JP2018128956A - Load distribution device and load distribution method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a load distribution device and a load distribution method capable of reducing processing TAT of a request signal while keeping a processing load of a processing server constant.SOLUTION: A load distribution device 100 includes: an information storage section 130 storing a load limit value as a limit value for processing of a low delay server; a server information acquisition section 120 acquiring load information of the low delay server; a take-over signal number calculation section 111 calculating a take-over signal number S which is the number of signals which should be assigned to a high delay server but are actually assigned to the low delay server, in a range not exceeding the load limit value, on the basis of the number of high delay servers, the number of low delay servers, a maximum value of the acquired load information, and a processing server load fluctuation rate A indicating load fluctuation of the low delay server; and an assigning function section 110 performing assignment of signals of the calculated take-over signal number S from the high delay server to the low delay server.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a load distribution apparatus and a load distribution method for determining a server that executes information processing requested by a client for a plurality of processing servers in a data center.

Due to the development of server configuration technology, virtualization technology, and severe disaster countermeasures, data centers are connected across regions and countries, and there are multiple processing servers in the data center. In order to distribute the load to the scattered processing servers, a load distribution device such as a load balancer is installed, and the load distribution device is used to distribute request signals (requests) to the processing servers. It is common to disperse.
For example, in Non-Patent Document 1, load distribution is realized by determining a processing server as a distribution destination in consideration of the load state (CPU usage rate, etc.) and response time of each processing server.

Introduction to Load Balancing (Introduction to Load Balancer), [online], [Search January 24, 2017], Internet <URL: http://fenics.fujitsu.com/products/ipcom/catalog/data/1/3. html> Yuta Kitano, Mitsuhiro Okamoto “Load Balancing Method Considering Inter-Communication Communication Delay” 2016 Electronic Information Communication Society Conference, B-6-13

However, when a signal is distributed to processing servers across regions and countries, a processing delay corresponding to the region and country is caused in the signal. When the processing delay occurs, for example, in the communication control server, it is considered that the subsequent signal may overtake the preceding signal.
On the other hand, in order to prevent processing delay, if a signal is assigned with priority to the processing server in the base, only the processing server in the base becomes heavy, leading to inefficient hardware resources and unnecessary scaling. End up.
The load distribution method described in Non-Patent Document 1 determines a processing server in consideration of response time and processing server load. However, since it is necessary to collect the load status and response time of all processing servers, if processing servers exist across regions and countries, the signal for collecting information increases and compresses the network bandwidth. there's a possibility that.

  The present invention has been made in view of such a background, and the present invention provides a load distribution apparatus capable of reducing a processing signal TAT (Turn Around Time) of a request signal while keeping the processing load of the processing server below a certain level. It is another object of the present invention to provide a load balancing method.

  In order to solve the above-described problem, the invention according to claim 1 is a load distribution apparatus that includes a plurality of processing servers in a base and determines the processing server that executes information processing requested by a client. A processing time measuring means for measuring the processing TAT of each of the processing servers, a classifying means for classifying the processing servers into a high delay server and a low delay server based on the measured processing TAT, A storage unit that stores a load limit value that is a processing limit value, a load information acquisition unit that acquires load information of the low-delay server, the number of high-delay servers, and the number of low-delay servers are acquired. Based on the maximum value of the load information and the load fluctuation rate of the processing server indicating the load fluctuation of the low-latency server, the high-delay server within a range not exceeding the load limit value A load comprising: a signal number calculating means for calculating the number of substitution signals to distribute signals to be distributed to the low delay server; and a distribution means for distributing the calculated number of substitution signals from the high delay server to the low delay server. A dispersion apparatus was used.

  The invention according to claim 7 is a load distribution method of a load distribution apparatus that includes a plurality of processing servers in a base and determines the processing server that executes information processing requested by a client. The load distribution apparatus includes a processing time measuring step for measuring a processing TAT of each of the processing servers, a classification step for classifying the processing servers into a high delay server and a low delay server based on the measured processing TAT, A storage step for storing a load limit value that is a limit value of processing of the low delay server, a load information acquisition step for acquiring load information of the low delay server, the number of the high delay servers, and the number of the low delay servers And exceeding the load limit value based on the acquired maximum value of the load information and the load fluctuation rate of the processing server indicating the load fluctuation of the low-latency server. A signal number calculating step for calculating the number of substitution signals to be distributed to the low delay server within a range, and the calculated number of substitution signals from the high delay server to the low delay server. The load distribution method executes the distribution step.

  By doing this, even when low-latency servers and high-latency servers coexist in the same base, or when the TAT of processing servers at other bases changes dynamically, the number of signals to be distributed is calculated appropriately. can do. As a result, the request signal processing TAT can be reduced while keeping the processing load of the processing server below a certain level.

  Further, the invention according to claim 2 is characterized in that the signal number calculation means receives the request based on a processing TAT when processed by the low-latency server and a processing TAT when processed by the high-latency server. An average signal processing TAT variation rate indicating a variation in processing response time when allocating to each processing server in a predetermined order is calculated, and a load variation rate of the low delay server and the average signal processing TAT variation rate are In addition, a score for evaluating the number of substitution signals is calculated, and the allocating means is a range that does not exceed the load limit value based on the acquired maximum value of the load information and the load fluctuation rate of the low-delay server. The number of substitution signals corresponding to the maximum score is distributed from the high delay server to the low delay server.

  In this way, it is possible to make the response time as fast as possible while limiting the processing load in consideration of both the load of the processing server and the response time. In other words, TAT can be reduced by increasing the processing load of the low-latency server to the limit.

  The invention according to claim 3 is characterized in that the load information acquisition means acquires the CPU usage rate of the low-latency server as load information.

  In this way, since the CPU usage rate of each processing server is relatively easy to acquire, the processing load of each processing server can be acquired in a timely manner and can be immediately reflected in the load distribution operation.

  In the invention according to claim 4, when the processing TAT when processed by the low delay server is 1, the processing TAT when processed by the high delay server is an average of the processing TAT of the high delay server. Is divided by the average of the processing TAT of the low-latency server.

  In this way, the signal processing TAT fluctuation rate in one round can be obtained from the communication delay magnification between the processing TAT when processed by the low-delay server and the processing TAT when processed by the high-delay server. Thereby, the substitution signal can be calculated using the load fluctuation rate of the low-delay server and the average signal processing TAT fluctuation rate.

  Further, the invention according to claim 5 is the range of the number of substitution signals that does not exceed a CPU threshold value by the distribution means multiplying the acquired maximum value of the load information by the load fluctuation rate of the low delay server. And the distribution is performed according to the value of the number of substitution signals that gives the maximum score in the calculated range of the number of substitution signals.

  In this way, since the distribution is performed based on the value of the shoulder signal number that is the maximum score in the calculated range of the shoulder signal number, the shoulder signal number S with higher reliability can be obtained. As a result, it is possible to distribute from the high-delay server to the low-delay server by the optimal number of substitution signals S.

  Further, in the invention according to claim 6, the classification means increases each processing server as a low-delay server one by one in ascending order of processing TAT to the total number −1, and the low delay in each case The number of servers and the number of high-delay servers are determined, and the signal number calculation means determines the number of the low-delay servers, the number of the low-delay servers, and processing when processing is performed by the low-delay servers Based on the TAT and the processing TAT when processed by the high-delay server, an average signal processing TAT variation rate indicating a variation in processing response time when the request is distributed to each processing server in a predetermined order is calculated. And calculating a score for evaluating the number of substitute signals based on the load variation rate of the low-delay server and the average signal processing TAT variation rate, The maximum value of the obtained load information is multiplied by the load fluctuation rate A of the low delay server, the range of the number of substitution signals not exceeding the CPU threshold value is calculated, and the maximum score in the calculated range of the number of substitution signals The distribution is performed according to the value of the number of substitution signals.

  In this way, the optimal low-latency server and high-delay server can be dynamically determined by looking at the overall processing delay and processing load of the processing server without using a constant response time threshold value. Since classification is performed, more appropriate processing server classification can be performed. As a result, load distribution can be further optimized, and the response time can be shortened as much as possible while further limiting the processing load.

  ADVANTAGE OF THE INVENTION According to this invention, the load distribution apparatus and load distribution method which can reduce a response time and the processing load of a processing server can be provided, reducing the signal amount for collecting the information of a processing server.

1 is a configuration diagram illustrating a distribution system to which a load distribution apparatus according to a first embodiment of the present invention is applied. It is a figure which shows the base information which the information storage part of the load distribution apparatus which concerns on the said 1st Embodiment stores. It is a figure explaining the load distribution operation | movement of the load distribution apparatus which concerns on the said 1st Embodiment. It is a figure explaining the load distribution operation | movement which processes the signal distributed to the high delay server of the load distribution apparatus which concerns on the said 1st Embodiment by a low delay server as a table | surface. It is a flowchart which shows the example of a load distribution sequence of the load distribution apparatus which concerns on the said 1st Embodiment. It is a figure explaining the example of a sequence of FIG. It is a figure which shows the example of calculation of the shoulder signal number calculated by the shoulder signal number calculation part of the load distribution apparatus which concerns on the said 1st Embodiment, the load fluctuation rate of a low-delay server, an average signal processing TAT fluctuation rate, and a score. . It is a block diagram which shows the distribution system with which the load distribution apparatus which concerns on the 2nd Embodiment of this invention is applied. It is a flowchart explaining the example of a sequence of the load distribution method of the load distribution apparatus which concerns on the said 2nd Embodiment. It is a figure explaining the example of a sequence of FIG. It is a figure which shows the example of calculation of the number of substitution signals calculated by the substitution signal number calculation part of the load distribution apparatus which concerns on the said 2nd Embodiment, the load fluctuation rate of a low-delay server, an average signal processing TAT fluctuation rate, and a score. . The number of substitution signals calculated by the substitution signal number calculation unit of the load distribution apparatus according to the second embodiment, the load fluctuation rate of the low-delay server, the average signal processing TAT fluctuation rate and score calculation, and the score C were used. It is a figure which shows the example of distribution by the number of shoulder substitute signals. It is a block diagram which shows the distribution system with which the load distribution apparatus of a nonpatent literature 2 is applied.

Hereinafter, a load balancer and the like in a mode for carrying out the present invention (hereinafter referred to as “the present embodiment”) will be described with reference to the drawings.
(Background explanation)
First, the background art will be described using Non-Patent Document 2.
FIG. 13 is a configuration diagram illustrating a distributed system to which the load distribution apparatus of Non-Patent Document 2 is applied.
As shown in FIG. 13, two data centers (DC: Data Center) 1 and 2 exist. The data center 1 (own site) and the data center 2 (other sites) can exist across regions and countries. For example, the data center 1 is based in Tokyo (Japan), and the data center 2 is based in Washington, DC (USA). The data center 1 and the data center 2 are connected by a network 3 having a wide band and a low delay. However, the data center 1 and the data center 2 have a signal-to-base delay.

In the present specification, the own site refers to a site where the load balancer 10 is placed. Here, the local base is the data center 1. The other base refers to a remote base other than the base where the load balancer 10 (self) is placed. Here, the other base is the data center 2.
In the data center 1, a plurality of (plural) processing servers 11 to 13 are accommodated, and each of the processing servers 11 to 13 is connected to the load distribution apparatus 10 via a data center network. In the data center 2, a plurality of processing servers 21 to 24 are connected by a data center network.
The data center 1 is configured by connecting clients (not shown) via a network.

  The load balancer 10 exists in the data center 1 (own site). The load balancer 10 distributes request signals (requests) from clients (not shown) to the processing servers 11 to 13 and 21 to 24 of the data center 1 or 2. Upon receiving a request from a client, the load distribution apparatus 10 selects one of the processing servers 11 to 13 and 21 to 24 of the data center 1 or the data center 2 and requests the selected processing server to perform processing. The selected processing server processes the application request from the client.

The load balancer 10 may be distributed processing middleware as an example of a program that operates in the distributed system, and is a load balancer that distributes processing to a plurality of processing servers 11 to 13 and 21 to 24 in the distributed system. Either is acceptable.
Moreover, the number of each apparatus which comprises the distributed system of FIG. 13 is not limited to the number illustrated in FIG. 13, It is good also as arbitrary numbers.

Non-Patent Document 2 describes a load distribution method that considers both the processing TAT of a request signal and the processing load of a processing server. The load distribution method described in Non-Patent Document 2 calculates the number of signals distributed to the low-delay server from the processing load and number of processing servers and the delay between the two sites.
In the load distribution method described in Non-Patent Document 2, two servers, a low communication delay server with a small communication delay (hereinafter referred to as a low delay server) and a high communication delay server with a large communication delay (hereinafter referred to as a high delay server), are used. It is divided into types and the number of signals to be distributed is calculated from a predetermined delay factor (the value obtained by dividing the delay time of the high delay server by the delay time of the low delay server). In the case of FIG. 13, among the processing servers 11 to 13 and 21 to 24 of the data center 1 or 2, the processing servers 11 and 12 of the data center 1 are low-latency servers, the processing server 13 of the data center 1 and the data center 2. The processing servers 21 to 24 are high-latency servers.

  By the way, as indicated by a symbol n in FIG. 13, the low-delay servers 11 and 12 and the high-delay server 13 may be mixed in the same base (data center 1 (own base)). Further, as indicated by a symbol o in FIG. 13, the TAT of the processing server may dynamically fluctuate at the base (data center 2). The dynamic change of the TAT of the processing server can specifically occur in the following state. (1) The TAT varies depending on the background processing of the processing server. (2) TAT fluctuates due to a specific route being blocked by a disaster or the like.

  In the load distribution method described in Non-Patent Document 2, when the TAT of the processing server changes dynamically, or when the low delay server and the high delay server coexist in the same base, the TAT changes. The processing load of each processing server cannot be accurately measured, and the number of signals to be distributed cannot be calculated appropriately.

  There is a need for a method that can appropriately calculate the number of signals to be distributed even when low-latency servers and high-latency servers coexist in the same base, or even when the TAT of a processing server at another base changes dynamically. It has been.

(First embodiment)
FIG. 1 is a configuration diagram showing a distributed system to which a load distribution apparatus according to a first embodiment of the present invention is applied. The same components as those in FIG. 13 are denoted by the same reference numerals.
As shown in FIG. 1, two data centers (DC) 1 and 2 exist. Here, an example will be described in which the data center 1 is based in Tokyo (Japan) and the data center 2 is based in Washington, DC (USA). The data center 1 and the data center 2 are connected by a broadband and low delay network 3.
In the data center 1, a plurality (multiple) of processing servers 11 to 17 are accommodated, and each of the processing servers 11 to 17 is connected to the load distribution apparatus 100 via a data center network. In the data center 2, a plurality of processing servers 21 to 24 are connected by a data center network.
The data center 1 is configured by connecting clients 31 to 34 via a network.
The clients 31 to 34 are user terminals or opposing systems, and request application processing to the processing servers 11 to 17 and 21 to 24 of the data centers 1 and 2.

  The load distribution apparatus 100 exists in the data center 1 (own site). The load distribution apparatus 100 distributes request signals (requests) from the clients 31 to 34 to the processing servers 11 to 17 and 21 to 24 of the data center 1 or the data center 2. In response to a request from the clients 31 to 34, the load distribution apparatus 100 selects one of the processing servers 11 to 17 and 21 to 24 of the data center 1 or the data center 2, and performs processing on the selected processing server. Ask. The selected processing server processes application requests from the clients 31-34.

  As described above, the load distribution apparatus 100 causes the processing servers 11 to 17 in which the load balancing apparatus 100 is installed or the processing servers 21 to 24 separated from the base to execute the information processing requested by the clients 31 to 34. Determine the server. The detailed configuration of the load distribution apparatus 100 will be described later.

The load distribution apparatus 100 may be distributed processing middleware as an example of a program that operates in the distributed system, or may be any load balancer that distributes processing to a plurality of processing servers 21 to 27 in the distributed system.
Further, the number of devices constituting the distributed system in FIG. 1 is not limited to the number illustrated in FIG. 1 and may be an arbitrary number.

Details of the load balancer 100 will be described.
The load distribution apparatus 100 distributes a signal to a processing server having as small a communication delay as possible while considering the processing load, the number of processing servers, and the communication delay. The load distribution apparatus 100 classifies each processing server into a low-delay server and a high-delay server based on a processing time measuring unit that measures a processing TAT (Turn Around Time) of each processing server and the measured processing TAT. It has a function as a classification means. The processing TAT includes a time required for transmitting a signal from the load balancer 100 to the processing server (communication delay time) and a signal processing time by the processing server.
As shown in FIG. 1, the load distribution apparatus 100 includes a distribution function unit 110 (distribution unit) having a substitution signal number calculation unit 111 (signal number calculation unit) and a server information acquisition unit 120 (load information acquisition unit). And an information storage unit 130 (storage means).

<Shoulder substitution signal number calculation unit 111>
The shoulder signal number calculation unit 111 calculates the values of the load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C (all described later) of the low-delay server for each shoulder signal number S, and stores them in the information storage unit 130. To do.
When executing the postscript (method 1), the shoulder substitution signal number calculation unit 111 assigns the number of low delay servers, the number of high delay servers, and the number of shoulder substitution signals S (postscript) to which signals to be distributed to the high delay servers are distributed to the low delay servers. ) To calculate the load fluctuation rate A of the processing server indicating the load fluctuation of the low-latency server. Here, the number of substitution signals S is n (n is a natural number), and the load fluctuation rate A of the low-delay server is obtained by sequentially applying from S = 0 to S = n to the formula (1) described later. The details of determining the substitution signal number S will be described in Example 1: (Method 1) (see FIGS. 3 and 4 to be described later).
Then, the substitution signal number calculation unit 111 is based on the acquired maximum value of the CPU usage rate (load information) and the load fluctuation rate A of the low-delay server in a range not exceeding the CPU limit value (load limit value). The number S of substitution signals to be processed by the low delay server is determined for the signal to be processed by the high delay server.

  In addition, when executing the postscript (method 2), the substitution signal number calculation unit 111 counts the number of low delay servers, the number of high delay servers, and the substitution signal number S that distributes signals to be distributed to the high delay servers to the low delay servers. Response time when the requests are distributed to each processing server in a predetermined order using the processing TAT (response time) when processed by the low delay server and the processing TAT when processed by the high delay server. An average signal processing TAT fluctuation rate B indicating fluctuation is calculated. Here, the number of substitution signals S is n (n is a natural number), and from S = 0 to S = n, the average signal processing TAT fluctuation rate B is obtained by sequentially applying to equation (2) described later.

  Then, the substitution signal number calculation unit 111 uses the load variation rate A and the average signal processing TAT variation rate B of the low delay server, and the substitution signal number S for processing the signal to be processed by the high delay server by the low delay server. The range of is calculated. Specifically, the substitution signal number calculation unit 111 refers to the base information 150 (see FIG. 2 to be described later) from the information storage unit 130, and determines the maximum value in the acquired load information and the load variation rate A of the low-delay server. Multiplication is performed to calculate the range of the substitution signal number S that does not exceed the CPU limit value (load limit value) of the low delay server set in advance, and the largest substitution signal number S in the calculated substitution signal number S range. The number S of shoulder substitute signals serving as a score for evaluating is determined. That is, the substitution signal number calculation unit 111 is a range of the maximum substitution signal number S that does not exceed the CPU limit value (load limit value) of the low-delay server even if the maximum value of the CPU usage rate of the low-delay server is multiplied by A. Is calculated, and the shoulder signal number S that gives the highest score in a range not exceeding the maximum shoulder signal number S is determined. The details of determining the substitution signal number S by calculating the score will be described later in Example 2: (Method 2) (see FIGS. 6 and 7).

  Here, when the load distribution apparatus 100 adopts round robin (one of static distribution methods) that distributes requests equally to each processing server, the average signal processing TAT fluctuation rate B is, for example, in one round of round robin. Average signal processing TAT variation rate B. In addition, it applies to the weighted round robin which changes the distribution ratio of requests according to the processing capacity of the processing server and equalizes the load of each processing server, and in this weighted round robin, the average signal processing TAT variation rate in one round A mode of calculating B may be used.

<Distribution function unit 110>
The distribution function unit 110 distributes the signal to the processing server according to the value of the shoulder signal number S calculated by the shoulder signal number calculation unit 111. Specifically, when the processing server that is the distribution destination is a high-delay server, the distribution function unit 110 distributes the signal originally distributed to the high-delay server by the number of substitution signals S to the low-delay server ( Example 1 (described later in (Method 1)).

In addition, the distribution function unit 110 obtains the maximum substitution signal number S corresponding to the maximum score in the range of the substitution signal number S that does not exceed the CPU limit value (load limit value) of the low delay server from the high delay server. Distribution to low-latency servers (Example 2: (described later in (Method 2)). Specifically, the distribution function unit 110 multiplies the maximum value of the acquired load information by the load fluctuation rate A of the low-delay server, calculates the range of the substitution signal number S that does not exceed the CPU threshold, and calculates Sorting is performed according to the value of the shoulder signal number S that gives the maximum score C within the range of the shoulder signal number S.
When the distribution function unit 110 receives request signals (requests) from the clients 31 to 34, the distribution function unit 110 determines a distribution destination processing server by a predetermined method (for example, a round robin method).

<Server information acquisition unit 120>
The server information acquisition unit 120 acquires the response time (TAT) of each processing server. In addition, the server information acquisition unit 120 periodically collects (acquires) load information (for example, CPU usage rate, memory usage rate, etc., taking CPU usage rate as an example in this embodiment) of each processing server. To do. The load information acquisition unit 120 stores the collected CPU usage rate (load information) in the information storage unit 130 and notifies the distribution function unit 110 thereof.

<Information storage unit 130>
The information storage unit 130 stores the CPU threshold, the response time (TAT) of each processing server, and the CPU usage rate of each processing server. The information storage unit 130 holds values of the load fluctuation rate A, the average signal processing TAT fluctuation rate B, and the score C (see FIG. 7) of the low-delay server for each substitution signal number S.
In addition, the information storage unit 130 includes a CPU limit value (load limit value) of the processing server at the local site, a processing delay time value between the data center 1 and the data center 2 (inter-base delay), and the number of processing servers at the local site (processing Number of servers), number of low-delay servers (number of low-delay servers) and number of high-delay servers (number of high-delay servers), processing server CPU usage (load information), etc. Stored as information 150 (see FIG. 2). The base information 150 held by the information storage unit 130 will be described later with reference to FIG.

FIG. 2 is a diagram illustrating the base information 150 stored in the information storage unit 130 of the load distribution apparatus 100 in FIG.
As shown in FIG. 2, the information storage unit 130 includes, as the base information 150, a CPU limit value (load limit value) of a low-delay server, an inter-base delay, a low-delay server number (the number of low-delay servers), and a high-delay server. Number (the number of high-delay servers) and usage rates of the low-delay servers 21 to 23 (CPU usage rate of the processing server 21, CPU usage rate of the processing server 22, and CPU usage rate of the processing server 23) are stored. By storing the CPU limit value, it is possible to determine the substitution signal number S from a comparison when the maximum value of the CPU usage rate is multiplied by A (load variation rate of the low delay server).

In the example of FIG. 2, the CPU limit value of the low-latency server is “80” (80% when the maximum CPU limit value is 100%).
The inter-base delay is “1000” (1000 [ms]).
The number of processing servers at the local site is “7” because there are seven processing servers 11 to 17 (see FIG. 1) in the local site (data center 1).
In addition, as a result of classification based on the measured processing time among the processing servers 11 to 17 at the local site, the number of low delay servers is “3” and the number of high delay servers is “4” (see FIG. 3).
The number of processing servers at other bases is “4” because there are four processing servers 21 to 24 (see FIG. 1) in the other bases (data center 2).
The CPU usage rate of the low-latency server 21 is “35” (35% when the maximum value of the CPU usage rate is 100%. The same applies hereinafter), and the CPU usage rate of the processing server 22 is “34”. The CPU usage rate of the server 23 is “34”.

Hereinafter, a load distribution method of the load distribution apparatus 100 configured as described above will be described.
[Principle explanation]
First, the basic concept of the present invention will be described.
The load balancer 100 measures the processing time (TAT) of each processing server, determines whether the measured processing time (TAT) of each processing server is larger or smaller than a preset threshold, and sets each processing server. Classify into high-latency servers and low-latency servers.
Here, the number of low-delay servers is Nm, the number of high-delay servers is Nn, and according to the round robin method, the number of signals per round for processing signals that are originally processed by the high-delay server by the low-delay server is S ( If 0 ≦ S ≦ Nn), the load fluctuation rate A of the low-latency server is expressed by the following equation (1).

  Load fluctuation rate A = (Nm + S) / Nm of the low delay server (1)

  Further, when the processing TAT (Turn Around Time) (response time) when processed by the low delay server is 1, and the processing TAT when processed by the high delay server is the communication delay factor D, the average signal in one round The processing TAT fluctuation rate B is expressed by the following equation (2).

  Average signal processing TAT fluctuation rate B = ((Nm + S) + D × (Nn−S)) / (Nm + D × Nn) (2)

  That is, the average signal processing TAT fluctuation rate B in one round is ((Nm + S) + D × (Nn−S)) / (Nm + D × Nn) times. D is a delay magnification that approximates a value obtained by dividing the average of the processing TAT of the high delay server by the average of the processing TAT of the low delay server.

  The load distribution apparatus 100 includes (Method 1) and (Method 2) as methods for determining the number S of substitution signals.

(Method 1): The substitution signal number S is determined using the load fluctuation rate A of the processing server at the local site.
The load distribution apparatus 100 acquires the CPU usage rate (load information) of each processing server at its own base, and multiplies the maximum value in the acquired load information by the load fluctuation rate A of the low delay server. The maximum substitution signal number S is calculated within a range not exceeding the CPU limit value (load limit value) of the processing server set in advance, and the signals distributed to the high delay server up to the calculated substitution signal number S are transmitted to the low delay server. To be processed.

(Method 2): The range of the maximum substitution signal number is calculated using the load fluctuation rate A of the low-delay server, and the substitution signal number S having the maximum score in the range of the maximum substitution signal number is determined.
The load distribution apparatus 100 acquires the CPU usage rate (load information) of each processing server at its own base, and multiplies the maximum value in the acquired load information by the load fluctuation rate A of the low delay server. A range of the substitution signal number S that does not exceed a preset load limit value of the processing server is calculated. The signal processed by the high delay server is processed by the low delay server using the value of the shoulder signal number S that is the maximum score in the calculated range of the shoulder signal number S. The score is a value represented by an evaluation function calculated using the load fluctuation rate A and the average signal processing TAT fluctuation rate B of the low-delay server. For example, as an example of the score, a decreasing function calculated from the load fluctuation rate A and the average signal processing TAT fluctuation rate B of the low-delay server is used as the evaluation function. This decrease function is a value (for example, (1 / A + 1 / B)) in which the score increases when the load fluctuation rate A or the average signal processing TAT fluctuation rate B of the low delay server decreases.

  After the shoulder signal number S is determined by the above (Method 1) or (Method 2), the load distribution apparatus 100 replaces the signal that was originally distributed to the high delay server (the signal that the high delay server was supposed to process). The number of signals is allocated to the low-delay server so that the low-delay server processes the signal. At this time, the load balancer 100 may select which processing server at its own site according to the distribution algorithm used by the load balancer 100 so far.

First, a specific operation example of the above (Method 1) will be described.
[Example 1: Operation example of (Method 1)]
The above (Method 1) provides a method for reducing the TAT by increasing the processing load of the low-latency server to the limit.
With reference to FIG. 3 and FIG. 4, the (method 1) operation of the load distribution method of the load distribution apparatus 100 will be described.
FIG. 3 is a diagram for explaining the load distribution operation of the load distribution apparatus 100. The same components as those in FIG. 1 are denoted by the same reference numerals.
As shown in FIG. 3, low delay servers 11 to 13 and high delay servers 14 to 17 are mixed in the processing servers 11 to 17 in the local base (data center 1). Arrows (thin solid line arrows, broken line arrows, thick solid arrows) extending from the load balancer 100 to the respective processing servers 11 to 17 indicate that the load balancer 100 distributes signals to the respective processing servers 11 to 17. For example, the load distribution apparatus 100 distributes request signals from the clients to the processing servers 11 to 13 and 15 to 17 as indicated by thin solid arrows (→ marks) in FIG.

  As described in the above (Method 1), the load distribution apparatus 100 calculates the shoulder signal number S using the load fluctuation rate A of the low delay server, and is distributed to the high delay server up to the calculated shoulder signal number S. The signals are distributed to the low-delay server by the number of signals S on the shoulder. As shown by the symbol a and the white arrow (⇒ mark) in FIG. 3, the signal (see the broken line arrow) that is originally distributed to the high delay server 14 is distributed to the low delay server 11 (see the thick solid arrow).

  FIG. 4 is a diagram illustrating the load distribution operation in which the low delay servers 11 to 13 process signals distributed to the high delay servers 14 to 17 by the load distribution apparatus 100 of FIG. FIG. 4 shows an example of distribution when the shoulder signal number S per round is 1. The vertical axis in FIG. 4 indicates signal sequence numbers 1 to 14 that are sequentially processed by the processing server, and the horizontal axis indicates processing servers 11 to 17 (low delay servers 11 to 13 and high delay servers 14 to 17). In FIG. 4, a solid white circle (◯ mark) indicates that the signal is distributed to the corresponding processing server. For example, a solid white circle (o mark) at a position where signal 1 and low-delay server 11 in FIG. 4 intersect indicates that signal 1 having a signal sequence number is distributed to low-delay server 11. Further, in FIG. 4, broken white circles (broken line circles) indicate that the signal originally distributed to the high delay server is distributed to the low delay server. For example, a solid white circle (o mark) at a position where the signal 4 and the low delay server 11 in FIG. 4 intersect each other indicates that the signal 4 having the signal sequence number originally assigned to the high delay server 14 is assigned to the low delay server 11. Is shown.

  As shown in FIG. 4, when the shoulder signal number S per round is 1, the shoulder signal number S is 1, so that one of the signals distributed to the high delay server per round is distributed to the low delay server. That is, as indicated by the symbol b in FIG. 4 and the white circle of the broken line (dashed circle mark), the signal that is originally distributed to the high delay server 14 (see FIG. 3) is distributed to the low delay server 11 (see FIG. 3). . In addition, as indicated by the symbol c in FIG. 4 and the white circle (dashed line circle), the signal that is originally distributed to the high delay server 15 (see FIG. 3) is distributed to the low delay server 12 (see FIG. 3). .

As shown in the symbol d in FIG. 4 and the dashed box, the effect of the distribution example when the shoulder substitution signal number S per round is 1 is as follows.
That is, by distributing one of the signals distributed to the high delay servers 14 to 17 to the low delay servers 11 to 13, the load on the low delay servers 11 to 13 is increased by 4/3. Further, the average signal processing TAT fluctuation rate B in one round is (4 + D × 3) / (3 + D × 4) times by substituting Nm “3” and Nn “4” into the equation (2). I understand. Since signals distributed to the high-delay servers 14 to 17 are distributed more to the low-delay servers 11 to 13, the signals can be acquired more quickly than when the load distribution operation is not performed, and more timely information can be obtained. With load distribution processing becomes possible.

Next, a specific operation example of the above (Method 2) will be described.
[Example 2: Operation example of (Method 2)]
The above (Method 2) provides a method for maximizing the score calculated from the processing load and TAT.
FIG. 5 is a diagram for explaining an example of a load distribution sequence in the (method 2) of the load distribution apparatus. The same components as those in FIG. 1 are denoted by the same reference numerals. FIG. 6 is a flowchart illustrating the sequence example of FIG.
5, the load distribution apparatus 100 sets the CPU threshold value and response time threshold value (threshold value for determining whether the server is a high delay server or a low delay server) for each of the processing servers 11 to 17. For example, the load distribution apparatus 100 sets various setting values (CPU threshold = 80, response time threshold = 10 ms).

  In step S102, the load distribution apparatus 100 periodically transmits test signals (such as ping) to all the processing servers 11 to 17 as processing time measuring means, and obtains response times (TAT) of the processing servers 11 to 17. (taking measurement. 6, the response time (TAT) “1.5 ms” of the processing server 11, the response time (TAT) “2.1 ms” of the processing server 12, and the response time (TAT) “of the processing server 13. 1.1 ms ", processing server 14 response time (TAT)" 10.1 ms ", processing server 15 response time (TAT)" 11.5 ms ", processing server 16 response time (TAT)" 23.5 ms ", The response time (TAT) “10.7 ms” of the processing server 17 is acquired.

  In step S103, the load distribution apparatus 100 compares the acquired response times (TAT) of the processing servers 11 to 17 with response time thresholds as classification means, and sets the processing servers 11 to 17 as high-delay servers 14 to 17. And low-latency servers 11-13. For example, as shown in FIG. 6, when the response time threshold is set to 10 ms, among the processing servers 11 to 17 at the local site, the processing servers 14 to 17 are high delay servers, and the processing servers 11 to 13 are low delay servers. Classify.

  In step S104, the server information acquisition unit 120 of the load distribution apparatus 100 periodically acquires load information (CPU usage rate) of the low-latency servers 11 to 13 (see FIG. 6). In the example of FIG. 6, the server information acquisition unit 120 has a CPU usage rate “35” of the processing server 11 that is a low-latency server, a CPU usage rate “34” of the processing server 12, and a CPU usage rate “34” of the processing server 13. To get.

  In step S <b> 105, the load distribution apparatus 100 stores the acquired information in the information storage unit 130 and notifies the distribution function unit 110.

  In step S106, the substitution signal number calculation unit 111 of the distribution function unit 110 extracts information from the information storage unit 130, and uses the number of low delay servers, the number of high delay servers, and the number of substitution signals S to calculate the number of processing server load fluctuations. A, average signal processing TAT fluctuation rate B, and score C are calculated, and the value of the substitution signal number S when the score C is maximized is determined. Specifically, the substitution signal number calculation unit 111 calculates the number of substitution signals for each substitution signal S based on the number of high delay servers, the number of low delay servers, the average response time of high delay servers, and the average response time of low delay servers. The load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C of the low-delay server are calculated (see FIG. 7). The score C is an evaluation value for evaluating the number S of substitute signals, and is a decreasing function calculated from the load fluctuation rate A and the average signal processing TAT fluctuation rate B of the low delay server (increases when A or B decreases). expressed.

  In step S107, the distribution function unit 110 of the load distribution apparatus 100 multiplies the acquired maximum value of the load information and the load fluctuation rate A of the low-delay server, and calculates the range of the substitution signal number S that does not exceed the CPU threshold value. Then, with the value of the substitution signal number S that gives the maximum score C in the calculated substitution signal number S range, the signal that is originally distributed to the high delay server is distributed to the low delay server. That is, the distribution function unit 110 transmits signals that are originally distributed to the high delay servers 11 to 14 by the amount corresponding to the maximum shoulder signal number S that has the highest score C within a range not exceeding the maximum shoulder signal number S. Sort to ~ 13.

Next, an example of signal distribution to the low-latency server will be described with reference to FIG.
FIG. 7 is a diagram illustrating a calculation example of the shoulder signal number S calculated by the shoulder signal number calculation unit 111, the load variation rate A, the average signal processing TAT variation rate B, and the score C of the low-delay server. The score C uses a decreasing function (1 / A + 1 / B).
The load fluctuation rate A of the low delay server is calculated by substituting Nm “3” into the equation (1), and the average signal processing TAT fluctuation rate B is calculated as Nm “3” and Nn in the equation (2). Calculated by substituting “4”. For example, when the number of substitution signals S is “1”, Nm “3” and S “1” are introduced into the equation (1) to calculate the load fluctuation rate A “1.33” of the low-delay server, Also, Nm “3”, Nn “4”, and S “1” are introduced into the equation (2) to calculate the average signal processing TAT fluctuation rate B “0.80”. The score for S “1” is calculated as “2.01” by substituting A “1.33” and B “0.80” into the decreasing function (1 / A + 1 / B). Is done. Similarly, in the case of S “2”, the load fluctuation rate A “1.67”, the average signal processing TAT fluctuation rate B “0.59”, and the score “2.29” of the low delay server are calculated. In the case of S “3”, the load fluctuation rate A “2”, the average signal processing TAT fluctuation rate B “0.39”, and the score “3.09” of the low delay server are calculated, and in the case of S “4” , The load fluctuation rate A “2.33”, the average signal processing TAT fluctuation rate B “0.18”, and the score “5.95” of the low delay server are calculated.

  In the case of this example, as shown in FIG. 6, the CPU usage rate (load information) of the acquired low-latency servers 11 to 13 is the CPU usage rate “35” of the processing server 11 and the CPU usage rate of the processing server 12. “34” is the CPU usage rate “34” of the processing server 13, and the maximum value of the CPU usage rate is “35” of the processing server 11. As shown in FIG. 2, the information storage unit 130 stores, as the base information 150, the CPU limit value “80” of the low delay server, the inter-base delay “1000”, the number of low delay servers “3”, and the high delay. The number of servers “4” is stored. Furthermore, the shoulder signal number calculation unit 111 calculates the shoulder signal number S, the load fluctuation rate A of the low delay server, the average signal processing TAT fluctuation rate B, and the score C shown in FIG.

  In this case, as shown by the symbol g in FIG. 7, the CPU limit value “80” is not exceeded even if the obtained maximum value “35” of the CPU usage rate is doubled. This is the range of the number of substitution signals. The distribution function unit 110 can distribute the three signals in the range of the maximum number of substitution signals to the low delay servers 11 to 13. Incidentally, a method of determining the maximum shoulder signal number S as it is as the shoulder signal number S corresponds to the above (Method 1).

  Further, as shown by the symbol h in FIG. 7, the highest score C in the range not exceeding the maximum shoulder signal number S is “3” because the shoulder signal number S is within the range of the shoulder signal number S “3”. The highest score “3.09” is selected, and S “3” indicated by the score “3.09” is determined. In this example, the distribution function unit 110 determines the number of substitution signals S “3” from the score “3.09”, and distributes the three signals to the low-delay servers 11 to 13.

In the example of FIG. 7, S “3” according to the method (method 1) of determining the shoulder signal number S using only the load fluctuation rate A of the low-delay server (see symbol g in FIG. 7), the maximum shoulder signal number. S "3" by the method (method 2) for determining the number S of substitution signals that gives the maximum score in the range of (2) (see symbol h in FIG. 7) has the same result, but the substitution to be decided The number S of signals may be different between the two. That is, depending on the score value, another shoulder signal number S having a higher score value may be determined from the range of the maximum shoulder signal number. When the shoulder signal number S differs between the two, priority is given to the determination of the shoulder signal number S by (Method 2).
In this way, in (Method 2), since the highest substitution signal number S is determined from the range of the maximum substitution signal number obtained in (Method 1), a more reliable substitution signal number S is obtained. be able to. As a result, since it is possible to distribute from the high-delay server to the low-delay server by the optimal number of substitution signals S, it is possible to further optimize the load distribution and further reduce the response time while limiting the processing load. It can be accelerated as much as possible.

  As described above, the load distribution apparatus 100 according to this embodiment includes the processing time measuring unit that measures the processing TAT of each processing server, and the processing servers are classified into the high delay server and the low delay server based on the measured processing TAT. And classifying means for classifying the data. Furthermore, the load distribution apparatus 100 includes an information storage unit 130 that stores a load limit value that is a limit value of processing of the low-delay server, a server information acquisition unit 120 that acquires load information of the low-delay server, Based on the number of servers, the number of low-latency servers, the maximum value of the acquired load information, and the load fluctuation rate A of the processing server that indicates the load fluctuations of the low-latency servers, A substitution signal number calculation unit 111 that calculates a substitution signal number S that distributes a signal to be distributed to a low delay server, a distribution function unit 110 that distributes the calculated substitution signal number S from a high delay server to a low delay server, Is provided.

  In the load distribution method of the load distribution apparatus 100, the processing time measurement step for measuring the processing TAT of each processing server, and the classification for classifying the processing servers into the high delay server and the low delay server based on the measured processing TAT. A step, a storage step for storing a load limit value that is a limit value of processing of the low delay server, a load information acquisition step for acquiring load information of the low delay server, the number of high delay servers, and the number of low delay servers And a signal to be distributed to the high-delay server based on the maximum value of the acquired load information and the load variation rate A of the processing server indicating the load variation of the low-delay server within a range not exceeding the load limit value A signal number calculating step for calculating the number of shoulder substitute signals S to be distributed to and a portion for distributing the calculated number of shoulder substitute signals S from the high delay server to the low delay server. And step, to run.

The disadvantages of the load balancing method of the existing technology load balancer are described.
The load balancing method of the load balancer of the existing technology includes “low load server priority” focusing on load distribution and “low delay server priority” focusing on average processing time.
(1) Low-load server priority The load on each processing server varies (the situation is that some servers are 100% but other servers are 0%, etc.). Load distribution distributes the load of each processing server evenly. The low-load server priority is assigned with priority to the processing server with the smallest CPU load. In the low load server priority, when the load of a server with a short TAT (low delay) increases, the server is distributed to a server with a long TAT (high delay). For this reason, the processing capability of the low-latency server cannot be used up. As a result, the average TAT can be high.

(2) Low-latency server priority In low-load server priority, priority is given to the server with the shortest TAT. Since the low load server priority does not consider the load on the server, the load on the server with a short TAT may temporarily increase. As a result, there is a possibility of being subject to CPU regulation.

The load distribution apparatus 100 according to the present embodiment has the following characteristics as compared with the existing technology.
The load distribution apparatus 100 according to the present embodiment performs distribution to return signals to be processed by the high-delay servers 14 to 17 (see FIG. 3) to the low-delay servers 11 to 13 within a range not exceeding the CPU limit value. Yes. In other words, since the load distribution apparatus 100 optimizes the signal distribution to the low-delay servers 14 to 17, the signal distribution to the high-delay servers 14 to 17 is further reduced, While preventing processing delay, it is possible to suppress inefficiency of hardware resources and unnecessary scaling by not applying an excessive processing load to the low-latency servers 11 to 13. As a result, the request signal processing TAT can be reduced while keeping the processing load of the processing server below a certain level. Both the request signal processing TAT and the processing load of the processing server can be satisfied.

(Second Embodiment)
FIG. 8 is a configuration diagram showing a distributed system to which the load distribution apparatus according to the second embodiment of the present invention is applied. The same components as those in FIG. 1 are denoted by the same reference numerals, and description of overlapping portions is omitted.
As illustrated in FIG. 8, the load distribution apparatus 200 includes a distribution function unit 210 having a substitution signal number calculation unit 211, a server information acquisition unit 120, and an information storage unit 130.

  The load distribution apparatus 200 measures the processing time of each processing server as processing time measurement means. Further, as the classifying means, the processing server is classified into a high delay server and a low delay server based on the measured processing time. In the present embodiment, the load distribution apparatus 200 increases each processing server as a low delay server in order of increasing processing TAT to the total number −1, and the number of low delay servers and the high delay in each case. Determine the number of servers.

The substitution signal number calculation unit 211 is based on the determined number of low-delay servers, the number of the low-delay servers, a process TAT when processed by the low-delay server, and a process TAT when processed by the high-delay server. The average signal processing TAT fluctuation rate B indicating the fluctuation of the processing response time when the request is distributed to each processing server in a predetermined order, and the load fluctuation rate A of the low delay server and the average signal processing TAT fluctuation rate Based on B, a score C for evaluating the shoulder signal number S is calculated.
The distribution function unit 210 multiplies the maximum value of the acquired load information by the load variation rate A of the low-delay server, calculates the range of the substitution signal number S that does not exceed the CPU threshold, and calculates the calculated substitution signal number S range The distribution is performed by the value of the shoulder signal number S that gives the maximum score C.

Hereinafter, a load distribution method of the load distribution apparatus 200 configured as described above will be described.
In the present embodiment, a signal is distributed to a processing server having a communication delay as small as possible while considering the processing delay, processing load, and number of processing servers.
The load distribution apparatus 200 measures the processing time (TAT) and load of each processing server, and sorts the processing servers in ascending order of TAT.
Then, the load distribution apparatus 200 first determines the number of low-delay servers in order of increasing TAT, and sets the other as the number of high-delay servers (total number N-1), and the load fluctuation rate of the low-delay server for each signal number S A, average signal processing TAT fluctuation rate B, and score C are calculated. Next, the number of low-delay servers is determined in order of increasing TAT, the other is the number of high-delay servers (N-2), and the load variation rate A and average signal processing TAT variation rate B of the low-delay server for each substitution signal number S , Score C is calculated. Similarly, the number of low-delay servers is increased to 1, 2, 3,... In order of increasing TAT, and the combinations of low-delay servers for every combination of low-delay servers and high-delay servers are assumed for each low-delay server number S. A load fluctuation rate A, an average signal processing TAT fluctuation rate B, and a score C are calculated.

Then, the load balancer 200 determines that the multiplication value of the load variation rate A and the load information (CPU usage rate) of the low-latency server exceeds the CPU threshold value among all combinations of the low-latency server count and the high-latency server count. The distribution is performed by the value of the shoulder signal number S that gives the maximum score C in a non-existing range.
The allocating function unit 110 multiplies the acquired maximum value of the load information by the load variation rate A of the low-delay server, calculates the range of the substitution signal number S that does not exceed the CPU threshold, and calculates the calculated substitution signal number S. The distribution is performed by the value of the shoulder signal number S that gives the maximum score C in the range.

FIG. 9 is a flowchart for explaining a sequence example of the load distribution method of the load distribution apparatus 200. FIG. 10 is a diagram illustrating the sequence example of FIG.
In step S201 of FIG. 9, the load distribution apparatus 200 sets the CPU threshold value of each of the processing servers 11-17. For example, the load balancer 200 sets CPU threshold = 80.

  In step S202 of FIG. 9, the load distribution apparatus 200 periodically transmits test signals (such as ping) to all the processing servers 11 to 17, and acquires (measures) the response times (TAT) of the processing servers 11 to 17. To do. As shown by the symbol i in FIG. 10, the response time (TAT) “1.1 ms” of the processing server 11, the response time (TAT) “1.5 ms” of the processing server 12, and the response time (TAT) “of the processing server 13. 2.1 ms ”, response time (TAT)“ 10.1 ms ”of the processing server 14, response time (TAT)“ 11.5 ms ”of the processing server 15, response time (TAT)“ 13.5 ms ”of the processing server 16, The response time (TAT) “118.7 ms” of the processing server 17 is acquired.

  In step S <b> 203, the server information acquisition unit 120 of the load distribution apparatus 200 periodically acquires load information (CPU usage rate) of all the processing servers 11 to 17. As shown by a symbol j in FIG. 10, the server information acquisition unit 120 includes a CPU usage rate “35” of the processing server 11, a CPU usage rate “34” of the processing server 12, a CPU usage rate “38” of the processing server 13, The CPU usage rate “45” of the processing server 14, the CPU usage rate “44” of the processing server 15, the CPU usage rate “31” of the processing server 16, and the CPU usage rate “24” of the processing server 17 are acquired.

In step S204, the substitution signal number calculation unit 211 of the distribution function unit 210 increases the number of low-delay servers to one, two,..., (Total number-1) in ascending order of response time (TAT). Then, the load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C of the low-delay server are calculated for each substitution signal number S in each case (see FIG. 11 described later).
In the example of FIG. 10, when the measurement results of the response times (TAT) of the processing servers 11 to 17 are viewed, the response time (TAT) of the processing server 11 is “1.1 ms” in ascending order of the response time (TAT). The response time (TAT) of the server 12 is “1.5 ms”, the response time (TAT) of the processing server 13 is “2.1 ms”, and so on. Therefore, first, among the processing servers 11 to 17, one of the processing servers 11 is a low-latency server, and six of the other processing servers 12 to 17 are high-latency servers. Load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C are calculated. Next, two processing servers 11 and 12 are low-delay servers, and five other processing servers 13 to 17 are high-delay servers. , Average signal processing TAT fluctuation rate B and score C are calculated.

  In the same manner, the low-delay servers are increased in ascending order of response time (TAT), (total number-1), that is, six processing servers 11 to 16 are set as low-delay servers, and one of the other processing servers 17 The value of the load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C of the low delay server for each substitution signal number S is calculated using the platform as a high delay server. As described above, the load variation rate A and the average signal processing TAT variation rate B of the low-delay server for each substitution signal number S for all combinations when the number of low-delay servers is changed in the order of decreasing response time (TAT). , The value of score C is calculated.

  In step S205, the distribution function unit 210 of the load distribution apparatus 200 performs distribution according to the value of the substitution signal number S when the score C is maximized. That is, the distribution function unit 210 distributes the signal that is originally distributed to the high-delay server to the low-delay server by the number of substitution signals S when the score C is maximized.

Next, an example of signal distribution to the low-latency server will be described with reference to FIG.
FIG. 11 is a diagram illustrating a calculation example of the shoulder signal number S calculated by the shoulder signal number calculation unit 211, the load variation rate A, the average signal processing TAT variation rate B, and the score C of the low-delay server. The score C uses a decreasing function (1 / A + 1 / B).
As shown by the symbol k in FIG. 11, the number of low-delay servers is increased, and in each case, the load variation rate A, average signal processing TAT variation rate B, and score C of the low-delay server for each substitution signal number S. Calculate the value. In the example of FIG. 11, when the number of low-delay servers is “1” and the number of high-delay servers is “6”, Nm “1” and S “0” are expressed by the above formula ( 1), the load fluctuation rate A “1” of the low-delay server is calculated, and Nm “1”, Nn “6”, and S “0” are introduced into the equation (2) to calculate the average signal processing TAT. The fluctuation rate B “1” is calculated. Also, the score “2” is calculated by substituting A “1” and B “1” into the decreasing function (1 / A + 1 / B) in the case of S “0”. Similarly, in the case of S “1”, the load fluctuation rate A “2”, the average signal processing TAT fluctuation rate B “0.86”, and the score “1.67” of the low delay server are calculated. In the case of “2”, the load fluctuation rate A “3”, average signal processing TAT fluctuation rate B “0.71”, and score “1.74” of the low delay server are calculated.
Then, as indicated by reference numeral 1 in FIG. 11, the range of the substitution signal number S that does not exceed the CPU threshold is calculated, and distribution is performed based on the substitution signal number S in which the score C is maximum.

Next, an example of signal distribution to the low-latency server will be described in more detail with reference to FIG.
FIG. 12 illustrates the calculation of the shoulder signal number S calculated by the shoulder signal number calculation unit 211, the load fluctuation rate A of the low delay server, the average signal processing TAT fluctuation rate B and the score C, and the number of shoulder signals using the score C. It is a figure which shows the example of distribution by S. The score C uses a decreasing function (1 / A + 1 / B). The method of calculating the load fluctuation rate A, average signal processing TAT fluctuation rate B, and score C of the low-delay server for each substitution signal number S has been described with reference to FIG.

  As shown in FIG. 12, in the combination of the number of low-delay servers and the number of high-delay servers, the load variation rate A, average signal processing TAT variation rate B of the low-delay server for each substitution signal number S in each combination, A value of score C is calculated. Then, as shown in FIG. 12, the distribution function unit 210 (see FIG. 8) of the load balancer 200 determines whether or not the CPU threshold is exceeded even if the maximum CPU value is multiplied by A among the low-latency servers. Determine. In the example of FIG. 12, for example, when the number of low-delay servers is “1” and the number of high-delay servers is “6”, when the number of substitution signals S is “0” and “1”, the CPU threshold value is exceeded and the substitution signal When the number S is “2” to “6”, the CPU threshold value is not exceeded. In the case of the number of low-delay servers “2” and the number of high-delay servers “5”, when the number of substitution signals S is “0”, “1”, and “2”, the CPU threshold is exceeded and the number of substitution signals When S “3” to “5”, the CPU threshold value is not exceeded. Further, when the number of low-delay servers is “3” and the number of high-delay servers is “4”, the CPU threshold is exceeded when the number of substitution signals S is “0”, “1”, “2”, and “3”. The CPU threshold value is not exceeded only when the number of substitution signals S is “4”. Hereinafter, when the number of low delay servers is “4” and the number of high delay servers is “3”, the number of low delay servers is “5” and the number of high delay servers is “2”, the number of low delay servers is “6”. When the number of high-delay servers is “1”, the CPU threshold value is exceeded regardless of the value of the substitution signal number S.

  Then, as shown by the symbol m in FIG. 12, among the low-delay servers, the maximum CPU value multiplied by A does not exceed the CPU threshold value, and the number of low-delay servers with the maximum score C and the high delay The distribution is performed by the combination of the number of servers and the number of signals S on the shoulder. The maximum score C is simply the score C “6.71” when the number of low-delay servers is “2” and the number of high-delay servers is “5” and the number of substitution signals S is “5”. There is. However, this combination is an example of distribution based on the number of substitution signals S “5” when the CPU threshold value exceeds 80. In other words, the processing time of the entire system increases because the distribution is performed when the CPU value of the processing server is exceeded. In the example of FIG. 12, signals that are originally distributed to the high-delay server are allocated to the low-delay servers by the number of low-delay servers “3”, the number of high-delay servers “4”, and the number of substitution signals S “3”.

  Here, in the first embodiment, the response time (TAT) of each processing server is compared with a response time threshold value, and each processing server is classified into a low delay server and a high delay server. On the other hand, in this embodiment, without setting a response time threshold as in the first embodiment, each processing server is increased as a low-delay server one by one in ascending order of processing TAT to the total number -1. Then, by determining the number of low latency servers and the number of high latency servers in each case, each processing server is classified into a low latency server and a high latency server. That is, in this embodiment, without using a response time threshold value that is a constant value, the low delay server and the high delay server that are optimal at that time are dynamically determined by looking at the overall processing delay and processing load of the processing server. Since classification is performed, more appropriate processing server classification can be performed. As a result, load distribution can be further optimized, and the response time can be shortened as much as possible while further limiting the processing load.

In each of the above embodiments, the score is calculated from the load fluctuation rate A and the average signal processing TAT fluctuation rate B of the low-delay server using a decreasing function (evaluation function). What is necessary is just to use, and the thing which does not calculate the score based on an evaluation function may be used. In other words, any method may be used as long as the shouldering signal number S is determined using the load fluctuation rate A and the average signal processing TAT fluctuation rate B of the low-delay server.
Further, only the average signal processing TAT fluctuation rate B may be used without using the load fluctuation rate A of the low delay server. In this case, the TAT fluctuation rate B target value is set in advance, and the shoulder signal number S can be obtained from the difference value from the actual measurement value or the like.

  In addition, the operation of distributing (returning) the signal distributed to the high-delay server to the low-delay server is an image of winding the signal from another site to its own site. In this respect, “sort (return)” to the low-latency server may be referred to as “winding”.

In addition, among the processes described in the above embodiments, all or a part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed All or a part of the above can be automatically performed by a known method. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above-described document and drawings can be arbitrarily changed unless otherwise specified.
Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

  Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Further, each of the above-described configurations, functions, and the like may be realized by software for interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function is stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), an IC (Integrated Circuit) card, an SD (Secure Digital) card, an optical disk, etc. It can be held on a recording medium. Further, in this specification, the processing steps describing time-series processing are not limited to processing performed in time series according to the described order, but are not necessarily performed in time series, either in parallel or individually. The processing (for example, parallel processing or object processing) is also included.

1 Data center (own site)
2 Data center (other bases)
11-17, 21-24 Processing server 31-34 Client 100, 200 Load balancer (processing time measuring means, classification means)
111, 211 Shoulder signal number calculation unit (signal number calculation means)
110, 210 Distribution function part (distribution means)
120 Server information acquisition unit (load information acquisition means)
130 Information storage unit (storage means)
150 Base information A Load variation rate of low-latency server B Average signal processing TAT variation rate S Number of shoulder signals C Score

Claims (7)

A load balancer that includes a plurality of processing servers in a base and determines the processing server that executes information processing requested by a client,
A processing time measuring means for measuring a processing TAT (Turn Around Time) of each of the processing servers;
Classification means for classifying the processing server into a high-latency server and a low-latency server based on the measured processing TAT;
Storage means for storing a load limit value which is a limit value of processing of the low-latency server;
Load information acquisition means for acquiring load information of the low-latency server;
Based on the number of the high-delay servers, the number of the low-delay servers, the maximum value of the acquired load information, and the load variation rate of the processing server indicating the load variation of the low-delay server, the load limit value is calculated. A signal number calculating means for calculating the number of substitution signals to be distributed to the low-delay server within a range not exceeding the signal to be distributed to the low-delay server;
Distribution means for distributing the calculated number of substitution signals from the high delay server to the low delay server;
A load balancer comprising: The signal number calculating means distributes the requests to the processing servers in a predetermined order based on a processing TAT when processed by the low-latency server and a processing TAT when processed by the high-latency server. Calculating the average signal processing TAT variation rate indicating the variation of the processing response time in the case,
Based on the load variation rate of the low-latency server and the average signal processing TAT variation rate, a score for evaluating the number of substitute signals is calculated,
The allocating means, based on the acquired maximum value of the load information and the load fluctuation rate of the low-latency server, from the range not exceeding the load limit value, the number of substitution signals serving as the maximum score The load distribution apparatus according to claim 1, wherein minutes are distributed from the high delay server to the low delay server. The load distribution apparatus according to claim 1, wherein the load information acquisition unit acquires a CPU usage rate of the low-latency server as load information. When the processing TAT when processed by the low delay server is 1, the processing TAT when processed by the high delay server is the average of the processing TAT of the high delay server by the average of the processing TAT of the low delay server. The load distribution apparatus according to claim 1, wherein the load distribution apparatus is approximated by a divided value. The allocating unit multiplies the acquired maximum value of the load information by the load variation rate of the low-latency server, calculates a range of the shoulder signal number that does not exceed the CPU threshold, and calculates the calculated shoulder signal number 2. The load distribution apparatus according to claim 1, wherein the distribution is performed based on the value of the number of substitution signals having the maximum score in the range. The classification means increases each processing server as a low-delay server in order from the smallest processing TAT to the total number -1, and the number of low-delay servers and the number of high-delay servers in each case And decide
The signal number calculating means includes the determined number of low-delay servers, the number of low-delay servers, a process TAT when processed by the low-delay server, and a process TAT when processed by the high-delay server; And calculating an average signal processing TAT variation rate indicating variation in processing response time when the requests are distributed to the processing servers in a predetermined order, and the load variation rate of the low delay server and the average signal Based on the processing TAT fluctuation rate, a score for evaluating the number of substitution signals is calculated,
The allocating means multiplies the maximum value of the acquired load information by the load fluctuation rate of the low-delay server, calculates the range of the substitution signal number not exceeding the CPU threshold value, and calculates the range of the calculated substitution signal number. The load distribution apparatus according to claim 1, wherein the distribution is performed based on the value of the number of substitution signals having the largest score. A load balancing method for a load balancer that includes a plurality of processing servers in a base and determines the processing server that executes information processing requested by a client,
The load balancer is:
A processing time measuring step of measuring a processing TAT (Turn Around Time) of each processing server;
A classification step of classifying the processing server into a high-latency server and a low-latency server based on the measured processing TAT;
A storage step of storing a load limit value that is a limit value of processing of the low-latency server;
A load information acquisition step of acquiring load information of the low-latency server;
Based on the number of the high-delay servers, the number of the low-delay servers, the maximum value of the acquired load information, and the load variation rate of the processing server indicating the load variation of the low-delay server, the load limit value is calculated. A signal number calculating step for calculating the number of substitution signals to be distributed to the low delay server within a range not exceeding the signal to be distributed to the low delay server;
A distribution step of distributing the calculated number of substitution signals from the high delay server to the low delay server;
Load balancing method to execute.

Images