JP2015108878A - Allocation determination device, control method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology for more efficiently using a processing server in a distributing system.SOLUTION: An allocation determination device 2000 includes: a resource featured value acquisition part 2020; a prediction time distribution acquisition part 2040; a statistical amount calculation part 2060; and an allocation determination part 2080. The resource featured value acquisition part 2020 is configured to acquire a resource featured value related to a candidate processing server as the candidate of a processing server 3000 to which a task is allocated. The prediction time distribution acquisition part 2040 is configured to acquire prediction time distribution corresponding to the resource featured value acquired by the resource featured value acquisition part 2020. The statistical amount calculation part 2060 is configured to calculate the statistical amount of the prediction time distribution acquired by the prediction time distribution acquisition part 2040. The allocation determination part 2080 is configured to determine the allocation of the task to the candidate processing server on the basis of the statistical amount calculated by the statistical amount calculation part 2060.

Description

  The present invention relates to an assignment determination device, a control method, and a program.

  A distributed system has been developed in which a plurality of tasks generated by dividing a job are distributed to a plurality of computers (hereinafter referred to as processing servers) and executed. When the processing server is ready to execute a task, it receives a new task assignment. In such a distributed system, new tasks are assigned in order from the processing server where the tasks have been completed, and therefore, it becomes easier to assign more tasks to a high-speed processing server. As a result, job execution efficiency increases.

  If there are a plurality of processing servers to which a new task is assigned, the job completion time can be shortened if a server that can process the task in a shorter time than the other processing servers can be selected. In order to select a server that can be processed in a short processing time, it is necessary to predict the task processing time for each processing server.

  Non-patent document 1 and patent documents 1 to 6 are prior art documents disclosing prediction of task processing time or technologies related thereto. Non-Patent Document 1 discloses a technique for shortening the time required for processing a task being executed. Specifically, first, the time required for processing a task being executed on the processing server is calculated for two cases: 1) when the task is continued to be executed, and 2) when the task is re-executed. . If the time required for re-execution is shorter, the task is re-executed.

  Patent Documents 1 and 2 disclose a technique for statically calculating a task allocation plan based on a task load index and a processing server performance index measured in advance.

  Patent Document 3 estimates the completion time from the progress status of the job, and if the job request completion time is not satisfied, the job is divided into a plurality of tasks, and the divided tasks are allocated to a plurality of processing servers. A technique for realizing the above is disclosed.

  Patent Document 4 discloses a technique for assigning a task to a high-performance processing server based on a static performance index value calculated based on the configuration of the processing server.

  Patent Document 5 discloses a technique for predicting whether processing is completed within a time when resources are secured, and interrupting processing when the processing is not completed within the time.

  Patent Document 6 discloses a technique for estimating a data segment reading time in a file system including a plurality of storage devices and selecting a storage device to be read based on an estimation of which storage device is the shortest.

JP 2008-243216 A JP 2010-277604 A JP 2008-123205 A JP 2007-317038 A JP 2010-152738 A JP 2009-59357 A

Ganesh Ananthanarayanan, 6 others, “Reining in the Outliers in Map-Reduce Clusters using Mantri”, Proceedings of the 9th USENIX Symposium on Operating Systems Design and Implementation (OSDI '10), October 4, 2010

  The present inventor has studied a method of using a processing server more efficiently in a distributed system. Among them, the present inventor has paid attention to the fact that the task processing time varies even when the resource usage rates of the processing servers are the same or close to each other. For example, when the resource usage rate of the processing server is high, task processing tends to take a long time, but task processing may be completed in a short time. That is, just because a processing server has a high resource usage rate, the processing server does not necessarily take a long time to process a task. None of the techniques described in the above-mentioned prior art documents considers that “the task processing time varies even when the resource usage rates are the same or close to each other”.

  The present invention has been made in view of the above problems. An object of the present invention is to provide a technique for more efficiently using a processing server in a distributed system.

  The allocation determination apparatus provided by the present invention includes a resource feature amount acquisition unit that acquires a resource feature amount related to a candidate processing server that is a candidate of a processing server to which a task is allocated, and a predicted value of task processing time corresponding to the resource feature amount Based on the statistical amount, a predicted time distribution acquiring unit that acquires a predicted time distribution that is a probability distribution of, a statistical amount calculating unit that calculates a statistical amount of the predicted time distribution acquired by the predicted time distribution acquiring unit, and Assignment determination means for determining task assignment to the candidate processing server.

  The control method provided by the present invention is executed by a computer that operates as an assignment determination device. The control method includes resource feature amount acquisition means for acquiring a resource feature amount related to a candidate processing server that is a candidate of a processing server to which a task is assigned, and a probability distribution of predicted values of task processing times corresponding to the resource feature amount. A prediction time distribution acquisition unit that acquires a prediction time distribution; a statistic calculation unit that calculates a statistic of the prediction time distribution acquired by the prediction time distribution acquisition unit; and the candidate processing server based on the statistic Assignment determination means for determining task assignment.

  The program provided by the present invention causes a computer to have a function of operating as an assignment determination device provided by the present invention. The program causes the computer to have the function of each functional component included in the assignment determination apparatus provided by the present invention.

  According to the present invention, a technique for more efficiently using a processing server in a distributed system is provided.

It is a block diagram which illustrates the allocation determination apparatus which concerns on Embodiment 1 with the use environment. 1 is a block diagram illustrating a configuration of an assignment determination apparatus according to a first embodiment. It is a graph which shows the result of having measured the relationship between the disk I / O utilization rate at the time of performing a task, and task processing time. It is a figure which shows notionally the prediction time distribution acquired by the prediction time distribution acquisition part. It is a block diagram which illustrates the hardware constitutions of the allocation determination apparatus. 3 is a flowchart illustrating an example of a flow of processing executed by the assignment determination apparatus according to the first embodiment. It is a figure which illustrates the structure of the allocation determination apparatus which has a resource feature-value calculation part. FIG. 6 is a block diagram illustrating an assignment determination device according to a second embodiment. 10 is a flowchart illustrating an example of a flow of processing executed by the assignment determination apparatus according to the second embodiment. It is a figure which shows notionally the method of determining assignment of a task using the 5th percentile value, 50th percentile value, and 95th percentile value of prediction time distribution. It is a block diagram which shows the assumption environment of Example 1. FIG. FIG. 2 is a diagram illustrating a processing server included in the distributed system according to the first embodiment. It is a figure which illustrates the prediction time distribution stored in the prediction time distribution storage part in a table format. It is a graph which shows one of the prediction time distribution acquired by the prediction time distribution acquisition part. It is a graph which shows one of the prediction time distribution acquired by the prediction time distribution acquisition part. It is a graph which shows one of the prediction time distribution acquired by the prediction time distribution acquisition part. It is a block diagram which shows the assumption environment of Example 2. FIG. FIG. 10 is a diagram illustrating a processing server included in the distributed system according to the second embodiment. It is a figure which illustrates the resource information input into a resource feature-value acquisition part in a table format.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In all the drawings, the same reference numerals are given to the same components, and the description will be omitted as appropriate.

[Embodiment 1]
FIG. 1 is a block diagram illustrating an allocation determination apparatus 2000 according to the first embodiment together with its usage environment. In FIG. 1, arrows indicate the flow of information. Further, in FIG. 1, each block represents a functional unit configuration, not a hardware unit configuration.

  The distributed system 5000 has a plurality of processing servers 3000. A task is assigned to each processing server 3000. The processing server 3000 is various computers such as a server machine, a PC (Personal Computer), and a tablet terminal. The processing server 3000 executes the assigned task. The assignment determination device 2000 determines task assignment to the processing server 3000. The allocation determination apparatus 2000 is a variety of computers, like the processing server 3000. Note that one or more tasks may be executed on the processing server 3000.

  The processing server 3000 and the allocation determination device 2000 are connected via a network. This network may be a network constituted by a wired line, a network constituted by a wireless line, or a network in which a wired line and a wireless line are mixed.

  FIG. 2 is a block diagram illustrating the configuration of the allocation determination apparatus 2000 according to the first embodiment. In FIG. 2, arrows represent the flow of information. Further, in FIG. 2, each block represents a functional unit configuration, not a hardware unit configuration. The allocation determination apparatus 2000 includes a resource feature amount acquisition unit 2020, a predicted time distribution acquisition unit 2040, a statistic calculation unit 2060, and an allocation determination unit 2080. Hereinafter, each functional component will be described.

<Resource feature amount acquisition unit 2020>
The resource feature amount acquisition unit 2020 acquires a resource feature amount related to the candidate processing server. The candidate processing server is a candidate for the processing server 3000 to which a task is assigned. The resource feature amount is a feature amount related to the resource used by the candidate processing server. The resources used by the candidate processing server are, for example, CPU, memory, disk bandwidth, network bandwidth, and the like.

<Predicted time distribution acquisition unit 2040>
The predicted time distribution acquisition unit 2040 acquires a predicted time distribution corresponding to the resource feature amount acquired by the resource feature amount acquisition unit 2020. The predicted time distribution is a probability distribution of predicted values of task processing times.

<Statistics calculation unit 2060>
The statistic calculation unit 2060 calculates the statistic of the predicted time distribution acquired by the predicted time distribution acquisition unit 2040.

<Allocation determination unit 2080>
The allocation determination unit 2080 determines task allocation to the candidate processing server based on the statistic calculated by the statistic calculation unit 2060.

<Action and effect>
As described above, even when the resource usage rate of the processing server is the same or close, the time required for task processing may differ. FIG. 3 is a graph showing the results of actual measurement of the relationship between the disk I / O usage rate and the task processing time when a task is executed in the processing server. Here, the task processing time means the time required for the processing server 3000 to process a task. The X axis shows the average disk I / O usage rate during task execution, and the Y axis shows the task processing time when the task is executed. Each plot shows a combination of measured disk I / O utilization and task processing time.

  Each of the graphs 10 to 40 is a function showing the relationship between the disk I / O usage rate and the predicted value of the task processing time corresponding to the disk I / O usage rate (hereinafter, predicted time). Graphs 10 to 40 are functions generated by different methods based on the relationship between the actually measured disk I / O usage rate and the task processing time. The graph 10 is a graph in which the average value of all task processing times is the predicted time. The graph 20 is a function generated by performing linear regression using all plots. Graph 30 is a graph generated by applying all plots to the M / M1 queue model. Graph 40 is a graph generated by applying all plots to the M / D / 1 queuing model.

  Here, attention is paid to a range 80 in FIG. The range 80 includes many plots. This indicates that even when the disk I / O usage rate is high (example: 70%), the task processing time may be short (example: 20 seconds).

  However, when the task processing time is predicted using the graphs 20 to 40, a value larger than the predicted time included in the range 80 is calculated. Here, normally, a new task is assigned to a processing server having a small estimated processing time. Therefore, when the task processing time is predicted using the graphs 20 to 40 in the situation of FIG. 3, if the disk I / O usage rate is high, no task is assigned to this processing server. As a result, this processing server is not used efficiently. Further, when the predicted time is calculated using the graph 10, the resource usage rate is not considered. Therefore, even when the graph 10 is used, this processing server is not efficiently used.

  On the other hand, the allocation determination apparatus 2000 acquires a predicted time distribution corresponding to the resource feature amount related to the candidate processing server. FIG. 4 is a diagram conceptually showing a predicted time distribution corresponding to each of a plurality of resource feature amounts. A graph 50 is a predicted time distribution corresponding to a case where the disk I / O usage rate is 40%, and a graph 60 is a predicted time distribution corresponding to a case where the disk I / O usage rate is 60%. Is a predicted time distribution corresponding to a case where the disk I / O usage rate is 80%.

  When the statistic of the predicted time distribution acquired by the predicted time distribution acquisition unit 2040 is calculated, the statistic reflects the variation in the predicted time when the resource feature value is the same value or a close value. Become. For example, it is assumed that the predicted time distribution acquisition unit 2040 acquires the predicted time distribution represented by the graph 70. The graph 70 shows that the predicted time varies under the condition that the disk I / O usage rate is 80%. Specifically, the graph 70 indicates that “there is a high probability that the prediction time will be about 70 seconds, but the prediction time may be 30 seconds or less”. Then, when the statistical amount of the predicted time distribution represented by the graph 70 is calculated, such a variation in the predicted time is reflected in the statistical amount.

  According to the assignment determination apparatus 2000 of the present embodiment, the task assignment is determined for the candidate processing server based on the statistic that reflects the variation in the prediction time represented by the prediction time distribution according to the resource feature amount. . In this way, determination regarding task assignment is performed in consideration of “variation in prediction time when resource feature values are the same or close to each other”. As a result, for example, it is determined that a task may be assigned to a processing server that “the processing time may be short even when the resource usage rate is high”. Therefore, according to the allocation determination apparatus 2000 of the present embodiment, the processing server can be used more efficiently.

<Hardware configuration>
FIG. 5 is a block diagram illustrating a hardware configuration of the allocation determination apparatus 2000. In FIG. 5, the assignment determination apparatus 2000 includes a bus 1020, a processor 1040, a memory 1060, a storage 1080, and a network adapter 1100.

  The bus 1020 is a data transmission path through which the processor 1040, the memory 1060, the storage 1080, and the network adapter 1100 transmit / receive data to / from each other. The processor 1040 is an arithmetic processing unit such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The memory 1060 is a memory such as a RAM (Random Access Memory) or a ROM (Read Only Memory). The storage 1080 is a storage device such as a hard disk, an SSD (Solid State Drive), or a memory card. The storage 1080 may be a memory such as a RAM or a ROM. The network adapter 1100 is an interface for connecting to a network.

  The resource feature amount acquisition module 1220 is a program for causing the allocation determination apparatus 2000 to have the function of the resource feature amount acquisition unit 2020. The processor 1040 implements the function of the resource feature quantity acquisition unit 2020 by executing the resource feature quantity acquisition module 1220.

  The predicted time distribution acquisition module 1240 is a program for causing the allocation determination apparatus 2000 to have the function of the predicted time distribution acquisition unit 2040. The processor 1040 implements the function of the predicted time distribution acquisition unit 2040 by executing the predicted time distribution acquisition module 1240.

  The statistic calculation module 1260 is a program for causing the allocation determination apparatus 2000 to have the function of the statistic calculation unit 2060. The processor 1040 implements the function of the statistic calculation unit 2060 by executing the statistic calculation module 1260.

  The assignment determination module 1280 is a program for causing the assignment determination apparatus 2000 to have the function of the assignment determination unit 2080. The processor 1040 implements the function of the assignment determination unit 2080 by executing the assignment determination module 1280.

  For example, the processor 1040 reads the above modules onto the memory 1060 and executes them. However, the processor 1040 may execute the modules without reading them onto the memory 1060.

  The storage 1080 stores the above modules.

  The hardware configuration of allocation determination apparatus 2000 is not limited to the configuration shown in FIG. For example, each module may be stored in the memory 1060. In this case, the allocation determination apparatus 2000 may not include the storage 1080.

<Process flow>
FIG. 6 is a flowchart illustrating the flow of processing executed by the assignment determination apparatus 2000 according to the first embodiment. In step S102, the resource feature amount acquisition unit 2020 acquires a resource feature amount related to the candidate processing server. In step S104, the predicted time distribution acquisition unit 2040 acquires a predicted time distribution corresponding to this resource feature amount. In step S106, the statistic calculation unit 2060 calculates the statistic of the predicted time distribution. In step S108, the assignment determination unit 2080 determines whether to assign a task to the candidate processing server based on this statistic.

  Hereinafter, the allocation determination apparatus 2000 of this embodiment will be described in more detail.

<Details of Resource Feature Quantity Acquisition Unit 2020>
The source feature amount acquired by the resource feature amount acquisition unit 2020 is, for example, the usage amount or usage rate of the resource used by the candidate processing server. Further, the resource feature amount may be the number of waiting resources. The number of waiting resources is the number of processes, requests, etc. waiting for the resource allocation. For example, the number of CPU waits is expressed by the queue length of the run queue of the process scheduler that manages the process assigned to the CPU. For example, the number of waiting for a disk is represented by the queue length of the I / O scheduler that manages requests for that disk.

  The resource feature amount may be a value related to a resource currently used by the candidate processing server, or may be a value related to a resource used in the past. Further, the resource feature amount may be a statistical value of a value related to the resource at a plurality of points in time. Furthermore, the resource feature amount is not limited to the actual measurement value. For example, the resource feature amount may be a predicted value related to the resource used by the candidate processing server.

  The resource feature amount acquisition unit 2020 acquires a resource feature amount for one type or a plurality of types of resources.

  Further, the resource feature amount may be a value calculated using the resource usage amount described above. In this case, for example, the resource feature amount is a value calculated as a value that is proportional to the time required for the candidate processing server to process the task (hereinafter, task processing time). This will be specifically described below.

Assuming the queuing model M / M / 1, the relationship expressed by the following formula (1) holds between the task processing time and the resource usage rate. Here, t represents task processing time, and x represents resource usage rate. K is a positive real number. For example, the resource feature amount acquisition unit 2020 acquires 1 / (1-x) calculated using Equation (1) as the resource feature amount.

Assuming the queuing model M / M / 1, the relationship expressed by the following formula (2) holds between the task processing time and the number of waiting resources. The meaning of each symbol in Formula (2) is the same as the meaning of each symbol in Formula (1). For example, the resource feature amount acquisition unit 2020 acquires x + 1 calculated using Equation (2) as a resource feature amount.

Further, the resource feature value acquired by the resource feature value acquisition unit 2020 may be one value calculated using the resource feature value for each of a plurality of types of resources. A specific method for calculating one resource feature amount from a plurality of resource feature amounts is, for example, a method using Equation (3) below. ri is a resource feature amount related to each resource, and R is a resource feature amount calculated using each resource feature amount. For example, the resource feature amount acquisition unit 2020 acquires R calculated using Equation (3) as a resource feature amount.

  Moreover, the allocation determination apparatus 2000 may have a function of calculating resource feature amounts using the above-described mathematical expressions. A functional configuration unit having this function is referred to as a resource feature amount calculation unit 2100. FIG. 7 is a diagram exemplifying a configuration of an allocation determination apparatus 2000 having a resource feature quantity calculation unit 2100.

  The resource feature quantity calculation unit 2100 acquires information (hereinafter referred to as resource information) indicating the resource usage status of the server, and calculates a resource feature quantity based on the resource information. The value indicated by the resource information is, for example, a resource usage rate, a resource usage amount, or a resource waiting number. Then, the resource feature amount calculation unit 2100 calculates the resource feature amount from the resource usage rate or the like. For example, the resource feature amount calculation unit 2100 acquires the resource usage rate and calculates the resource feature amount using Expression (1).

  For example, the resource feature amount calculation unit 2100 determines a resource feature amount calculation method according to the type of value indicated by the resource information. In this case, the resource information indicates a combination of a value and a type of the value (eg, resource usage rate, waiting number, etc.). Then, the resource feature amount calculation unit 2100 calculates the resource feature amount using a method corresponding to the type of value indicated by the resource information. However, the type of value acquired by the resource feature quantity calculation unit 2100 may be fixed in advance. In this case, the resource feature amount calculation unit 2100 calculates the resource feature amount by a calculation method according to the type of value fixed in advance.

  There are various methods by which the resource feature amount acquisition unit 2020 obtains the resource feature amount. For example, the resource feature amount acquisition unit 2020 acquires a resource feature amount input from an external device. In addition, for example, the resource feature amount acquisition unit 2020 acquires a manually input resource feature amount. Furthermore, the resource feature amount acquisition unit 2020 may access an external device to acquire the resource feature amount. For example, the external device is a processing server. When there is a management server that manages the processing server in the distributed system, the resource feature amount acquisition unit 2020 may acquire the resource feature amount from the management server.

  When the allocation determination apparatus 2000 includes the resource feature amount calculation unit 2100, the resource feature amount calculation unit 2100 acquires resource information in the same manner as the resource feature amount acquisition unit 2020 acquires the resource feature amount. In this case, the resource feature value acquisition unit 2020 acquires the resource feature value calculated by the resource feature value calculation unit 2100. Further, the resource feature value acquisition unit 2020 may acquire both the resource feature value calculated by the resource feature value calculation unit 2100 and the resource feature value provided from the outside.

<Details of predicted time distribution acquisition unit 2040>
The predicted time distribution acquisition unit 2040 acquires a predicted time distribution according to the resource feature amount. For example, in the example illustrated in FIG. 4, it is assumed that the resource feature amount acquisition unit 2020 acquires a resource feature amount “disk I / O usage rate is 80%”. In this case, the predicted time distribution acquisition unit 2040 acquires the predicted time distribution represented by the graph 70. In the same example, it is assumed that the resource feature amount acquisition unit 2020 acquires a resource feature amount “disk I / O usage rate is 40%”. In this case, the predicted time distribution acquisition unit 2040 acquires the predicted time distribution represented by the graph 50.

  In addition, for example, the predicted time distribution acquisition unit 2040 acquires a conditional distribution of predicted time using the resource feature quantity as a condition. Then, the predicted time distribution acquisition unit 2040 acquires the predicted time distribution corresponding to the resource feature amount by applying the resource feature amount acquired by the resource feature amount acquisition unit 2020 to the acquired conditional distribution.

  The predicted time distribution may be prepared for each processing server 3000 or may be prepared in common for a plurality of processing servers 3000. The predicted time distribution may be prepared for each group of processing servers 3000 having a similar configuration. When the predicted time distribution is prepared for each processing server 3000 or each group of the processing server 3000, for example, the predicted time distribution acquisition unit 2040 includes the resource feature amount and the ID of the candidate processing server or the ID of the group to which the candidate processing server belongs. Get the predicted time distribution corresponding to.

  There are various methods by which the predicted time distribution acquisition unit 2040 acquires the predicted time distribution. For example, the predicted time distribution acquisition unit 2040 acquires the predicted time distribution corresponding to the resource feature amount from the predicted time distribution storage unit that stores the predicted time distribution in association with the resource feature amount. The predicted time distribution storage unit may be provided inside the allocation determination apparatus 2000 or may be provided outside the allocation determination apparatus 2000. Further, for example, the allocation determination apparatus 2000 inputs a resource feature amount to an apparatus that generates a predicted time distribution (hereinafter, predicted time distribution generation apparatus), and acquires the predicted time distribution output by the predicted time distribution generation apparatus. May be.

<Details of Allocation Determination Unit 2080>
There are various “determinations regarding task allocation to candidate processing servers” performed by the allocation determination unit 2080. For example, the assignment determination unit 2080 determines that a task is assigned to a candidate processing server, or determines that a task is not assigned to a candidate processing server. Further, for example, the assignment determination unit 2080 may determine that “tasks may be assigned” to the candidate processing server. A candidate processing server determined as “may be assigned a task” is assigned a task when other conditions are satisfied, for example. Arbitrary conditions can be set as “other conditions”.

<Determination of candidate processing server>
There are various methods by which the assignment determination apparatus 2000 selects a candidate processing server from among the plurality of processing servers 3000. For example, the allocation determination apparatus 2000 receives a task allocation request from the processing server 3000 and selects the processing server 3000 that transmitted the request as a candidate processing server. In addition, for example, the assignment determination apparatus 2000 sequentially selects processing servers 3000 as candidate processing servers from the processing servers 3000 included in the distributed system 5000, or randomly selects the processing servers 3000. Further, for example, the assignment determination device 2000 may accept designation of the processing server 3000 by an external device or manually. Further, for example, the assignment determination device 2000 may acquire information indicating whether to access the external device and make the processing server 3000 a candidate processing server.

  Here, the assignment determination apparatus 2000 may not select the processing server 3000 determined by the assignment determination unit 2080 as “not assigning a task” as a candidate processing server until a predetermined period elapses. By doing so, the processing servers 3000 to be determined by the allocation determination apparatus 2000 can be narrowed down, and the load on the allocation determination apparatus 2000 is reduced.

[Embodiment 2]
FIG. 8 is a block diagram illustrating an assignment determination apparatus 2000 according to the second embodiment. In FIG. 8, arrows indicate the flow of information. Further, in FIG. 8, each block represents a functional unit configuration, not a hardware unit configuration.

  The assignment determination apparatus 2000 according to the second embodiment includes a task type acquisition unit 2120. The task type acquisition unit 2120 acquires the type of task (task type) assigned to the candidate processing server. Furthermore, the predicted time distribution acquisition unit 2040 according to the second embodiment acquires the predicted time distribution corresponding to the resource feature amount acquired by the resource feature amount acquisition unit 2020 and the task type acquired by the task type acquisition unit 2120.

  The task type of a task is represented by, for example, the ID of the job to which the task belongs. Further, for example, the task type may be represented by a job phase ID or a combination of a job ID and a phase ID.

<Process flow>
FIG. 9 is a flowchart illustrating the flow of processing executed by the assignment determination apparatus 2000 according to the second embodiment. In FIG. 9, in the steps having the same reference numerals as in FIG. 6, the same processing as in FIG. 6 is performed.

  In step S202, the task type acquisition unit 2120 acquires the task type of the task to be assigned to the candidate processing server. In step S204, the predicted time distribution acquisition unit 2040 acquires a predicted time distribution corresponding to the resource feature amount and the task type.

<Action and effect>
According to the present embodiment, tasks are allocated using the predicted time distribution corresponding to the task type and the resource feature amount. Here, it is considered that the tasks included in the same task type usually have similar processing speed distributions. Therefore, a predicted time distribution used by the assignment determination device 2000 is generated for each task type. By doing so, the distribution of the predicted time represented by the predicted time distribution represents the predicted value distribution of the time required for processing the task that is to be assigned to the candidate processing server with higher accuracy, and thus more efficient. Thus, the processing server 3000 can be used.

[Embodiment 3]
The configuration of the allocation determination apparatus 2000 according to the third embodiment is represented in FIG. 2 as with the allocation determination apparatus 2000 according to the first embodiment. Except for the points described below, the allocation determination apparatus 2000 of the third embodiment is the same as the allocation determination apparatus 2000 of the first or second embodiment.

<Statistics calculation unit 2060>
The statistic calculation unit 2060 according to the third embodiment calculates a representative value of the prediction time distribution as the statistic of the prediction time distribution. There are various representative values calculated by the statistic calculator 2060. For example, the representative value calculated by the statistic calculation unit 2060 is the maximum value, minimum value, average value, X percentile value (0 <X <100), or quartile of the prediction time indicated by the prediction time distribution.

In addition, for example, the statistic calculation unit 2060 calculates a representative value of the predicted time distribution using a method called a direct method. Specifically, first, the statistic calculator 2060 generates a uniform random number y. Then, the statistic calculation unit 2060 calculates the value of x in Equation (4) so that the uniform random number y satisfies Equation (4) below. Here, f (x) in Expression (4) is the predicted time distribution acquired by the predicted time distribution acquisition unit 2040. Then, the statistic calculation unit 2060 sets the calculated x as a representative value of the predicted time distribution.

<Allocation determination unit 2080>
The assignment determination unit 2080 uses the representative value calculated by the statistic calculation unit 2060 to determine the task assignment to the candidate processing server. For example, if the representative value exceeds a predetermined value, the assignment determination unit 2080 determines that no task is assigned to the candidate processing server.

  Further, for example, the assignment determination unit 2080 may determine task assignment for a candidate processing server by comparing representative values calculated for each of the plurality of candidate processing servers. In this case, the resource feature value acquisition unit 2020 acquires a resource feature value for each of the plurality of candidate processing servers. Further, the predicted time distribution acquisition unit 2040 acquires, for each candidate processing server, a predicted time distribution corresponding to the resource feature amount of the candidate processing server. Then, the statistic calculation unit 2060 calculates a representative value for each of the predicted time distributions acquired for each of the plurality of candidate processing servers.

  There are various methods for comparing the representative values calculated for a plurality of candidate processing servers. For example, the assignment determination unit 2080 determines to assign a task to the candidate processing server having the smallest representative value. For example, the assignment determination unit 2080 determines that tasks may be assigned to the top N candidate processing servers when the candidate processing servers are arranged in ascending order of representative values. Also, for example, the assignment determination unit 2080 generates a representative value probability distribution using each representative value. Then, the assignment determination unit 2080 determines that tasks can be assigned only to candidate processing servers whose representative values are included in the predetermined range of the probability distribution. The predetermined range is represented by [μ−σ, μ + σ], for example, when the average value of the probability distribution is μ and the standard deviation is σ.

<< When there are multiple candidate processing servers to which a task may be assigned >>
When there are a plurality of candidate processing servers for which it is determined that a task may be allocated, there are various methods by which the allocation determining unit 2080 determines a candidate processing server to which a task is allocated. For example, the allocation determination unit 2080 allocates a task to a candidate processing server having the highest load (resource usage rate or resource usage) among candidate processing servers determined to be able to allocate a task. As a result, the processing server 3000 having a high load is more actively used. Conversely, the assignment determination unit 2080 may assign a task to a candidate processing server with the lowest load among candidate processing servers determined to be able to assign a task.

  Also, for example, the assignment determination unit 2080 compares the number of tasks already assigned to each candidate processing server, and determines the candidate processing server to which the task is assigned so that the number of assigned tasks is as uniform as possible. To do.

<Action and effect>
The statistic calculation unit 2060 according to the third embodiment calculates a representative value of the prediction time distribution as the statistic of the prediction time distribution. By using the representative value of the prediction time distribution as the statistic of the prediction time distribution, it is possible to realize task assignment determination that more reflects the characteristics of the shape of the probability distribution by a simple method. For example, if the 95th percentile value is used as the representative value, the processing time is shown to be less than this representative value with a probability of 95%. Therefore, for example, when it is determined that “assign a task if the representative value is less than or equal to the threshold value”, the determination result is shown to be correct with a probability of 95%. This increases the probability that task assignment determination accuracy will be higher.

  For example, when the prediction processing time using the direct method is used as the representative value, the prediction processing time according to the probability distribution can be directly obtained. Therefore, when task assignment is determined from this representative value, task assignment can be performed more finely. This increases the probability that task processing time will be shortened.

[Embodiment 4]
The configuration of the allocation determination apparatus 2000 according to the fourth embodiment is represented in FIG. 2 as with the allocation determination apparatus 2000 according to the first embodiment. Except as described below, the allocation determination apparatus 2000 of the fourth embodiment is the same as the allocation determination apparatus 2000 of the first or second embodiment.

<Statistics calculation unit 2060>
The statistic calculation unit 2060 according to the fourth embodiment calculates the distribution degree of the predicted time distribution as the statistic of the predicted time distribution. The degree of dispersion calculated by the statistic calculator 2060 varies. For example, the dispersity calculated by the statistic calculator 2060 is the standard deviation or variance of the predicted time indicated by the predicted time distribution, or the difference between the X and Y percentile values (0 <X <100, 0 <Y <100), etc. It is. As the difference between the X and Y percentile values, for example, the difference between the 95th and 5th percentile values is used.

<Allocation determination unit 2080>
For example, the assignment determination unit 2080 determines assignment of tasks to candidate processing servers using the distribution degree calculated by the statistic calculation unit 2060. For example, the assignment determination unit 2080 does not assign a task to the candidate processing server when the distribution degree exceeds a predetermined value. This is because when the spread degree is large, it is considered that there is a large risk that the processing time becomes long.

  Further, for example, the assignment determination unit 2080 determines that a task may be assigned when the predicted time distribution is biased toward a small value. This is because if the predicted time distribution is biased to a small value, it is considered that there is a high probability that the processing time will be shortened. Specifically, processing is performed by the following method. First, the statistic calculation unit 2060 calculates the (50−X) percentile value, 50th percentile value, and (50 + X) percentile value of the predicted time distribution, respectively. Here, X is a real number satisfying 0 <x <50. For example, use 45 as the value of X. In this case, the 5th, 50th and 95th percentile values will be used.

  Then, the statistic calculation unit 2060 calculates the difference between the (50 + X) percentile value and the 50 percentile value as the first distribution. Further, the statistic calculation unit 2060 calculates the difference between the 50th percentile value and the (50−X) percentile value as the second distribution degree.

  The allocation determination unit 2080 compares the first spreading degree and the second spreading degree. Only when the first distribution degree is smaller than the second distribution degree, it is determined that the task is assigned to the candidate processing server.

  FIG. 10 is a diagram conceptually illustrating a method of determining task assignment using the 5th percentile value, 50th percentile value, and 95th percentile value of the predicted time distribution. In FIG. 10, the 5th percentile value is 10, the 50th percentile value is 35, and the 95th percentile value is 90. Therefore, the first spread is 25 and the second spread is 65. Therefore, the graph of FIG. 10 is a graph biased toward a smaller prediction time.

  Further, for example, the assignment determination unit 2080 may make a determination regarding task assignment by comparing the distribution degrees calculated for each of the plurality of candidate processing servers. In this case, the resource feature value acquisition unit 2020 acquires a resource feature value for each of the plurality of candidate processing servers. Further, the predicted time distribution acquisition unit 2040 acquires, for each candidate processing server, a predicted time distribution corresponding to the resource feature amount of the candidate processing server. Then, the statistic calculation unit 2060 calculates the degree of dispersion for each predicted time distribution acquired for each of the plurality of candidate processing servers.

  There are various methods for comparing the distribution degrees calculated for a plurality of candidate processing servers. For example, the assignment determination unit 2080 determines that a task is not assigned to a candidate processing server having the largest distribution degree. Further, for example, the assignment determination unit 2080 determines that no task is assigned to the top N candidate processing servers when the candidate processing servers are arranged in descending order of spread degree. Further, for example, the assignment determination unit 2080 generates a probability distribution of the distribution degree using each distribution degree. Then, the assignment determination unit 2080 determines that a task may be assigned only to a candidate processing server whose distribution degree is included in the predetermined range of the probability distribution. For example, the predetermined range of the spread degree is the same as the predetermined range of the representative value described in the third embodiment.

<< When there are multiple candidate processing servers to which a task may be assigned >>
The assignment determination unit 2080 according to the third embodiment performs a method for determining a candidate processing server to which a task is assigned when there are a plurality of candidate processing servers for which it is determined that the assignment determination unit 2080 according to the fourth embodiment may assign a task. It is the same as the method.

<Action and effect>
The statistic calculation unit 2060 according to the fourth embodiment calculates the distribution degree of the predicted time distribution as the statistic of the predicted time distribution. By using the distribution degree of the prediction time distribution as the statistic of the prediction time distribution, it is possible to realize task assignment determination that more reflects the degree of dispersion of the processing time by a simple method. For example, if skewness is used as the spread degree, it is possible to determine whether the distribution of the prediction processing time at a certain load tends to be small or large by examining the absolute value of the skewness and the sign of positive and negative become. This increases the probability that processing time will be shortened when tasks are assigned.

[Example 1]
<Assumed environment>
A specific operation example of the assignment determination apparatus 2000 according to the first embodiment will be described as a first example. FIG. 11 is a diagram illustrating an assumed environment of the first embodiment. In FIG. 11, arrows indicate the flow of information. Further, in FIG. 11, each block represents a functional unit configuration, not a hardware unit configuration.

<< Allocation Determination Device 2000 >>
The assignment determination apparatus 2000 according to the first embodiment has the same configuration as the assignment determination apparatus 2000 according to the second embodiment. Further, the statistic calculation unit 2060 according to the first embodiment calculates a representative value of the predicted time distribution as the statistic of the predicted time distribution. Here, the statistic calculator 2060 calculates the representative value of the predicted time distribution using the direct method described above. When the representative value of the predicted time distribution acquired for the combination of the resource feature amount and task type of the candidate processing server exceeds 90 seconds, the assignment determination unit 2080 of the first embodiment will determine the candidate processing server for the task type task. Is determined not to be assigned.

<< Distributed system 5000 >>
The distributed system 5000 includes a master server 4000 and a plurality of processing servers 3000. The master server 4000 is a server that manages the processing server 3000.

<< Processing Server 3000 >>
The processing server 3000 includes a task execution unit 3020, a resource information storage unit 3040, and a task information storage unit 3060. The task execution unit 3020 executes one or a plurality of tasks. For example, the task execution unit 3020 is a process, a thread, a virtual machine, or the like. The processing server 3000 may include a plurality of task execution units 3020.

  The resource information storage unit 3040 stores resource information. As described above, the resource information is information representing the resource usage status of the server, such as a resource usage rate. A method for measuring the usage amount of each resource in the processing server 3000 is a known technique. Therefore, description of the method is omitted.

  The task information storage unit 3060 stores information related to tasks (task information). The task information stored in the task information storage unit 3060 is task information related to a task executed by the task execution unit 3020 or a task executed in the past by the task execution unit 3020.

<< Master server 4000 >>
The master server 4000 includes a task execution management unit 4020, a resource information management unit 4040, and a task information management unit 4060. The task execution management unit 4020 performs processing for assigning a task to the processing server 3000, processing for instructing the processing server 3000 to start or stop the task, and the like. The resource information management unit 4040 collects and manages resource information from the resource information storage unit 3040 of each processing server 3000. The task information management unit 4060 manages tasks executed in the distributed system 5000. For this purpose, the task information management unit 4060 collects and manages task information from the task information storage unit 3060 of each processing server 3000. The task information management unit 4060 also manages information related to tasks that are not yet assigned to the processing server 3000.

  The task execution management unit 4020 acquires information regarding the resources of the processing server 3000 from the resource information management unit 4040 and inputs the information to the resource feature amount acquisition unit 2020. Further, the task execution management unit 4020 acquires the task type of the target task (hereinafter referred to as the target task) to be assigned to the processing server 3000 from the task information management unit 4060 and inputs it to the task type acquisition unit 2120. The assignment determination device 2000 makes a determination on task assignment to the processing server 3000 and notifies the task execution management unit 4020 of the determination result. The task execution management unit 4020 performs task allocation to the processing server 3000 based on the determination result acquired from the allocation determination apparatus 2000.

<< Predicted time distribution storage unit 6000 >>
The predicted time distribution storage unit 6000 stores the predicted time distribution in association with the task type and the resource feature amount.

  FIG. 12 is a diagram illustrating the processing server 3000 included in the distributed system 5000 according to the first embodiment. The distributed system 5000 according to the first embodiment includes computers c1 and c2 as the processing server 3000. Computers c1 and c2 are connected via a single network switch. The computer c1 has processes P1 and P2 as a task execution unit 3020. Further, the computer c2 has processes P3 and P4 as the task execution unit 3020. Process P1 is executing task T1, and process P3 is executing task T2. In FIG. 12, the master server 4000 is omitted.

<Specific examples of processing>
Hereinafter, a specific example of processing performed in the assumed environment described above will be described. Here, since the process P2 is free, the computer c1 requests the master server 4000 to assign a task. Then, the master server 4000 determines a task to be assigned next. The master server 4000 first tries to allocate a Type A task for the purpose of balancing the resource allocation amount between jobs. Therefore, the task execution management unit 4020 inputs a task type of Type A to the task type acquisition unit 2120. In addition, the task execution management unit 4020 inputs the resource feature amount related to the resource of the computer c1 that is the candidate processing server to the resource feature amount acquisition unit 2020. Here, it is assumed that “disk I / O issue rate = 90%” is input as the resource feature amount of the computer c1.

  FIG. 13 is a diagram illustrating the predicted time distribution stored in the predicted time distribution storage unit 6000 in a table format. The table shown in FIG. 13 is referred to as a predicted time distribution table 100. The predicted time distribution table 100 has three columns of a task type 102, an average value 104, and a standard deviation 106. Each record of the predicted time distribution table 100 indicates a normal distribution determined by the average value 104 and the standard deviation 106 in association with the task type indicated by the task type 102.

  The predicted time distribution acquisition unit 2040 acquires the average value 104 and the standard deviation 106 corresponding to “task type 102 = Type A” from the predicted time distribution storage unit 6000. As a result, the predicted time distribution acquisition unit 2040 acquires 10 / (1-ρ) +5 as the average value 104, and acquires 100ρ / 3 as the standard deviation 106. Here, ρ represents the disk I / O issue rate.

  Then, the predicted time distribution acquisition unit 2040 applies “disk I / O issue rate (ρ) = 90%” to the acquired average value 104 and standard deviation 106. This gives an average value of 105 seconds and a standard deviation of 30 seconds. As a result, the predicted time distribution acquisition unit 2040 obtains a normal distribution having an average value of 105 seconds and a standard deviation of 30 seconds. That is, the predicted time distribution acquisition unit 2040 has an average value of 105 seconds and a standard deviation of 30 seconds as a predicted time distribution corresponding to “task type = Type A” and “disk I / O issue rate = 90%”. Get a normal distribution. FIG. 14 is a graph showing the predicted time distribution.

  The statistic calculator 2060 calculates the representative value of the predicted time distribution using the direct method. Here, it is assumed that the representative value calculated by the statistic calculator 2060 is 110 seconds.

  The assignment determination unit 2080 determines the assignment of the Type A task to the computer c1 using the calculated representative value. As described above, when the representative value of the predicted time distribution acquired for the combination of the resource feature amount and the task type of the candidate processing server exceeds 90 seconds, the allocation determination unit 2080 performs the candidate processing for the task type task. It is determined not to allocate to the server. Since the representative value calculated by the statistic calculation unit 2060 is 110 seconds and exceeds 90 seconds, the assignment determination unit 2080 does not assign a task whose type of task is Type A to the computer c1. judge.

  Next, it is assumed that the master server 4000 tries to assign a Type B task to the computer c1. The task execution management unit 4020 inputs “task type = Type B” to the task type acquisition unit 2120. Also, the task execution management unit 4020 inputs “disk I / O issue rate = 90%” to the resource feature quantity acquisition unit 2020.

  The predicted time distribution acquisition unit 2040 acquires a predicted time distribution corresponding to “task type = Type B” and “disk I / O issue rate = 90%”. As shown in FIG. 13, “task type = Type B” corresponds to a normal distribution having an average value of 1 / (1-ρ) +10 and a standard deviation of 5ρ. The predicted time distribution acquisition unit 2040 obtains a normal distribution having an average value of 20 seconds and a standard deviation of 4.5 seconds by applying “disk I / O issue rate = 90%” to this normal distribution. That is, the predicted time distribution acquisition unit 2040 has an average value of 20 seconds and a standard deviation of 4.5 seconds as a predicted time distribution corresponding to “task type = Type B” and “disk I / O issue rate = 90%”. Get a normal distribution that is. FIG. 15 is a graph showing the predicted time distribution.

  The statistic calculation unit 2060 calculates a representative value of the predicted time distribution. Assume that the calculated representative value is 19 seconds.

  The assignment determination unit 2080 determines the assignment of the Type B task to the computer c1. The representative value calculated by the statistic calculator 2060 is 19 seconds, which is smaller than 90 seconds. Therefore, the assignment determination unit 2080 determines to assign a task whose task type is Type B to the computer c1. As a result, a type B task is assigned to the process P2 of the computer c1.

  Furthermore, it is assumed that the computer c2 requests the master server 4000 to assign a task for the process P4. Assume that the master server 4000 attempts to assign a Type A task. Also assume that the disk I / O issue rate of computer c2 at this time is 50%. Then, the task execution management unit 4020 inputs “disk I / O issue rate = 50%” to the resource feature amount acquisition unit 2020 and inputs “task type = Type A” to the task type acquisition unit 2120. To do.

  As a result, the predicted time distribution acquisition unit 2040 acquires a normal distribution having an average value of 25 seconds and a standard deviation of 30 seconds as the predicted time distribution. FIG. 16 is a graph showing the predicted time distribution.

  The statistic calculation unit 2060 calculates a representative value of the predicted time distribution. Here, it is assumed that the calculated representative value is 50 seconds. Since this representative value is smaller than 90 seconds, the assignment determination unit 2080 determines that a task whose task type is Type A is assigned to the computer c2. As a result, the task execution management unit 4020 assigns a task whose task type is Type A to the process P4 of the computer c2.

  Further, it is assumed that the process P2 of the computer c1 completes the execution of the task, and the computer c1 requests the master server 4000 for task assignment. Here, it is assumed that the master server tries to assign a Type A task to c1. The task execution management unit 4020 inputs “disk I / O issue rate = 90%” to the resource feature amount acquisition unit 2020 and inputs “task type = Type A” to the task type acquisition unit 2120. As a result, the predicted time distribution acquisition unit 2040 acquires a normal distribution having an average value of 105 seconds and a standard deviation of 30 seconds as the predicted time distribution (see FIG. 14).

  The statistic calculation unit 2060 calculates a representative value of the predicted time distribution. Here, it is assumed that the statistic calculation unit 2060 calculates 50 seconds as a representative value. In this case, since the representative value of 50 seconds is smaller than 90 seconds, the assignment determination unit 2080 determines that a task is assigned to the computer c1. As a result, a task whose type is Type A is assigned to process P2 of computer c1.

  Here, in the above example, when “disk I / O issuance rate = 90%” and “task type = Type A”, it is determined that no task is allocated and when it is determined that a task is allocated. There were both. This is because the statistic calculation unit 2060 calculates a random value according to the predicted time distribution as a representative value using the direct method. From this, it can be seen that “variation in task processing time when the resource usage rate is the same or close” is taken into consideration.

[Example 2]
Another embodiment of the allocation determination apparatus 2000 will be described as a second embodiment. Except as described below, the assumed environment of the second embodiment is the same as the assumed environment of the first embodiment. FIG. 17 is a diagram illustrating an assumed environment of the first embodiment. In FIG. 11, arrows indicate the flow of information. Further, in FIG. 11, each block represents a functional unit configuration, not a hardware unit configuration.

  The second embodiment is different from the first embodiment in that the assignment determination device 2000 does not have the task type acquisition unit 2120. Therefore, the assignment determination apparatus 2000 acquires the predicted time distribution corresponding to the resource feature amount without considering the task type in the assumed environment of the second embodiment.

  The predicted time distribution acquisition unit 2040 acquires the predicted time distribution for each of the plurality of candidate processing servers. The statistic calculation unit 2060 calculates a representative value of the predicted time distribution for each of the plurality of candidate processing servers. The allocation determination unit 2080 generates a probability distribution using the calculated representative values. Then, for each candidate processing server, the allocation determination unit 2080 only applies to the candidate processing server when the representative value calculated for the candidate processing server is in the range [μ−2σ, μ + 2σ] in the probability distribution. It is determined that the task can be assigned. Here, μ is an average value of the probability distribution, and σ is a standard deviation of the probability distribution.

  FIG. 18 is a diagram illustrating the processing server 3000 included in the distributed system 5000 according to the second embodiment. The distributed system 5000 according to the second embodiment includes three computers c1 to c3 as the processing server 3000. Each computer is connected via a network switch. Each computer has a plurality of processes as the task execution unit 3020. Specifically, computer c1 has processes P1 and P2, computer c2 has processes P3 and P4, and computer c3 has processes P5 and P6. In FIG. 18, the master server 4000 is omitted.

  Here, it is assumed that the computer c1 requests the master server 4000 for a task to be executed in the process P2. The master server 4000 determines whether to assign a task to the computer c1.

  The resource feature amount acquisition unit 2020 receives resource information indicating the resource feature amount of each processing server 3000. In this embodiment, the network bandwidth utilization rate is used as the resource feature amount. FIG. 19 is a diagram illustrating resource information input to the resource feature amount acquisition unit 2020 in a table format. The table shown in FIG. 19 is referred to as a resource information table 200. The resource information table 200 has two columns of a processing server identifier 202 and a network bandwidth utilization rate 204. Each record of the resource information table 200 indicates the utilization rate of the network bandwidth used in the processing server 3000 in association with the processing server 3000 represented by the processing server identifier 202. From FIG. 19, the resource feature amount related to the computer c1 is “network bandwidth utilization rate = 50%”, the resource feature amount related to the computer c2 is “network bandwidth utilization rate = 25%”, and the resource feature amount related to the computer c3 is “ Network bandwidth utilization rate = 75% ".

  The predicted time distribution acquisition unit 2040 acquires the predicted time distribution for each of the computers c1, c2, and c3. Then, the statistic calculation unit 2060 calculates a representative value for each predicted time distribution of the computers c1, c2, and c3. Here, it is assumed that the calculated statistical values are 20 seconds, 20 seconds, and 80 seconds, respectively.

  The assignment determination unit 2080 determines whether or not task assignment is possible for each of the computers c1 to c3. The allocation determination unit 2080 calculates an average value μ and a standard deviation σ of all representative values calculated by the statistic calculation unit 2060. The calculation results are assumed to be μ = 40 and σ = 35. In this case, [μ−2σ, μ + 2σ] is [5, 75]. Therefore, the assignment determination unit 2080 determines that a task can be assigned to a computer only when the representative value x calculated for the computer satisfies 5 <= x <= 75.

  The typical value calculated for computer c1 is 20, and 5 ≦ 20 ≦ 75. Therefore, the assignment determination unit 2080 determines that a task can be assigned to the computer c1. Similarly, since the representative value calculated for the computer c2 is 20, the assignment determination unit 2080 determines that a task may be assigned to the computer c2. On the other hand, the representative value calculated for computer c3 is 80, which is outside the range of [5, 75]. Therefore, the assignment determination unit 2080 determines that no task is assigned to the computer c3.

  As mentioned above, although embodiment of this invention was described with reference to drawings, these are the illustrations of this invention, Various structures other than the above are also employable.

10, 20, 30, 40, 50, 60, 70 Graph 80 Range 100 Prediction time distribution table 102 Task type 104 Average value 106 Standard deviation 200 Resource information table 202 Processing server identifier 204 Network bandwidth utilization 1020 Bus 1040 Processor 1060 Memory 1080 Storage 1100 Network adapter 1220 Resource feature acquisition module 1240 Predictive time distribution acquisition module 1260 Statistics calculation module 1280 Allocation determination module 2000 Allocation determination device 2020 Resource feature acquisition unit 2040 Predictive time distribution acquisition unit 2060 Statistics calculation unit 2080 Allocation determination unit 2100 Resource feature amount calculation unit 2120 Task type acquisition unit 3000 Processing server 3020 Task execution unit 3040 Resource information storage unit 3060 Task information storage unit 4000 Master server 4020 Task execution management unit 4040 Resource information management unit 4060 Task information management unit 5000 Distributed system 6000 Predictive time distribution storage unit

Claims (24)

Resource feature amount acquisition means for acquiring a resource feature amount related to a candidate processing server that is a candidate of a processing server to which a task is assigned;
A predicted time distribution acquisition unit that acquires a predicted time distribution that is a probability distribution of a predicted value of task processing time corresponding to the resource feature amount;
A statistic calculating means for calculating a statistic of the predicted time distribution acquired by the predicted time distribution acquiring means;
An assignment determination unit configured to determine assignment of tasks to the candidate processing server based on the statistics;
An allocation determination device.   The allocation determination apparatus according to claim 1, wherein the statistic calculation unit calculates a representative value of the predicted time distribution as the statistic.   The allocation determination apparatus according to claim 1, wherein the statistic calculation unit calculates a distribution degree of the predicted time distribution as the statistic. The statistic calculation means includes a first distribution degree which is a difference between a (50 + X) percentile value and a 50 percentile value of the predicted time distribution, and a 50 percentile value and a (50-X) percentile value of the predicted time distribution. And calculate the second spreading degree which is the difference from (0 <x <50)
The allocation determination unit according to claim 3, wherein the allocation determination unit compares the first distribution degree and the second distribution degree, and assigns a task to the candidate processing server when the first distribution degree is smaller. apparatus.   The allocation determination unit determines whether or not the statistic exceeds a predetermined value, and if the statistic exceeds the predetermined value, does not allocate a task to the processing server. The allocation determination device according to item. The statistic calculation means calculates a statistic of the predicted time distribution for each of the plurality of candidate processing servers based on the predicted time distribution corresponding to the resource feature amount of each of the plurality of candidate processing servers,
The allocation determination unit compares the statistics calculated for each of the plurality of candidate processing servers and allocates a task to any of the plurality of candidate processing servers based on the comparison result. The allocation determination device according to one item. Task type acquisition means for acquiring a task type that is a type of task to be assigned to the candidate processing server;
The allocation determination apparatus according to claim 1, wherein the predicted time distribution acquisition unit acquires the predicted time distribution corresponding to the resource feature amount and the task type.   The predicted time distribution acquisition unit corresponds to the resource feature amount acquired by the resource feature amount acquisition unit from the predicted time distribution storage unit that stores the predicted time distribution in association with each of a plurality of resource feature amounts. The allocation determination apparatus according to claim 1, which acquires a predicted time distribution. A control method executed by a computer operating as an allocation determination device,
A resource feature amount acquisition step of acquiring a resource feature amount related to a candidate processing server that is a candidate of a processing server to which a task is assigned;
A predicted time distribution acquisition step of acquiring a predicted time distribution corresponding to the resource feature quantity, which is a probability distribution of predicted values of task processing times;
A statistic calculating step of calculating a statistic of the predicted time distribution acquired in the predicted time distribution acquiring step;
An assignment determination step for determining assignment of tasks to the candidate processing server based on the statistics;
A control method.   The control method according to claim 9, wherein the statistic calculation step calculates a representative value of the predicted time distribution as the statistic.   The control method according to claim 9, wherein the statistic calculation step calculates a distribution degree of the predicted time distribution as the statistic. The statistic calculation step includes a first distribution degree that is a difference between a (50 + X) percentile value and a 50 percentile value of the predicted time distribution, and a 50 percentile value and a (50-X) percentile value of the predicted time distribution. And calculate the second spreading degree which is the difference from (0 <x <50)
The control method according to claim 11, wherein the assignment determination step compares the first spreading degree and the second spreading degree, and assigns a task to the candidate processing server when the first spreading degree is smaller. .   The assignment determination step determines whether or not the statistic exceeds a predetermined value, and if the statistic exceeds the predetermined value, a task is not allocated to the processing server. The control method according to item. The statistic calculation step calculates a statistic of the predicted time distribution for each of the plurality of candidate processing servers based on a predicted time distribution corresponding to the resource feature amount of each of the plurality of candidate processing servers,
The assignment determination step compares statistics calculated for each of the plurality of candidate processing servers, and assigns a task to any of the plurality of candidate processing servers based on the comparison result. The control method according to one item. A task type acquisition step of acquiring a task type that is a type of task assigned to the candidate processing server;
The control method according to claim 9, wherein the predicted time distribution acquisition step acquires the predicted time distribution corresponding to the resource feature amount and the task type.   The predicted time distribution acquisition step corresponds to the resource feature amount acquired by the resource feature amount acquisition step from the predicted time distribution storage means storing the predicted time distribution in association with each of a plurality of resource feature amounts. The control method according to claim 9, wherein the predicted time distribution is acquired. A program for causing a computer to have a function of operating as an assignment determination device,
A resource feature amount acquisition function for acquiring a resource feature amount related to a candidate processing server that is a candidate of a processing server to which a task is assigned;
A predicted time distribution acquisition function for acquiring a predicted time distribution that is a probability distribution of predicted values of task processing times corresponding to the resource feature amount;
A statistic calculating function for calculating a statistic of the predicted time distribution acquired by the predicted time distribution acquiring function;
An assignment determination function for determining assignment of tasks to the candidate processing server based on the statistics;
A program to give   The program according to claim 17, wherein the statistic calculation function calculates a representative value of the predicted time distribution as the statistic.   The program according to claim 17, wherein the statistic calculation function calculates a distribution degree of the predicted time distribution as the statistic. The statistic calculation function includes a first distribution degree that is a difference between a (50 + X) percentile value and a 50 percentile value of the predicted time distribution, and a 50 percentile value and a (50-X) percentile value of the predicted time distribution. And calculate the second spreading degree which is the difference from (0 <x <50)
The program according to claim 18, wherein the assignment determination function compares the first distribution degree and the second distribution degree, and assigns a task to the candidate processing server when the first distribution degree is smaller.   21. The assignment determination function determines whether or not the statistic exceeds a predetermined value, and does not allocate a task to the processing server when the statistic exceeds the predetermined value. The program described in the section. The statistic calculation function calculates a statistic of a predicted time distribution for each of the plurality of candidate processing servers based on a predicted time distribution corresponding to a resource feature amount of each of the plurality of candidate processing servers.
The allocation determination function compares statistics calculated for each of the plurality of candidate processing servers, and allocates a task to any of the plurality of candidate processing servers based on the comparison result. The program according to one item. The computer has a task type acquisition function for acquiring a task type that is a type of task assigned to the candidate processing server,
The program according to any one of claims 17 to 22, wherein the predicted time distribution acquisition function acquires the predicted time distribution corresponding to the resource feature amount and the task type.   The predicted time distribution acquisition function corresponds to a resource feature amount acquired by the resource feature amount acquisition function from a predicted time distribution storage unit that stores the predicted time distribution in association with each of a plurality of resource feature amounts. The program according to any one of claims 17 to 23, wherein a predicted time distribution is acquired.

Images