the time axis) is represented as T _i .

Also, the service load at time t is the number of user requests per second arriving at the service at time t. Here, a set of data centers is C = [c _i | 1 ≦ i ≦ N _c ], a set of primitive services is P = [p _i | 1 ≦ i ≦ N _p ], and occurs within the time slot T _i D (T _i ) = [d (T _i , j) | 1 ≦ j ≦ n (T _i )], and a set of k evaluation indexes for evaluating DC. _{[m 1, m 2, ...} , m k], k -number of a set of k evaluation values of the evaluation of the DC with metric _{[e 1, e 2, ...} , e k], each constituting the NDC series A sum of NDC e _i over the entire NDC sequence is a sequence evaluation value E _i , and a set of NDCs belonging to the time slot T _i is Γ (T _i ) = [γ (T _i , j) | 1 ≦ j ≦ ν ( T _i )], and let S be the set of all NDC sequences.

Further, the NDC belonging to the time slot T _i in NDC constituting a certain NDC series _{s∈S γ s (T i, j} ) written as, also NDCγ (T _i, j) the value of the evaluation value e _i calculated for Write e _i (γ (T _i , j)). Thus, one instance s∈S of the NDC sequence is given by s = <γ _s (T ₁ , j ₁ ), γ _s (T ₂ , j ₂ ), γ _s (T ₃ , j ₃ ) _,. (T _Nt−1 , j _Nt−1 ), γ _s (T _Nt , j _Nt )>.

In addition, p _j ⇒ c _i indicates that the data center c _i is the original center of the primitive service p _j , and the load (number of requests per second) that can be generated using the total resource amount of c _i is R (c _i ).

Here, three indexes m ₁ , m ₂ , and m ₃ defined below are adopted as evaluation indexes for evaluating DC. Here, m ₁ represents the degree of load equality of the entire data center group when the derived service is arranged according to the DC to be evaluated. This is because, in order to minimize the peak required resource amount to be prepared in each data center of the data center group, the load amount should be evenly distributed among all the data centers throughout the prediction period.

M ₂ represents the traffic volume between derived services. When each primitive service is an XML-based WEB service, there may be a case where the primitive services communicate with each other using the SOAP protocol. In such a case, derived services generated from these primitive services also perform SOAP communication.

  Therefore, when each derived service is arranged in the arrangement data center according to a certain DC, each of a plurality of derived services that frequently exchange large amounts of data with each other by SOAP communication is arranged in different data centers. In this case, the amount of communication between data centers is greatly increased.

If the communication delay of the SOAP communication increases due to a shortage of the bandwidth of the communication path, an unnecessary load is applied to each data center as a result. Therefore, the inter-derivative service traffic m ₂ was adopted as the DC evaluation index.

M ₃ represents the degree of similarity of NDC. The definition of the similarity of NDC γ (T _i , j ₁ ) belonging to the time slot T _i is as follows. When a certain NDC sequence is determined, NDC corresponding to the time slot T _i-1 is γ (T _i−1 , j ₀ ) and NDC corresponding to the time slot T _i is γ ( T _i , j ₁ ) and NDC corresponding to time slot T _{i + 1} is γ (T _{i + 1} , j ₂ ), when the time slot transitions from T _{i -1} to T _i as time elapses Includes a derived service set [D (T _i ) as the arrangement of the derived service set in the data center group changes from γ (T _i−1 , j ₀ ) to γ (T _i , j ₁ ). _-1 ) ∩D (T _i )], some derived services may be moved from one data center to another.

Similarly, when the time slot transitions from T _i to T _{i + 1} , the derived service arrangement is changed from γ (T _i , j ₁ ) to γ (T _{i + 1} , j ₂ ). , Several derived services belonging to [D (T _i ) ∩D (T _{i + 1} )] may be moved between data centers.

Therefore, the similarity of NDC γ (T _i , j ₁ ) in the NDC sequence corresponding to time slot T _i moves between data centers for the above reason when the time slot transitions from T _{i -1} to T _i. It is assumed that the number of derived services to be moved and the number of derived services to be moved between data centers when the time slot transitions from T _i to T _{i + 1} are totaled.

  That is, the similarity degree of a certain NDC is large. When a derived service is arranged according to the NDC, the number of times the derived service moves along with the time slot transition increases, and the data center group is moved accordingly. The control load will be applied.

As a result, each data center is burdened with a large amount of load that normally does not need to be processed. Therefore, it is necessary to employ NDC similarity as one of the NDC evaluation indexes. Note that, unlike m ₁ and m ₂ , m ₃ cannot be evaluated by a single NDC, but can be evaluated only when an NDC sequence is designated.

FIG. 13 is a flowchart showing a processing procedure of the transaction planning unit 110 shown in FIG. As shown in the figure, in the transaction planning unit 110, first, the prediction unit 111 performs an ARIMA model prediction or the like for the load fluctuation prediction over the whole prediction period of p _i for all integers i satisfying 1 ≦ i ≦ N _p. Is predicted using the existing prediction method (step S101). The load fluctuation prediction of p _i is expressed as a function of time of the load amount of p _i, load of p _i at time t will be denoted as θ _i (t).

Then, the derived service generation unit 112 generates a derived service based on the load fluctuation prediction and the original center resource amount (step S102). Details of how to generate a derived service based on load fluctuation prediction will be described later. Then, to produce a 1 ≦ i ≦ N _t become set D derived services that are predicted to occur during the T _i for each integer i (T _i) (step S103).

Then, the DC generation unit 113 sets all candidates (DC) of possible arrangement combinations to the data center group C of each derived service belonging to D (T _i ) for each integer i satisfying 1 ≦ i ≦ N _t. A set of arrangement combination candidates generated by the tree search algorithm and generated for the time slot T _i is defined as Γ (T _i ) (step S104).

Then, the NDC generation unit 114 calculates γ (T _i , j) ∈Γ (T for all combinations of each integer _i satisfying 1 ≦ i ≦ N _t and each integer j satisfying 1 ≦ j ≦ ν (T _i ). _i) was evaluated by the evaluation index m _1, m _2, obtaining the evaluation value e _1, e ₂ (step S105). Details of the procedure for calculating e ₁ and e ₂ for γ (T _i , j) will be described later.

Then, for all combinations of each integer i where 1 ≦ i ≦ N _t and each integer j where 1 ≦ j ≦ ν (T _i ), e ₁ (γ (T _i , j))> 0.7 or e ₂ If (γ (T _i , j))> 0.8, γ (T _i , j) is taken as VDC and removed from the set Γ (T _i ). Otherwise, γ (T _i , j) is set as NDC and left in the set Γ (T _i ) (step S106).

Then, the optimum solution search unit 116 defines the NDC sequence set S (step S107), and k sequence evaluation values E ₁ (s), E ₂ (s), E in the NDC sequences included in the NDC sequence set S All NDC sequences s in which at least one of ₃ (s) has the minimum value are searched for and used as elements of the initial sequence set (step S108).

とする。なお、ＮＤＣ系列集合Ｓから初期系列集合(Pareto集合）を抽出する手順の詳細は後述する。 However,

And Details of the procedure for extracting the initial sequence set (Pareto set) from the NDC sequence set S will be described later.

とする。また、K₁=1.0, K₂=1.0, K₃=0.1とした。 Then, the NDC sequence belonging to the initial sequence set is evaluated by the sequence evaluation function ξ (s), and the NDC sequence having the smallest evaluation value is set as a sub-optimal NDC sequence σ (step S109). However, with k = 3

And Further, K ₁ = 1.0, K ₂ = 1.0, and K ₃ = 0.1.

Then, NDC replacement processing is performed for the time slot T _i corresponding to each integer i satisfying 1 ≦ i ≦ N _t (step S110), and the sub-optimal NDC sequence before NDC replacement is replaced by σ ′ and NDC. The improvement rate | ξ (σ ′) − ξ (σ) | / ξ (σ ′) is calculated using σ as the quasi-optimal NDC sequence (step S111). The details of the NDC replacement process will be described later.

  Then, it is determined whether or not the improvement rate is 0.03 or less (step S112). If the improvement rate is 0.03 or less, σ is output as a resource plan for the data center group C, and all The process ends (step S113). Otherwise, σ is set as a new σ ′, and the process returns to step S110.

  In this way, the optimal solution search unit 115 generates an optimal NDC sequence based on the Pareto optimization theory for the derived service generated by the derived service generation unit 112, thereby creating an optimal resource plan for the data center group. be able to.

  Next, details of derivation service generation by the derivation service generation unit 112 will be described. The derived service generation unit 112 performs the following (S1) to (S3) for each measurement period for all measurement periods.

を求め、集合[Θ_i / R(c_i)|1≦i≦N_c ]を標本分布と見なした時の標準偏差を計算し、当該標準偏差が0.1以下であれば（Ｓ３）へ、そうでなければ（Ｓ２）へ行く。 (S1) For the central time t of the measurement cycle and the integer i satisfying 1 ≦ i ≦ N _c , for all _j satisfying p _j ⇒ c _i

And calculate the standard deviation when the set [Θ _i / R (c _i ) | 1 ≦ i ≦ N _c ] is regarded as a sample distribution. If the standard deviation is 0.1 or less, go to (S3). Otherwise, go to (S2).

(S2) For each integer i satisfying 1 ≦ i ≦ N _c , a part of the load amount at time t of p _i is divided as a derived service. Here, how to divide the load amount into the derived service is as follows.

Assuming that N _c −1 derived services are generated from each primitive service, a temporary center arrangement of each derived service is determined. The distribution ratio between the assumed service load and the source service load is obtained as follows.

The above distribution ratio is optimized using linear programming so that the following conditions are satisfied. When each derived service is arranged in the above temporary center arrangement, for each data center c _i (1 ≦ i ≦ N _c ), the total load amount of the source service and the derived service operated there is defined as g _i The standard deviation of the sample set [g _i / R (c _i ) | 1 ≦ i ≦ N _c ] is set to be 0.1 or less.

  If the division ratio that satisfies the above conditions cannot be obtained by integer programming, the provisional center layout of each derived service should be changed at random and the division ratio optimized by integer programming should be performed again. By repeating, the division ratio of the load amount to be divided into derived services for each primitive service is obtained.

However, even if N _c > 5, the upper limit of the number of derived services that can be generated simultaneously from one primitive service is four. When the distribution ratio determined by the above method for a derived service is 0, the derived service is treated as not existing.
(S3) Move to the next measurement cycle and start (S1) for a new measurement cycle.

Next, details of the procedure for calculating e ₁ for γ (T _i , j) will be described. A set of center times of the measurement periods included in the time slot T _i is defined as [t ⁱ _h | 1 ≦ h ≦ N _m ]. That is, it is assumed that the central time of the h-th measurement cycle from the start time of the time slot T _i is t ⁱ _h, and N _m measurement cycles are included in one time slot.

Assuming that the total load amount at time t of all services arranged in the data center _cr (1 ≦ r ≦ N _c ) is g _r (t), the value of e _i for γ (T _i , j) is It is obtained in this way.

を求める。 First, time series ^{_{[g r (t i h)}} | 1 ≦ h ≦ N m] of the center each load sum value for each integer r which is a 1 ≦ r ≦ N _c seek, is 1 ≦ r ≦ N _c Average value on time axis of time series [g _r (t ⁱ _h ) / R (c _r ) | 1 ≦ h ≦ N _m ] for each integer r

Ask for.

Then, the standard deviation when the set [μ (0, i), μ (1, i), μ (2, i),…, μ (N _c , i)] is regarded as the sample distribution is calculated, Let this standard deviation be the value of e _i for γ (T _i , j).

Next, details of the procedure for calculating e ₂ for γ (T _i , j) will be described. First, for each data center for each measurement cycle in the time slot T to which the evaluation target NDC belongs, the communication amount by SOAP communication between services is estimated in units of bps.

  Then, for each data center, a value (hereinafter referred to as a communication path occupation ratio) obtained by dividing the estimated communication amount by the bandwidth value (in bps) of the communication path connecting the data center and the external network is obtained.

And the value which averaged the communication channel occupation rate for every measurement period over the whole time slot T is calculated about each data center. The average value of the data center to which the average value is maximized for a value of e _2.

An example of a procedure for estimating the amount of SOAP communication between services when the measurement cycle in the data center c _i is t when the relationship between primitive services communicating with each other is as shown in FIG. Is shown below.

Here, a set of derived services arrangement destination data center is designated in the NDC that is evaluated by _{e 2 D = [d k |} 1 ≦ k ≦ N d], the placement destination data center d _k that it is a c _i expressed as d _k ⇒ c _i like. A value representing the load amount of the service X by the number of requests per second is represented as L (X, t). In addition, p _j ↑ d _k or the like indicates that the primitive service p _j is a primitive service from which d _k is derived.

First, the value B ₁ (c _i , t) at the data center c _i at the measurement cycle t is obtained by estimating the communication amount between the first stage and the second stage in FIG. First, B ₁ (c _i , t) is initialized to 0, and the following four cases are considered.

(1) When p ₁ ⇒ c _i and p ₂ ⇒ c _i
Let B ₁ (c _i , t) = B ₁ (c _i , t) + | L (p ₁ , t) × ω ₁ −L (p ₂ , t) |.

(2) When p ₁ ⇒ c _i and not p ₂ ⇒ c _i
Let H be the set of all subscripts h such that (p ₂ ↑ d _h ) ∧ (d _h ⇒c _i )

(3) When p ₂ ⇒ c _i and not p ₁ ⇒ c _i
Let H be the set of all subscripts h such that (p ₁ ↑ d _h ) ∧ (d _h ⇒c _i )

(4) When p ₁ ⇒ c _i and not p ₂ ⇒ c _i
Let H be the set of all subscripts h such that (p ₁ ↑ d _h ) ∧ (d _h ⇒c _i ), and the set of all subscripts _{k such} that (p ₂ ↑ d _k ) ∧ (d _k ⇒c _i ) Let K be K

Next, the value B ₂ (c _i , t) at the data center c _i at the measurement cycle t is obtained by estimating the communication volume between the second and third stages in FIG. First, B ₁ (c _i , t) is initialized to 0, and the following four cases are considered.

(1) When p ₂ ⇒ c _i and p ₃ ⇒ c _i
Let B ₂ (c _i , t) = B ₂ (c _i , t) + | L (p ₂ , t) × ω ₂ −L (p ₃ , t) |.

(2) When p ₂ ⇒ c _i and not p ₃ ⇒ c _i
Let H be the set of all subscripts h such that (p ₃ ↑ d _h ) ∧ (d _h ⇒c _i )

(3) When p ₃ ⇒ c _i and not p ₂ ⇒ c _i
Let H be the set of all subscripts h such that (p ₂ ↑ d _h ) ∧ (d _h ⇒c _i )

(4) When p ₂ ⇒ c _i and not p ₃ ⇒ c _i
Let H be the set of all subscripts h such that (p ₂ ↑ d _h ) ∧ (d _h ⇒c _i ), and the set of all subscripts _{k such} that (p ₃ ↑ d _k ) ∧ (d _k ⇒c _i ) Let K be K

Here, by assuming that ω (c _i , t) = ω ₁ (c _i , t) + ω ₂ (c _i , t), the estimated communication amount ω (c at the data center c _i at the measurement period t _i , t) is obtained.

Next, details of a procedure for extracting an initial sequence set (Pareto set) from the NDC sequence set S will be described. First, sεS that minimizes the series evaluation value E ₁ (s) is obtained. There are generally a plurality of such s.

Specifically, for each integer i satisfying 1 ≦ i ≦ N _t , all NDCs belonging to Γ (T _i ) are stored in the array, and the array is sorted using the value of e ₁ as a key, and Γ (T the value of e ₁ in _i) to produce the smallest set of NDC Γ ^b (T _i). Then, s∈S that minimizes the sequence evaluation value E ₁ (s) of each NDC sequence belonging to the Cartesian product set Γ ^b (T ₀ ) × Γ ^b (T ₁ ) ×... × Γ ^b (T _Nt ) To do.

Next, sεS that minimizes the series evaluation value E ₂ (s) is obtained. That is, for each integer i satisfying 1 ≦ i ≦ N _t , all the NDCs belonging to Γ (T _i ) are stored in the array, the array is sorted using the value of e ₂ as a key, and Γ (T _i ) To generate an NDC set Γ ^b (T _i ) that minimizes the value of e ₂ . Further, each NDC sequence belonging to the Cartesian product set Γ ^b (T ₀ ) × Γ ^b (T ₁ ) ×... × Γ ^b (T _Nt ) is set to s∈S such that the sequence evaluation value E ₂ (s) is minimized. To do.

Next, sεS that minimizes the series evaluation value E ₃ (s) is obtained. Here, a set Γ of NDCs over the entire time axis is defined as follows.

Then, the following processes are performed for all combinations of each integer i to be 1 ≦ k ≦ N _A to become the each integer k 1 ≦ i ≦ N _t. Search for NDCs in the set Γ (T _i ), and γ (T _i , j) ∈Γ (T _i ), 1 ≦ j ≦ ν (T _i ), and γ _k and γ (T _i , j) The value of the subscript j ^k _i that minimizes the relative similarity when compared is obtained.

  Here, the relative similarity refers to an arrangement specified by one NDC among the derived services that are commonly included in both derived service arrangements designated by the two NDCs when two NDCs are compared. This represents the number of derived services in which the destination data center and the placement destination data center designated by the other NDC are different.

Then, 1 ≦ k ≦ N NDC sequence s expressed as follows composed for each integer _{k A (k) = <γ} (T 0, j k 0), γ (T 1, j k 1), ... , γ (T _Nt , j ^k _Nt )>.

Finally, the NDC sequence s (k) included in the set [s (k) | 1 ≦ k ≦ N _A ] is selected so as to minimize the value of E ₂ (s (k)).

As described above, all NDC sequences in which _{one of} E ₁ (s), E ₂ (s), and E ₃ (s) has the minimum value are selected from S, and the Pareto set (initial sequence set) is selected. Element.

Next, details of the NDC replacement process will be described. As the NDC replacement processing, the following (S1) to (S3) are performed.
(S1) The following relationship is established between γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) (where j ₁ ≠ j ₂ ) by searching the set Γ (T _i ). seek j ₂ such as.
e ₁ (γσ (T _i , j ₁ ))> e ₁ (γ (T _i , j ₂ ))
e ₂ (γσ (T _i , j ₁ ))> e ₂ (γ (T _i , j ₂ ))
e ₃ (γσ (T _i , j ₁ ))> e ₃ (γ (T _i , j ₂ ))

If j ₂ satisfying the above conditions is found, γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) are interchanged to change γ (T _i , j ₂ ) to the time within the NDC sequence σ. The NDC corresponding to the slot T _i is set, and the process proceeds to (S3). When j ₂ satisfying the above conditions is not found, the process proceeds to (S2).

(S2) Searching within the set Γ (T _i ), the following relationship holds between γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) (where j ₁ ≠ j ₂ ) seek j ₂ such as.

Here, I1 and I2 are subsets obtained by dividing the integer set [h | 1 ≦ h ≦ 3] into two having no common part, and e _h (γσ (T _i , j ₁ )) ≦ e when h ∈ I1 _h (γ (T _i , j ₂ )), and when h ∈ I2, e _h (γσ (T _i , j ₁ ))> e _h (γ (T _i , j ₂ )) and Δe _h = e _h (γσ (T _i , j ₁ )) − e _h (γ (T _i , j ₂ )), δe _h = e _h (γ (T _i , j ₂ )) − e _h (γσ (T _i , j ₁ )).

If j ₂ satisfying the above conditions is found, γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) are interchanged to change γ (T _i , j ₂ ) to the time within the NDC sequence σ. It is assumed that the NDC corresponds to the slot T _i . If j ₂ satisfying the above condition is not found, NDC replacement is not performed for time slot T _i . Here, the definitions of λ ₁ , λ ₂ , λ ₃ , Λ ₁ , Λ ₂ , Λ ₃ and the values of φ ₁ , φ ₂ , φ ₃ are as follows.

λ ₁ (x) = x λ ₂ (x) = x λ ₃ (x) = 0.1 × exp (1.3x)
Λ ₁ (x) = x Λ ₂ (x) = x Λ ₃ (x) = 0.1 × exp (1.3x)
φ ₁ = 1.0 φ ₂ = 1.0 φ ₃ = 1.0
Here, exp () is an exponential function with the base of the natural logarithm.

(S3) As i = i + 1, the process moves to the next time slot and returns to (S1).

  Next, the processing procedure of the batch planning unit 120 shown in FIG. 3 will be described. As an example of the batch-type service, a calculation service for collecting sales information of each store of a convenience store can be considered.

Hereinafter, in this case, a set of batch type primitive service _{P = [p i | 1 ≦} i ≦ N p], the turnaround time of each job p _i tau _i, the data center c _i, is located c _i R _max (c _i ) represents the required resource amount of ci in the number of requests per second when the total load amount of all services is at a peak, and the WEB service at the time t among the resources of the data center ci R (c _i , t) is the surplus computational resource amount that is not allocated to the server, the value of the unit server performance is f, and the definitions of other symbols are used to explain the processing procedure of the transaction planning unit 110 The same shall apply.

Further, here, the time difference between the start time of the closing time and the predicted period of each job p _i is set to an integer multiple of tau _i, tau _i is assumed to be an integer multiple of the measurement cycle loads transactional services .

Also, in generating a resource plan for a batch-type service, only m ₁ and m ₃ are adopted as evaluation indexes used for generating a resource plan for a WEB service. This is because it is not normally possible for batch services to perform SOAP communication, and m _2, which is an evaluation index related to the amount of SOAP communication between services, is not meaningful.

FIG. 15 is a flowchart showing a processing procedure of the batch planning unit 120 shown in FIG. As shown in the figure, in the batch planning unit 120, first, the load amount calculation unit 121 obtains the least common multiple of τ ₁ , τ ₂ ,..., Τ _Np and sets the value as the length of the time slot (step S201).

Then, the time axis is divided into time slots, and for each integer i satisfying 1 ≦ i ≦ N _c , η (c _i , t) = R (c _i , t) / f is set, and a job scheduling algorithm described later Thus, future load fluctuations are planned for each batch-type primitive service p _i (1 ≦ i ≦ N _p ) (step S202). The future load fluctuation related to p _i is generated as time-series data, and the load amount of p _i at the future time t is expressed as θ (i, t).

  Then, the batch type derived service generation unit 122 generates a batch type derived service based on the load amount calculated by the load amount calculation unit 121 (step S203). Details of the generation of the batch type derived service by the batch type derived service generation unit 122 will be described later.

Then, 1 for ≦ i ≦ N _t become each integer i, to generate a set D of predicted derived service to occur during the T _i (T _i) (step S204).

Then, the batch type DC generation unit 123, for each integer i satisfying 1 ≦ i ≦ N _t , all candidates (DC) of possible arrangement combinations to the data center group C of each derived service belonging to D (T _i ) ) Is generated by the tree search algorithm (step S205), and a set of arrangement combination candidates generated for the time slot T _i is defined as Γ (T _i ).

Then, the batch type NDC generation unit 124 calculates γ (T _i , j) ∈Γ for all combinations of each integer _i satisfying 1 ≦ i ≦ N _t and each integer j satisfying 1 ≦ j ≦ ν (T _i ). (T _i) was evaluated by the evaluation index m ₁ to obtain an evaluation value e ₁ (step S206).

Then, for all combinations of each integer i satisfying 1 ≦ i ≦ N _t and each integer j satisfying 1 ≦ j ≦ ν (T _i ), if e ₁ (γ (T _i , j))> 0.7, then γ Let (T _i , j) be a VDC and remove it from the set Γ (T _i ). Otherwise, γ (T _i , j) is set as NDC and left in the set Γ (T _i ) (step S207).

Then, the batch-type optimum solution searching unit 125 creates an NDC sequence set S (step S208), and k sequence evaluation values E ₁ (s) and E ₃ (s) among the NDC sequences included in the NDC sequence set S. All NDC sequences s having one or more of the minimum values are searched for and used as elements of the initial sequence set (step S209).

とする。 However,

And

Then, the NDC sequence belonging to the initial sequence set is evaluated by the sequence evaluation function ξ (s), and the NDC sequence having the smallest evaluation value is set as a sub-optimal NDC sequence σ (step S210). However, ξ (s) = K ₁ E ₁ (s) + K ₂ E ₂ (s). Further, K ₁ = 1.0 and K ₃ = 0.1.

Then, NDC replacement processing described later is performed for the time slot T _i corresponding to each integer i satisfying 1 ≦ i ≦ N _t (step S211). Then, the improvement rate is calculated using the evaluation value at ξ () of the sub-optimal NDC sequence before NDC replacement and the evaluation value at ξ () of the sub-optimal NDC sequence after NDC replacement (step S212). ).

  Then, it is determined whether or not the improvement rate is 3% or less (step S213). If the improvement rate is 3% or less, the current sub-optimal NDC sequence is output as a resource plan (step S214), and all the processes are finished. . Otherwise, the process returns to step S211.

Next, a job scheduling algorithm for planning future load fluctuations of the batch type primitive service will be described. When the job input schedule, turnaround time, and deadline time of each batch type primitive service p _i (1 ≦ i ≦ N _p ) are as shown in FIG. 16, each batch type primitive service p _i (1 ≦ i ≦ N _p) ), The following (S1) to (S9) are executed.

However, if the start time of the prediction period is t ₀ , t (i, j) −t ₀ = TA (i) × N, TD (i) −t0 = TA (i) × N for any i and j '(Where N and N' are integers), and TA (i) is an integral multiple of the length of the measurement cycle. The unit server performance is assumed to be f.

(S1) j = ρ (i). Also, θ _i (t) = 0 at each time t within the prediction period.
(S2) η (c when R (c _k , t) is the surplus resource amount in the data center _ck that is not assigned to any transaction-type or batch-type service at each time t in the forecast period _k , t) = R (c _k , t) / f is obtained.

(S3) The time period from time t (i, j) to time TD (i) is divided into time quantum lengths of TA (i), and the start time from time t (i, j) in ascending order. The h-th time quantum is counted as Q (i, j, h). N _Q (i, j) = (TD (i) −t (i, j)) / TA (i).

(S4) The following processing is performed for combinations of each integer h satisfying 1 ≦ h ≦ N _Q (i, j) and each integer k satisfying 1 ≦ k ≦ N _c . A measurement cycle that minimizes the value of η (c _k , t) at the center time t of the measurement cycle among a plurality of measurement cycles included in the time quantum Q (i, j, h) is obtained. The value of η (c _k , t) at the central time t is defined as the surplus resource amount η (Q (i, j, h), c _k ) in Q (i, j, h).

を計算する。 (S5) The following processing is performed for each integer h satisfying 1 ≦ h ≦ N _Q (i, j). Find the value of the index k such that η (Q (i, j, h), c _k ) is maximum, and let this be k (i, j, h)

Calculate

(S6) The value of the index h that maximizes the value of η (i, j, h) is obtained, 1 is subtracted from η (c _{k (i, j, h)} , t), and R (c _k Subtract f from _{(i, j, h)} , t).
(S7) θ _i (t) = θ _i for each time t where t (i, j) + TA (i) × h ≦ t ≦ t (i, j) + TA (i) × (h + 1) (t) + f.

(S8) Subtract 1 from n (i, j) and if n (i, j)> 0, return to (S5).
(S9) If j is 1, the process is terminated. Otherwise, j = j−1 and return to (S2).
The value of θ _i (t) obtained as a result of the above procedure execution is the future load at the future time t of the batch-type primitive service p _i .

Next, details of batch type derivation service generation by the batch type derivation service generation unit 122 will be described. The batch-type derived service generation unit 122 executes the following (S1) and (S2) for each integer i satisfying 1 ≦ i ≦ _Nc . Here, the total load amount by the WEB service of the data center c _i at time t is expressed as g (c _i , t).

ならば、p_h ⇒ c_iとなるp_hの時刻ｔでの負荷量θ'(h,t)を

とする。 (S1) The following processing is executed in each measurement cycle within the prediction period. When J is the set of all subscripts j that satisfy p _j ⇒ c _{i at} the center time t of the measurement cycle

If, load amount in the p _h ⇒ c _i to become p _h of time t θ 'the (h, t)

And

である。 However,

It is.

ならば、p_h⇒c_iとなる各バッチ型原始サービスp_hから時刻ｔにおいてN_c-1個のバッチ型派生サービスを発生させ、当該各バッチ型派生サービスの時刻ｔでの負荷量を以下のようにする。

(S2) The following processing is executed in each measurement cycle within the prediction period. When t is the central time of the measurement cycle, J is the set of all subscripts j that satisfy p _j ⇒ c _i

Then, N _c −1 batch-type derived services are generated at time t from each batch-type primitive service p _{h where} p _h ⇒c _i, and the load amount at time t of each batch-type derived service is as follows: Like this.

Next, details of the NDC replacement process will be described. As the NDC replacement process, the following (S1) to (S3) are performed.
(S1) The following relationship is established between γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) (where j ₁ ≠ j ₂ ) by searching the set Γ (T _i ). seek j ₂ such as.
e ₁ (γσ (T _i , j ₁ ))> e ₁ (γ (T _i , j ₂ ))
e ₃ (γσ (T _i , j ₁ ))> e ₃ (γ (T _i , j ₂ ))

If j ₂ satisfying the above conditions is found, γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) are interchanged to change γ (T _i , j ₂ ) to the time within the NDC sequence σ. The NDC corresponding to the slot T _i is set, and the process proceeds to (S3). When j ₂ satisfying the above conditions is not found, the process proceeds to (S2).

(S2) The set Γ (T _i ) is searched, and the following equation (S2) is set between γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) (where j ₁ ≠ j ₂ ). ) Or formula (S3), j ₂ is determined so that one of the relations is established.

However, λ (x) = 0.1 × exp (1.3 × x). If j ₂ satisfying the above conditions is found, γσ (T _i , j ₁ ) and γ (T _i , j ₂ ) are interchanged to change γ (T _i , j ₂ ) to the time within the NDC sequence σ. It is assumed that the NDC corresponds to the slot T _i . If j ₂ satisfying the above condition is not found, NDC replacement is not performed for time slot T _i .
(S3) As i = i + 1, the process moves to the next time slot and returns to (S1).

  Next, a simulation result of resource plan creation by the resource plan creation device 100 according to the present embodiment will be described. FIG. 17 is a diagram illustrating a simulation result of resource plan creation by the resource plan creation device 100 according to the present embodiment.

  As shown in the figure, the Consumer traffic of the data center X is divided into three at high load in the morning and night, and the processing is also performed in the data center Y and the data center Z. Further, the store traffic of the data center Z is divided into three at the peak of morning and evening, and processing is also performed in the data center X and the data center Y.

  As described above, in this embodiment, the prediction unit 111 of the transaction planning unit 110 performs the load prediction of the source service for a plurality of transaction type services operated by a plurality of data centers, and the derived service generation unit 112. Generates a derived service based on the load prediction by the prediction unit 111 and the calculation resource amount of the data center, and the optimum solution search unit 115 creates an optimal resource plan for the derived service generated by the derived service generation unit 112.

  In addition, for a plurality of batch type services operated by a plurality of data centers, the load amount calculation unit 121 predicts the load of the original service, and the batch type derived service generation unit 122 of the batch planning unit 120 sets the load amount calculation unit. The derived service is generated based on the amount of computational resources not used for the load prediction and transaction type service processing by 121, and the batch type optimal solution search unit 125 determines the optimal resource plan for the derived service generated by the derived service generation unit 122. create.

  Therefore, since it is possible to perform the optimum load allocation corresponding to the calculation resource for each data center, it is possible to prevent the data center from being overloaded and to efficiently use the data center calculation resource.

  In addition, the resource plan generation process determines the load distribution ratio between each derived service and the originating source service, the time at which each derived service occurred, and the location data center at that time. For each center, a temporal variation of the total amount of the original service load and the derived service load within the prediction period is obtained.

  Based on this, the peak load amount of the data center can be estimated for each data center when the derived service is appropriately generated according to the load fluctuation. It can be set to a value smaller than the peak load amount of each data center when no service is generated or when each derived service is improperly arranged.

  Further, in the present embodiment, a case has been described in which only the resource plan creation device 100 performs load prediction, creation of an optimal resource plan, and transmission of the resource plan to the data center, but the present invention is not limited to this. For example, the same can be applied to the case where the load prediction unit 111 is performed by another device.

  In the present embodiment, the resource plan creation device has been described. However, by realizing the configuration of the resource plan creation device with software, a resource plan creation program having the same function can be obtained. Therefore, a computer system that executes this resource plan creation program will be described.

  FIG. 18 is a diagram illustrating a computer system that executes a resource plan creation program according to the present embodiment. As shown in the figure, the computer system 200 includes a main body 201, a display 202 that displays information on a display screen 202a according to an instruction from the main body 201, and various information input to the computer system 200. It has a keyboard 203, a mouse 204 for designating an arbitrary position on the display screen 202a of the display 202, a LAN interface connected to the LAN 206 or a wide area network (WAN), and a modem connected to the public line 207. Here, the LAN 206 connects the computer system 200 to another computer system (PC) 211, a server 212, a printer 213, and the like.

  FIG. 19 is a functional block diagram showing the configuration of the main body 201 shown in FIG. As shown in the figure, the main body 201 includes a CPU 221, a RAM 222, a ROM 223, a hard disk drive (HDD) 224, a CD-ROM drive 225, an FD drive 226, an I / O interface 227, and a LAN. An interface 228 and a modem 229 are included.

  The resource plan creation program executed in the computer system 200 is stored in a portable storage medium such as a floppy disk (FD) 208, a CD-ROM 209, a DVD disk, a magneto-optical disk, an IC card, and the like. And installed in the computer system 200.

  Alternatively, the resource plan creation program is stored in a database of the server 212 connected via the LAN interface 228, a database of another computer system (PC) 211, and the like, read from these databases, and stored in the computer system 200. Installed.

  The installed resource plan creation program is stored in the HDD 224 and is executed by the CPU 221 using the RAM 222, the ROM 223, and the like.

  In this embodiment, the case where the service is arranged in the data center has been described. However, the present invention is not limited to this, and the present invention can be similarly applied to the case where the service is arranged in another computer cluster. it can.

  For example, there is a case where a set of computing resources constituting the inside of each data center is divided into a plurality of sections, and independent operation management is performed for each section. In this case, for each partition, optimize which derived service is placed when, and which primitive service is to be operated in which partition. It is also possible to determine the future time point to be generated.

(Supplementary note 1) A resource plan creation program for creating a resource plan for allocating a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters. And
Derived service generation procedure for generating a derived service in which a part of the operation is transferred from each computer cluster to another computer cluster based on the load prediction of the original service from the original service which is a service in which the operated computer cluster is determined; ,
A plan creation procedure for creating a resource plan for allocating the derived service group generated by the derived service generation procedure to the plurality of computer clusters based on a predetermined evaluation index;
A resource plan creation program characterized by causing a computer to execute.

(Supplementary Note 2) The derived service generation procedure generates a derived service for each time slot obtained by dividing the prediction period of the load prediction into a plurality of time slots at a predetermined time interval.
The supplementary plan 1 is characterized in that, in the plan creation procedure, an arrangement plan sequence in which time slot arrangement plans, which are derived service arrangement plans for each time slot, are arranged in the time order of the time slots is created as the resource plan. Resource planning program.

(Supplementary note 3) The resource plan creation program according to supplementary note 2, wherein the plan creation procedure creates a resource plan for optimally allocating a derived service group to the plurality of computer clusters based on Pareto optimality theory. .

(Additional remark 4) The said plan preparation procedure uses the load equality between computer clusters, and the similarity of the said time slot arrangement | positioning plan between the time slots adjacent on a time axis as an evaluation parameter | index, It is characterized by the above-mentioned. Resource planning program described in 1.

(Additional remark 5) The said plan preparation procedure makes the arrangement plan series with the best evaluation when the similarity and the load uniformity are used as evaluation indexes as an initial series set, and for the arrangement plan series included in the initial series set The resource plan creation program according to appendix 4, wherein the resource plan is created by performing correction so that the evaluation of the load uniformity is improved.

(Additional remark 6) The resource plan as described in any one of additional remarks 1-5 which makes a computer further perform the resource plan transmission procedure which transmits the resource plan produced by the said plan creation procedure to each computer cluster Creation program.

(Supplementary note 7) The computer cluster is a data center, and the computer cluster system is a system in which a plurality of data centers are connected via a network. Resource planning program.

(Supplementary Note 8) The derived service generation procedure is as follows.
A non-real-time batch service characterized in that each job can be processed anytime and at any time if all the jobs that make up the service have been processed by the specified deadline time. When a cluster is operated as a set of computer clusters, transactions are performed in order to preferentially allocate computer cluster resources to transaction-type services that require real-time performance and to allocate remaining surplus resources to batch-type services. The resource plan creation program according to appendix 1, wherein a resource plan is generated for the type service, and then a resource plan for the batch type service is generated for the surplus resource that has not been assigned.

(Supplementary note 9) The plan creation procedure includes all jobs that make up each batch-type primitive service as targets of execution plans in order from the one whose input time is close to the deadline time, and the surplus that is not used by any service The resource plan creation program according to appendix 1, wherein the execution plan of the job to be scheduled is generated in a period in which the total sum of the amount of computing resources over the entire computer cluster set is maximum.

(Supplementary note 10) Records a resource plan creation program for creating a resource plan for allocating a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters A computer-readable recording medium,
Derived service generation procedure for generating a derived service in which the operation is transferred from each computer cluster to another computer cluster based on the load prediction of the original service from the original service which is a service in which the operated computer cluster is determined;
A plan creation procedure for creating a resource plan for allocating the derived service generated by the derived service generation procedure to the plurality of computer clusters based on a predetermined evaluation index;
A computer-readable recording medium in which a resource plan creation program for causing a computer to execute is recorded.

(Supplementary Note 11) A resource plan creation method for creating a resource plan for allocating a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters. And
A derived service generation step of generating a derived service in which a part of the operation is transferred from each computer cluster to another computer cluster based on a load prediction of the original service from the original service which is a service in which the operated computer cluster is determined; ,
A plan creation step of creating a resource plan that optimally arranges the derived service group generated by the derived service generation step in the plurality of computer clusters based on a predetermined evaluation index;
A resource planning method characterized by including

(Supplementary note 12) A resource plan creation device for creating a resource plan for allocating a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters. And
Derived service generation means for generating a derived service in which a part of the operation is transferred from each computer cluster to another computer cluster based on the load prediction of the original service from the original service which is a service in which the operated computer cluster is determined ,
Plan creation means for creating a resource plan for optimally arranging the derived service group generated by the derived service generation means in the plurality of computer clusters, based on a predetermined evaluation index;
A resource plan creation device characterized by comprising:

  As described above, the resource plan creation program according to the present invention is useful for allocating services to a plurality of data centers, and is particularly suitable for data centers that need to eliminate overload and effectively use computing resources. .

It is explanatory drawing for demonstrating the concept of the resource plan which concerns on a present Example. It is explanatory drawing for demonstrating derivative service generation | occurrence | production and arrangement | positioning optimization. It is explanatory drawing for demonstrating arrangement | positioning optimization of the derivative service for every time slot. It is a functional block diagram which shows the structure of the resource plan preparation apparatus concerning a present Example. It is explanatory drawing (1) for demonstrating the generation conditions of a derivative service. It is explanatory drawing (2) for demonstrating the generation conditions of a derivative service. It is explanatory drawing for demonstrating the equal degree of load. It is explanatory drawing for demonstrating the similarity of NDC. It is explanatory drawing for demonstrating the optimization procedure outline. It is explanatory drawing for demonstrating the concept of a NDC table. It is explanatory drawing for demonstrating an optimization step (1). It is explanatory drawing for demonstrating an optimization step (2). It is explanatory drawing for demonstrating an optimization step (3). It is a figure which shows replacement conditions. It is explanatory drawing for demonstrating an optimization step (4). It is a flowchart which shows the process sequence of the transaction plan part shown in FIG. It is a figure which shows the example of the service set which carries out SOAP communication mutually. It is a flowchart which shows the process sequence of the batch plan part shown in FIG. It is a figure which shows the example of the job input schedule of a batch type primitive service. It is a figure which shows the simulation result of the resource plan preparation by the resource plan preparation apparatus 100 which concerns on a present Example. It is a figure which shows the computer system which performs the resource plan preparation program concerning a present Example. It is a functional block diagram which shows the structure of the main-body part shown in FIG.

Explanation of symbols

10 ₁ to 10 _n Data Center 20 Internet 100 Resource Plan Creation Device 110 Transaction Planning Unit 111 Prediction Unit 112 Derived Service Generation Unit 113 DC Generation Unit 114 NDC Generation Unit 115 Optimal Solution Search Unit 120 Batch Planning Unit 121 Load Amount Calculation Unit 122 Batch Type derivation service generation unit 123 Batch type DC generation unit 124 Batch type NDC generation unit 125 Batch type optimal solution search unit 130 Resource plan transmission unit 200, 211 Computer system 201 Main unit 202 Display 202a Display screen 203 Keyboard 204 Mouse 206 LAN
207 Public line 208 Floppy disk 209 CD-ROM
212 Server 213 Printer 221 CPU
222 RAM
223 ROM
224 Hard disk drive 225 CD-ROM drive 226 Floppy disk drive 227 I / O interface 228 LAN interface 229 Modem

Claims (2)

A resource plan creation program for creating a resource plan for allocating a plurality of services operated by a computer cluster system configured by connecting a plurality of computer clusters composed of a plurality of computers to the plurality of computer clusters,
A derived service in which a part of the operation is transferred from each computer cluster to another computer cluster based on the load prediction of the source service from the source service, which is a service for which the computer cluster to be operated is determined, is used for the prediction period of the load prediction. Derived service generation procedure for generating each time slot divided into a plurality of time slots at a predetermined time interval ;
As a resource plan for optimally allocating the derived service group generated by the derived service generation procedure to the plurality of computer clusters , time slot allocation plans that are derived service allocation plans for each time slot are arranged in the order of the times of the time slots. A plan creation procedure for creating an arrangement plan series using load equality between computer clusters and similarity of time slot arrangement plans between adjacent time slots on a time axis as an evaluation index ;
A resource plan creation program characterized by causing a computer to execute. In the plan creation procedure, an arrangement plan sequence that has the best evaluation when the similarity and load uniformity are used as evaluation indexes is set as an initial sequence set, and the load equality is calculated with respect to the arrangement plan sequence included in the initial sequence set. The resource plan creation program according to claim 1, wherein the resource plan is created by performing correction so as to improve the evaluation of the resource plan.

Images