JP2003150570A - Distributing method and distributing device for connection to server, and program and recording medium therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To minimize disk access, and to enhance the content distributing speed by maximally utilizing cashe on servers. SOLUTION: Single or a plurality of the same content groups are arranged in a plurality of servers 13, 14, and 15, and the connection is allocated to the respective servers to disperse a load. Time-out time after finally making access to the content is set with every content. When newly requiring the connection, if it is within the time-out time, the connection is distributed to the server to which the connection is finally distributed in distribution of the content. A data acquiring part 121 of a connection distributing device 12 acquires various data (such as round trip time, the present connection number, a CPU utilization factor, a disk I/O quantity, and total throughput) of the respective servers. A data storage part 122 stores the time and the server to which the connection is finally distributed. A calculating part 123 calculates the server to be distributed from recording.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for distributing connections to a server, which utilizes a cache on a server to minimize disk access and thereby increase content distribution speed. And its program and recording medium.

[0002]

2. Description of the Related Art When a huge file is loaded through the Internet, an image is viewed on a streaming magnetic tape device, or a Web page is browsed, the problem of concentration of distribution requests often occurs. To avoid this, a plurality of servers are prepared and connections are distributed. The distribution of connections to the server will be briefly described below with reference to FIG.

In FIG. 8, DNS (Domain Name) is used.
e System) 12 is made to function as a distribution device. The same www. ntt. co. The domain name jp is given. From client 11 to DNS 12 w
ww. ntt. co. In response to the inquiry that the IP address of jp is requested, the DNS (functions as a distribution device in this case) 12 compares the IP address with the client 1 by referring to the connection distribution policy.
Reply to 1. One client 11 has 192.168.1.
I sent back an answer of 1 and gave another client 11 '19
You will get a reply of 2.168.1.2.

By operating in this way, www. n
tt. co. jp connection to three servers 13,14,
It is possible to distribute the load to the server by distributing the load to 15 servers. Although the method of using the DNS 12 as a connection distribution device has been shown here, there is also a method of performing distribution by replacing the IP address of the request destination described in the packet requesting connection by the distribution device. Exists, but details are omitted.

Conventionally, the following techniques exist as methods of distributing connection servers. (Connection Round Robin Method) This method is a method of sequentially specifying servers to be distributed in the distribution device. For example, as shown in FIG. 8, three servers 1
When there are 3, 14, and 15, the first connection is assigned to the server 13, the next connection is assigned to the server 14, the next connection is assigned to the server 15, and the next connection is assigned to the server 13. In this method, the servers to which the connection is distributed are sequentially specified.

(Minimum number of connected servers selection method) This method is a method in which the distribution device distributes connections to the server having the smallest number of connections at present. For example, there are three servers 13, 14, 15 as shown in FIG.
The current number of connections is 4 for server 13 and 1 for server 1, respectively.
When 4 is 3 and the number of servers 15 is 2, the connection is distributed to the server 15 having the smallest current number of connections in response to the connection request generated at that time. In this case, the distribution device 12 grasps the number of connections.

(Server selection method based on land trip time) This method uses a ping in a distribution device.
It is a method of allocating to the server with the shortest response time when a connection request is made by using the active server response measurement method such as. Where ping is I
ICMP (Inter
net Control Message Proto
col) is used to send a reply request to the destination host. It is often used to confirm the connectivity between hosts connected by an IP-based network. In addition, there is also a function to check the turn-around time until a reply is returned, which can be used as a guide for the congestion level of the route to reach the other party.

In this example, ping is used as an active response measuring method. When there are three servers 13, 14, and 15 as shown in FIG. 8, the distribution device 12 receives a connection request and uses ping to each of the servers 13, 14, and 15 to send the request. The round trip time to the servers 13, 14, 15 is measured. As a result, the server with the shortest round trip time is determined as the server with the best response, and the connection is distributed to that server. Note that depending on the implementation, there is also a method in which an active server response measurement method such as ping is regularly used for measurement, and when a connection request occurs, a server is selected based on the latest measurement result. .

(Distribution method by server resource monitoring) In this method, in the distribution device, resources of the distribution destination server (for example, CPU utilization rate, disk I
/ O amount, current throughput), and any one or a combination of a plurality of them is used as a criterion for determination, and the connection is distributed to the server having the most resources.

[0010]

Normally, in the server, the data once read from the disk is stored in the cache area on the memory. These data remain in the cache area until they are deleted from the cache area by reading new data from the disk. By doing so, it is possible to avoid reading and accessing a disk with a slow access speed as much as possible (a cache function). When content is frequently accessed in content distribution, the disk is likely to become a system bottleneck. However, the conventional method of allocating connections to a plurality of servers does not consider effective use of the cache, and as a result, it cannot be said that the performance can be sufficiently exerted in content distribution.

FIG. 9 is a diagram showing an image of data and a cache on a disk in the conventional method. FIG. 9 shows a specific example in which the performance cannot be exhibited when the distribution is performed using the conventional technique. That is, in FIG. 9, the arrow indicates the file read position. Data is read out as these arrows progress from left to right. The horizontal line area is an area that is not in the cache, and the gray area is an area that is in the cache. When there is no data in the cache, the data is read from the disk, and when there is data in the cache, the data arranged in the cache is read without accessing the disk. In the example of FIG. 9, a total of 6 connections are distributed to 3 servers, 2 connections each, but it can be seen that none of the connections read data from the cache. Since the conventional distribution method does not take into account the effective use of the cache, such a situation is likely to occur.

Therefore, an object of the present invention is to solve these conventional problems and to utilize a cache on the server to minimize the disk access, thereby making it possible to improve the content delivery speed. To provide a connection distribution method and a distribution device, a program thereof, and a recording medium.

[0013]

In order to achieve the above object, a method of distributing connections to servers according to the present invention is to arrange a single or a plurality of identical contents groups in a plurality of servers and connect the connections to each server. It is a system that distributes the load by allocating it, and sets a timeout time after the last access to that content for each content, and when a new connection request occurs, that timeout If it is within the time, the connection will be assigned to the server that last gave the connection for the delivery of the content. (This method works effectively when the same content is placed on multiple servers.)

Further, when it is not possible to distribute to a server having a cache, that is, when there is no server within the timeout time, an active server response time measuring method such as ping which is one of the conventional distributing methods. Use to distribute to the server with the shortest response time. Also, when it is not possible to allocate to the server that has the cache, that is, when there is no server within the timeout time,
Connections are distributed using a method of selecting a server with the smallest number of connections, which is one of conventional distribution methods. Furthermore, it is one of the conventional distribution methods when it cannot be distributed to the server that has the cache, that is, when there is no server within the timeout time,
Allocate connections using the method of measuring server resources and selecting the server with the most resources.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings. (Embodiment of claim 1) FIG. 4 is a diagram showing the relationship between the data transfer rate of the disk and the memory and the access time.
In the content distribution server, there is a clear difference between the speed of reading the data from the disk and the speed of reading the data arranged in the software cache. That is, as shown in FIG. 4, when the disc is in the best state, the access time is 0 msec, and the read speed is 43.3 Mbyte / sec. In the worst state of the disc, the access time is 26.4 msec, which is considerably slow. And the reading speed is 2.6 Mbyte / s.
ec and low speed. On the other hand, the access time of the memory including the cache is 0.006 msec, and the read speed is 938 Mbyte / sec, which is extremely fast.

According to the first aspect of the present invention, a system is provided in which a single or a plurality of identical content groups are arranged in a plurality of servers and the connections are allocated to the respective servers, thereby distributing the load. Set the timeout time since the last access to that content, and if a new connection request occurs, if the connection time is within that timeout time, the connection is distributed last for the delivery of the content. Try to shake the connection to the server.

According to the present invention, the distribution device for a server selects a server having a high possibility that data is placed in the cache, thereby improving the delivery performance of contents and the like. By using the method according to the first aspect, it is possible to select a server that is more likely to have data arranged in the cache, so that the frequency of access to the disk can be reduced. From the results of FIG. 4, there is a clear difference in speed between the case of reading data from the cache and the case of reading data from the disk, so the present invention is an extremely effective technique.

FIG. 1 is a diagram of an experimental environment relating to distribution of connections to servers showing an embodiment of the present invention, and FIG.
It is a figure which shows the performance difference of the method of Claim 1, and the conventional method. As shown in FIG. 1, the client 11 is connected to the switch 12A, and the switch 12A is connected to the three servers 1.
It was connected to 3,14,15. The client 11 and the switch 12A are connected by Giga Ethernet (registered trademark), and the servers 13, 14, 15 and the switch 12A are connected by 1
It is connected by 00 Mega Fast Ethernet (registered trademark). Eight 1 Gbyte contents are arranged in each of the servers 13, 14, and 15. Also, client 1
The connection request generation intervals from 1 to the servers 13, 14 and 15 are generated according to the exponential distribution of λ = 0.1, and the content to be connected is selected at random. For example, the first connection is Content 1, then some time later Content 3,
Next, the connection request is generated in the order of the content 1 and then the content 7.

In this experiment, since there is no distribution device, the client 11 behaves as if it has a connection device. Specifically, create a connection request list in advance that describes the connection request time and the content of the connection destination,
As to which server each connection request is distributed to, the method of claim 1, the method of selecting the one with the shortest conventional round trip time, the method of selecting the server with the fewest number of connections, and the method using round robin. As a result, a list was created in which the time of the connection request, the content of the connection destination, and the server of the connection destination were described, and the connection was generated according to the list. Since the switch 12A is in the experimental stage, the connection is manually switched based on the calculation result.

As objects of comparison and examination, a method based on round robin, a method for connecting to a server with the minimum number of connections,
The method of measuring the round trip time using ping and selecting the best one was used. In addition, in the experiment of the present invention, the timeout time for distribution was set to 30 seconds. The evaluation is performed by the total throughput of the client 11. The total throughput refers to the total amount of data obtained by the client machine per unit time. From the results shown in FIG. 2, it became clear that the method of claim 1 is effective.

FIG. 3 is a block diagram of a connection distribution device showing an embodiment of the present invention. The connection distribution device 12 sends a connection request from the client 11 to the server 1
It has the function of distributing to 3, 14, and 15. This device 12
Includes a data acquisition unit 121, a data storage unit 122, and a calculation unit 123. The data acquisition unit 121 acquires various data (round trip time, current number of connections, CPU utilization rate, disk I / O amount, total throughput) of the distribution destination server. The data storage unit 122 stores the server to which the connection was last assigned for each content and the time at that time. The calculation unit 123 calculates the server to be distributed from the data acquired by the data acquisition unit 121 and the connection record accumulated in the data storage unit 122.

(Operations of Claims 1 to 5) FIG. 5 is a processing flowchart of the connection distribution device in FIG. When the device is activated, the data acquisition unit 121 detects a new connection request from the client 11 (step 101),
Whether or not there is a server that has delivered the same content within the timeout time is determined (step 102), and if there is, the connection is distributed to the server to which the last connection was made (step 103). If there is no server to which the same content has been distributed, the server to be connected by round robin is selected. Alternatively, the server with the shortest round trip time is selected (claim 3). Alternatively, the server with the smallest current number of connections is selected (claim 4).
Alternatively, the server having the most available resources is selected (claim 5) (step 104). This step 10
In 4, it is assumed that any one of the above processes 1 to 4 is performed.

(Embodiment of Claim 2) Claim 2 shows a method for quantitatively calculating a timeout time required when the method of Claim 1 is used. Since the algorithm for allocating cache data in the cache area is FIFO, it is possible to estimate the time when the data continues to exist in the cache area, and the time since the last access to that content. By paying attention to time, it becomes possible to select a server to be distributed.
If there is no connection that is reading data located in the cache, the total throughput of the server is T (M
bps) and the cacheable area is Mc (Mbyte), the cache replacement time Tc (sec)
Is expressed by the following equation. Tc = 8 Mc / T (sec) (1) According to claim 2, the dynamic change necessary when using the invention of claim 1 It will be possible to estimate the time at which the cached data will be overwritten.

FIG. 6 is an operational flowchart of the processing according to claim 2. Calculation unit 12 of connection distribution device 12
When the server 3 is activated, the server 3 measures the throughput of the server (step 201) and sets the timeout time by referring to the formula (1) of claim 2 (step 202). This operation is repeated at a certain fixed cycle.

(Embodiment of claim 3) In claim 3, in the method of claim 1, when the method of claim 1 is unsuccessful, a conventional round trip time method is additionally performed. Is. That is, when it is not possible to distribute to the server having the cache, that is, when there is no server within the timeout time, an active server response time measuring method such as ping, which is one of the conventional allocation methods, is used. The server with the shortest response time.

For example, as shown in FIG. 3, when three servers 13, 14 and 15 are connected to the connection distribution device 12, the connection distribution device 12 receives a connection request from the client 11, Each server 13,
A ping is used for 14 and 15 to measure the round trip time to each server 13, 14, and 15. That is, the ping can specify a host name or an IP address, hit a packet, and check the response time to the packet. Thereby, the round trip time can be measured. What is round trip time?
The time it takes for a computer to send a packet to another computer and get a response back. The data acquisition unit 121 acquires the measurement result, acquires the measurement result, and transfers the measurement result to the data storage unit 122. The data storage unit 122 stores the sent measurement results and also stores the server to which the connection was last assigned for each content and the time at that time. The calculation unit 123 calculates from the data stored in the data storage unit 122 using a calculation formula, determines that the server with the shortest round trip time is the server with the best response, and allocates the connection to that server.

(Embodiment of claim 4) According to the method of claim 4, when it is not possible to allocate to the server having the cache, that is, when there is no server within the timeout time, the conventional allocation method is used. Connections are distributed using a method of selecting the server with the smallest number of connections, which is one. FIG. 7 is an explanatory diagram of the processing showing the embodiment of claim 4. This is a method of combining the method of claim 1 and the method of allocating the connection to the server with the smallest number of connections in the prior art, and FIG. 7 shows a specific example of the actual processing in the method of claim 4.

The time-out time required when using the method of claim 1 was estimated from the data to be 30 seconds. First, connect from the client ((1) Content 1
Request to connect to) is coming. In the initial state, there is no connection to any server, so the connection is distributed to the server 1. Next, after 10 seconds, a connection ((2) connection request to the content 2) is received, but since the distribution of the content 2 is the first time, there is no server to be distributed preferentially. After 10 seconds, a connection comes (a connection request to (1) content 1). The distribution of the content 1 is performed by connection ((1) request for connection to the content 1), and 20 seconds have elapsed from the start of the distribution. Since the timeout time is 30 seconds, it is considered that there is still data to be distributed in the cache of the server 1, so the connection ((3) connection request to the content 1) is distributed to the server 1.

Connect 10 seconds later ((4) content 4
Connection request to the server 3 is received, but since it is the first time to deliver the content 4, the server 3 with the smallest number of connections
Allocate connections to. A connection (connection request to (5) content 4) comes 10 seconds later, but since content 4 was distributed to server 3 10 seconds ago, connection to server 3 (connection request to (5) content 4) ). Connect 20 seconds later ((6) Request to connect to content 1)
But content 1 was last delivered 5
Since it is 0 second ago and the time-out period has passed, there is no data to be distributed in the cache of the server 1, so the connection is distributed to the server 2 with the smallest number of connections. In this way, each server 1,
It means that two are connected to each of 2 and 3.

Next, an example of a method for the data acquisition unit 121 of the connection distribution device 12 to acquire the data regarding the current number of connections of the server will be described. There are a plurality of methods for acquiring the data of the number of connected servers, but the method mainly used is the following method. (A) SNMP (Simple Network Ma)
It is possible to centrally acquire and manage various information of the server by using the management protocol). (B) Method of independently constructing management system: SNM mentioned above
If management using P is not possible (for example, SN
In the case of low efficiency in MP), there is a method in which a daemon that reports resources is arranged in each server and a program that periodically reports the number of connections acquires necessary data. This is because when an inquiry is sent from the distribution device 12 to the server resource monitoring program of the server, this program returns a reply to the distribution device 12 to the inquiry.

(Embodiment of claim 5) According to claim 5, in the method of claim 1, when it is not possible to distribute to a server having a cache, that is, when there is no server within the timeout time, One of the allocation methods is to measure the server resource and allocate the connection using the method of selecting the server with the most available resources. As shown in FIG. 3, in the connection distribution device 12, the data acquisition unit 121 determines that the distribution destination server 13,
14 or 15 resources (for example, CPU utilization rate, disk I / O amount, current throughput, etc.) are monitored, and one of them is used as a criterion for connection to the server with the most available resources. Sort. Data acquisition unit 12
1 monitors the CPU utilization rate, the disk I / O amount, the current throughput, etc., and acquires information. Data storage unit 1
22 stores the data acquired by the data acquisition unit 121, and also stores the server to which the connection was last assigned for each content and the time at that time. The calculation unit 123
The server to be distributed is calculated from the connection records accumulated in the data storage unit 122, and the connection is distributed.

An example of a method in which the data acquisition unit 121 acquires data relating to the CPU utilization rate, disk I / O amount, current throughput, etc. will be described. Similarly to the case described in claim 4, the CPU utilization rate of the server and the disk I / O
Although there are a plurality of methods for acquiring data such as quantity and throughput, the following method is mainly used among them. (A) SNMP (Simple Network Ma)
It is possible to centrally acquire and manage various information of the server by using the management protocol). (B)
Method of independently constructing management system: When management using the above-mentioned SNMP is impossible (for example, when SNMP is inefficient), a daemon for reporting resources to each server is arranged, and CPU utilization or There is a method of acquiring required data by a program that regularly reports the disk I / O amount and total throughput. This is because when an inquiry is sent from the distribution device 12 to the server resource monitoring program of the server, this program returns a reply to the distribution device 12 to the inquiry.

(Program and its recording medium) The processing flow of the connection distribution device of FIG. 5 or the processing flow of claim 2 of FIG. 6 is converted into a computer-executable program, and the converted program is a CD-RO.
It is stored in a recording medium such as M. Thus, the present invention can be easily realized by mounting the recording medium on a DNS or other computer and installing the program in the computer or downloading the program to another computer via a network to execute the program. can do.

[0034]

As described above, according to the present invention,
By utilizing the cache on the server as much as possible, the disk access can be minimized, and as a result, the content delivery speed can be increased.

[Brief description of drawings]

FIG. 1 is a block diagram (claim 1) of a device for allocating connections to a server according to an embodiment of the present invention.

FIG. 2 is a diagram showing a performance difference between the method of the present invention (claim 1) and the conventional method.

FIG. 3 is a block configuration diagram of a connection distribution device in FIG.

FIG. 4 is a diagram showing a data transfer rate and an access time of a disk and a memory.

5 is a processing flowchart of the connection distribution device in FIG.

FIG. 6 is a processing flowchart (claim 2) showing a method of setting a timeout time according to an embodiment of the present invention.

FIG. 7 is an explanatory diagram (claim 4) showing an example of actual processing using the present invention.

FIG. 8 is a diagram of distribution load distribution by a plurality of servers in the related art.

FIG. 9 is a diagram showing an image of data on a disk and a cache in a conventional method.

[Explanation of symbols]

11 ... Client, 12A ... Switch, 13, 14,
15 ... Server, 12 ... Connection distribution device, 121 ... Data acquisition unit, 122 ... Data storage unit, 123 ... Calculation unit.

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Katsunori Ori             2-3-1, Otemachi, Chiyoda-ku, Tokyo             Inside Telegraph and Telephone Corporation F term (reference) 5B045 BB28 BB48 GG01 GG04                 5B082 HA02 HA05 HA08                 5B089 GA11 GA21 HA10 JB22 KA05                       KA06

Claims (8)

[Claims] 1. A method of allocating connections to servers for distributing loads by allocating a single or a plurality of identical content groups to a plurality of servers and allocating the connections to the respective servers. Set the timeout time after the access to the content, and if a new connection request arrives, if it is within the above timeout time, the server that last connected to the delivery of the content A method of allocating connections to a server, characterized in that the connections are categorized to each server. 2. The method of allocating connections to servers according to claim 1, wherein when the timeout time is quantitatively calculated, the total server count is set in a state where there is no connection reading data arranged in the cache. Throughput is T (Mbp
s), assuming that the cacheable area is Mc (Mbyte), the time Tc (sec) for changing the cache is T
Since it can be calculated by c = 8 Mc / T (sec), it can be estimated by the time Tc since the last access to the content, and a server characterized by selecting a server to be distributed based on this Connection distribution method. 3. The method of distributing the connection to the server according to claim 1, wherein when there is no server to be distributed within the timeout time, an active server response time measuring method such as ping is used to obtain the most response time. A method of allocating connections to servers characterized by allocating to short servers. 4. The method of distributing a connection to a server according to claim 1, wherein when there is no server to be distributed within the timeout time, the connection to the server having the smallest number of connections is made. Distribution method. 5. The method of allocating a connection to a server according to claim 1, wherein when there is no server to be distributed within the timeout time, the server resource is measured and the server having the most resource is selected. The method for allocating connections to the server is characterized by allocating connections to the server. 6. A connection distribution device for distributing a connection request to a server, wherein various data including one of a round trip time of a distribution destination server, a current number of connections, a CPU utilization rate, a disk I / O amount, and a total throughput. Data acquisition means for acquiring, a server that lastly distributes connections for each content and a data storage means that stores the time of allocation to the server, data acquired by the data acquisition means, and connections stored in the data storage means From the record,
A connection distribution device, comprising: a calculating unit that calculates a server to be distributed. 7. A connection distribution program for causing a computer to execute the procedure of the method for distributing the connection to the server according to any one of claims 1 to 5. 8. A computer-readable recording medium on which the connection distribution program according to claim 7 is stored.