JP3510623B2 - Network server allocation device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to allocation of server resources.

[0002]

2. Description of the Related Art In recent years, with the rapid spread of the Internet / Intranet, efficient use of network service servers and service stability have been demanded.
Optimal allocation of services to servers is indispensable for efficient use of servers and stable supply of services, and it is necessary to accurately recognize the load on the servers.

The following methods are known as conventional server load recognition methods. (1). Agent method This is a method to put a program to measure the resource usage rate of CPU, memory, etc. on the server. The agent interferes with the load measurement accuracy, such as consumption. Also, since the agent program has to be installed on the server, there is a problem that the instruction cost is high because of lack of versatility. (2). Load measurement communication method This is a method of pinging the server and simulating service communication to obtain the server load from the response time, but the communication for measurement consumes the bandwidth of the route, Since the server also bears the load for the response, it interferes with the load measurement. In addition, there is a problem that the server needs to support the protocol used for measurement and the like and lacks versatility. (3) .Number of VCs, connection time, connection frequency, connection error rate, response time measurement These are VCs to the server measured at the time of relay on the device that relays packets from the client to the server.
This method calculates the server load from the number, connection time, connection frequency, connection error rate, and response time, but the error is large because it is based on the behavior of the server during connection. Since it requires a large number of connections to improve the accuracy, it is not suitable for services that perform a large amount of communication with few connections. Further, since the relay is essential, there is a problem that the throughput of the server is limited by the throughput of the measuring device. (4). Hit count and hit ratio calculation method The packets to the WWW server etc. are checked, the access count (hit count) and access frequency (hit ratio) are measured for each content such as the file to be accessed, and the server is calculated from the results. Although this is a method of obtaining the load, it cannot support a new service because it requires packet analysis processing for each protocol in order to specify the access target file. In addition, the server performance must be known. The only way to give server performance in advance is to obtain the server performance from the catalog value and experience, but since the server performance is greatly influenced by the system configuration and operation form, the catalog performance value based on the standard configuration and form is not accurate. However, there is a problem in that it is impossible to avoid at least one failure when empirically requested. As described above, none of the methods can detect the server load quickly and efficiently without burdening the server.

Further, since the load on the server cannot be recognized accurately, it is difficult to allocate the service provided by the server. The following methods have been proposed only from the viewpoint of service allocation. (5). Round robin DNS method In the DNS (Domain Name System) service, the entry table is set so that one domain name corresponds to the IP addresses of multiple servers, and the client IP address inquiry request is made. On the other hand, each server is allocated cyclically (round robin) according to the entry table, and the IP address of the allocated server is selected to respond to the client, whereby the service is distributed to a plurality of servers.

However, in this round-robin DNS method, the service distribution rate can be equal or only a simple ratio, and each server is assigned regardless of its capacity or dynamic load condition. Since the service must be provided according to the distribution rate, there was a difference in the load status of each server, resulting in inefficiency as a whole. In addition, since the DNS inquiry information is usually cached on the client side, even if the ratio is changed, it cannot be immediately reflected. (6). Distributing method using hash table In this method, the entries of the hash table that manages connections are allocated to the servers, and the services are distributed to the servers at a ratio according to the number of allocated entries.

In this method, first, when a service request is made from a client, an entry is determined from the client address and service, and the request is sent to the server to which the entry is assigned. Then, since the number of services according to the ratio of the allocated number of entries is distributed to each server, a large number of entries are allocated to the high-performance server, or the allocated entries to the server with a heavy load are not so heavy in load. Efficient use of the server is realized by reallocating it to the server.

However, in the distribution method using this hash table, a hash function that generates a biased hash value is necessary to correctly reflect the ratio of the number of hash entries to the service distribution ratio. It is impossible to find a hash function that produces an unbiased hash value for any distribution of keys (such as client addresses or port numbers). Further, since the accuracy of the distribution rate is proportional to the number of hash entries, a large number of entries are required to improve the accuracy, and a large amount of storage resources (buffer) will be consumed. As a result, the storage resources (buffers) that can be used for connection management are reduced, and there is a problem that a large amount of access cannot be handled. (7) .Distribution method according to server status and performance Measure the response time by pinging the server, or relay the packet from the client and measure the connection time, connection error rate, etc. at the time of relaying the server. It is a method of distributing the amount of service according to the load and the performance ratio by predicting the level of the load of the server or predicting the performance ratio between the servers.

However, in this method, since the service to any client is equally distributed to the server regardless of the processing capacity of the client, the length of the path to the client, the bandwidth, etc., the utilization efficiency of the server can be maximized. There wasn't.

For example, for a client whose route is a bottleneck due to a narrow or long route bandwidth or a client with low processing capability, the difference in server performance (especially speed) or load does not appear in the quality of service.

On the other hand, for a client connected by a near or high-speed line or a client having a high processing capability, the difference in server performance or the load greatly affects the service quality. Therefore, if the services to all the clients are to be equally distributed, there is a problem that the clients have no choice but to allocate more server resources than necessary or allocate insufficient server resources.

As described above, both the server load recognition method and the server allocation method in the prior art have problems.

The present invention has been proposed to solve the above problems, and an object thereof is to recognize the load of a server quickly and efficiently without burdening the server, and
By distributing the service according to the dynamic load situation in the server, accurately reflecting the service distribution ratio obtained by setting and adjustment in the service distribution, and distributing the service by estimating the necessary server resources for each client. It is about maximizing server utilization efficiency.

[0012]

SUMMARY OF THE INVENTION The present invention uses a network for data.
Transfer from client to multiple servers via network
Data sent from the client in the device
Relaying means for changing the destination and transferring to one of the servers
And the correspondence between the data and the server is retained and the destination is sent to the relay means.
Connection management means for instructing the
Measures and calculates the processing capacity on the client side, and
Data and server using functions according to service distribution ratio
Server allocation means that determines the correspondence of and notifies the connection management means
In the server allocation means,
The client side processing capacity is different from the probability distribution according to the force.
The lower the correction, the closer the distribution is to a uniform distribution.
Network server allocating device with accurate rate distribution as distribution rate
Is. The present invention also allows data to be transmitted over a network.
On a device that transfers data from a client to multiple servers
And change the destination of the data sent from the client
Relay means to transfer to one of the servers, data and server
A connection pipe that retains the correspondence relationship of the bar and instructs the destination to the relay means
Management means, server processing capacity and client side processing capacity
Measure the force and calculate the service distribution rate
The correspondence between the data and server is determined using the
Server allocation means for notifying the connection management means,
To the client currently in service,
For the client side processing power distribution,
Continued, the lower the client-side processing power is for the distribution, the more
The probability distribution according to the processing capacity of the server close to a uniform distribution,
On the contrary, the higher the number, the more the correction is made to emphasize the processing capacity of each server.
Network with the distribution probability
Server allocation device. The present invention also includes data
Client to multiple servers via network.
In the sending device, the data sent from the client
Relay device that transfers the data to one of the servers by changing the destination
And the correspondence between the data and the server is retained and the destination is sent to the relay means.
Connection management means for instructing the
Measures and calculates the processing capacity on the client side, and
Data and server using functions according to service distribution ratio
Server allocation means that determines the correspondence of and notifies the connection management means
Currently in service in the server allocating means including
Minutes of client-side processing power for each client
Demand for cloth, new connection client side processing capacity to distribution
The lower the value, the better the probability distribution according to the processing capacity of the server.
Like distribution, and conversely the higher the processing power of each server
Make corrections to make them stand, find the correction probability distribution, and calculate the distribution ratio.
And has a plurality of the server allocation means,
This method distributes services to each service and each client.
Clients and services can be switched so that policies can be switched.
A network server allocation device selected for each server.

The first means is a device for transferring data from a client to a plurality of servers via a network, and a relay means for transferring the data transmitted from the client to one of the servers while changing the destination, and the data. The connection management means that maintains the correspondence relationship between the servers and indicates the destination to the relay means, and the processing capacity of the server, the client, and the route are measured and obtained, and the function according to the service distribution ratio based on the measurement is used to A network server allocating device comprising a server allocating unit that determines correspondence and notifies the connection managing unit.

Since the service is distributed according to the method obtained by measuring the performance / load of the server and the performance / load of the client side, it is possible to automatically respond to the dynamic change of the load condition of the server, and further, from the client. As many servers as necessary to maintain the visible quality of service can be allocated, which has the effect of maximizing server utilization efficiency. Furthermore, since the server allocation is determined using the function, there is an effect that the service distribution rate can be accurately reflected in the distribution. Furthermore, a sufficient effect can be obtained with only a single connection management means.

The second means is the server allocation means in said first means, with respect to the probability distribution according to the server's processing capability, corrects closer the lower processing power of the client and the path to uniform distribution This is a network server allocating device that uses the modified probability distribution obtained by the above as a distribution rate.

Since the proportional relationship between the high processing capacity of the client and the route and the processing capacity of the server on the quality of service is reflected in the service distribution ratio, the client having a large effect on the service quality by the influence of the server processing capacity. There is an effect that a server with high processing capacity can be preferentially assigned.

The third means is the server allocation means in said first means obtains a distribution of the processing capacity of the client and the path of the clients currently in service, the processing capabilities of the new client connections and routes the distribution On the other hand, the lower the value, the closer the probability distribution according to the processing capacity of the server becomes to a uniform distribution. did.

Since the service distribution ratio is adjusted in relation to the clients currently in service and the distribution of the processing capacity of the route, it is possible to automatically cope with the case where the ratio of clients from distant places and from near places changes. is there.

The fourth means is at said first means,
Storage resources (buffers) because the server allocation means is divided into multiple parts and the scale of connection management means does not depend on service distribution.
Has the effect of increasing the utilization efficiency of.

There is an effect that various server allocations can be performed by a single device, for example, by selectively allocating the allocation target server group for each service or client and switching the service distribution policy . A fifth means is a step of monitoring the communication from the client to the server and measuring the communication data size per connection as the load of the server, and a step of detecting a change in the communication data size per connection and recording the maximum value. And a step of determining that the server has a high load if the communication data size per connection at that time with respect to the maximum value is small.

Here, in TCP and the like, the server evenly allocates a storage resource (buffer) for holding the packet data sent from the client for each connection.
The server notifies the client of the data size that can be held in the storage resource (buffer) at the next reception, and the client sends the data of the size notified from the server to the server.

Therefore, when the server becomes heavily loaded, the data sent from the client cannot be processed immediately, so that all or part of the data remains in the storage resource (buffer) of the server. As a result, the server It becomes necessary to notify the client of a size smaller by the amount of residual data in the storage resource (buffer).

Therefore, by detecting the data size per connection time on the communication line, it becomes possible to detect the high load state of the server . A sixth means is, in the fifth means, using the minimum number of monitoring communications and the minimum monitoring time until the number of monitored communications reaches the minimum number of monitoring communications and the measurement time reaches the minimum monitoring time. The number of connections and communication data size were measured.

The seventh means is in said fifth means,
The connection start and connection end communication is recognized, and the connection start and connection end communication data size is excluded from the load detection target.

Since the communication data for starting and ending the connection is small and does not depend on the load on the server, excluding it from the total data size of communication has the effect of improving the accuracy of load measurement and high load determination.

The eighth means is the same as the fifth means,
A step of holding connection start communication information until connection end or connection establishment, a step of detecting connection start communication for reconnection which the client judges as a connection failure based on the held information, and a connection start The ratio of the reconnection communication to the number of times of communication is set as the load of the server, and when this ratio is high, it is determined that the server has a high load.

When the load on the server is large, the server does not return a response notice to the connection request from the client. In response to this, the client retransmits the connection request. Therefore, the high load of the server can be determined by detecting the retransmission of the connection request of the client on the communication line.

The ninth means is in said fifth means,
Further, a step of obtaining a distribution of the communication data size from the client, a step of identifying extremely small communication data not related to the load of the server from the distribution, and a step of excluding the extremely small communication data from the load judgment. It includes.

Excluding extremely small communication data not related to the server load from the measurement has the effect of improving the accuracy of load measurement and high load detection . A tenth means, in the fifth means, obtains at least a sequence number from the communication from the client to the server, holds the maximum value of the sequence number from the start to the end of the connection, and Comparing the sequence number with the sequence number held above,
When the sequence number obtained from the communication is smaller than the held sequence number, the communication is excluded from the measurement.

The sequence numbers are normally in ascending order, but if the ordering of communication is destroyed or lost due to congestion on the communication line, the order will not be in ascending order. Since the server cannot process the data after the data that has not arrived, the receivable data size of the server becomes small regardless of the server load, and the communication data size of the client also becomes small. By avoiding the influence of the route by the above method, there is an effect of improving the accuracy of server load measurement and high load detection.

The eleventh means, in the fifth means, when the sequence number obtained from the communication is smaller than the held sequence number, the communication data is counted after weighting processing, or The communication data size when there is no problem on the route is predicted from both sequence numbers, and the predicted size is added to the load detection.

The twelfth means monitors communications from the server to the client, the server asks the steps of measuring the receivable data size and the number of connections and notifies to the client, the receivable data size per connection as a server load And a maximum value of receivable data size per connection, and a network server load detection method that judges that the server has a high load by reducing the receivable data size per current connection with respect to the maximum value. .

The thirteenth means is a server load detecting device for monitoring the communication from the client to the server and detecting the load state of the server, and a data size calculating means for calculating the size of communication data per connection, A storage unit that detects a change in the communication data size per connection and stores the maximum value, and detects a high load on the server when the communication data size per connection at that point in time with respect to the maximum value becomes a certain value or less. A network server load detection device comprising a load detection means for

[0034]

DETAILED DESCRIPTION OF THE INVENTION

[Embodiment 1] FIG. 1 shows a functional configuration of a server load detection device 4 according to the first embodiment. As shown in the figure, the server load detection device 4 is connected to the communication line 3 of the client 1 and the server 2, and more specifically, can be mounted on a router or the like.

As shown in FIG. 3, the server load detecting device 4 is provided with packet data (TC) transmitted through the communication line 3.
P packet: Transmission Control Protocol Packet)
It has a communication data capturing section 5 for capturing. A connection number detection unit 6, a packet number calculation unit 8 and a packet size calculation unit 7 are connected to the communication data acquisition unit 5.

The connection number detecting unit 6 has a function of detecting the number C of connections per unit time from the TCP packet captured by the communication data capturing unit 5. This connection number detection unit 6
Is +1 when a SYN packet which means the first packet is detected, and -1 when a FIN packet which means the last packet is detected. This allows the number of clients currently connected to the server to be detected.

The packet number calculation unit 8 has a function of counting the number N of TCP packets taken in by the communication data acquisition unit 5 per unit time, and the packet size calculation unit 7 has the communication data acquisition unit 5. It has a function of calculating the total size S of TCP packets taken in by the loading unit 5 per unit time.

The calculation / count data of each of these units is sent to the load detecting unit 10 and the load is determined based on a predetermined calculation process described later. The total packet size S calculated by the packet size calculator 7 is set to 0 at the start of measurement, and when a packet arrives, it is sequentially increased by the packet size. It should be noted that the SYN and FIN packets are smaller in size than the data packets and have a smaller effect on the server load, and may be ignored.

The number of packets N counted by the packet number calculator 8 is set to 0 at the start of measurement and is set to +1 every time a packet arrives. In addition, also here, SYN, FIN
For packets, the count may be ignored for the reasons mentioned above.

Counting by the packet number calculation unit 8 continues until N exceeds a certain value Nmin.
If the time from the start of measurement is shorter than the preset time Tmin even if the time exceeds, the counting is continued until the time Tmin elapses.

Here, Nmin and Tmin are set in the packet number calculation unit 8 in advance. Thus N
By using min and Tmin together, it is possible to reduce the calculation error that occurs due to the small number of packet samples for load detection, and to avoid the overflow that occurs due to too many samples. The detection accuracy can be improved.

The load detection in the load detection unit 10 is performed by performing the following arithmetic processing. First, when the load detection unit 10 receives the connection number C from the connection number detection unit 6 and the packet size S from the packet size calculation unit 7, the load detection unit 10 obtains the server load index value L based on the following formula.

Here, T is the measurement time measured by the timer 11. When the set count number N after the elapse of Tmin exceeds Nmin, T = Tmin.

L = (S / C) / T Here, L means the data transfer amount per connection per unit time. The load of the server 2 can be detected using this L.

Further, the load detection unit 10 updates the processing capacity limit predicted value Lmax of the server 2. Where Lmax
Is set to 0 as an initial value, and when L exceeds Lmax, Lm
Let the value of ax be L. Here, if the following relationship is established between L and Lmax, the server can be determined to have a high load. L <αLmax where 0 <α <= 1 (1) In the above formula (1), α is a preset constant. FIG. 3 is a flow chart showing load detection in the load detection unit 10 described above.

First, when the measurement is started, the count number N
And the server load index value L is reset, and the timer 11
Is started (step 301). Then, when packet reception is started via the communication data capturing unit 5 (302), is it a connection start packet SYN (30
3) It is determined whether the packet is the connection end packet FIN (305). Here, in the case of the connection start packet SYN, the variable V is incremented by 1 (304). Also,
When the packet is the connection end packet FIN, the variable V is decremented by 1 (306).

Then, each time a new packet is received, N
Is incremented by 1 and the server load index value L is calculated by the load detection unit 10 (307). This calculation is performed based on the calculation formula described above. When the server load index value L is smaller than αLmax using the above equation (1),
It is determined that the server is in a high load state.

Such a high load determination ends when the timer value is equal to or greater than the preset Tmin and the packet count number N is equal to or greater than the preset Nmin (308).

Here, in TCP, the server 2 evenly allocates a storage resource (buffer) for holding the packet data sent from the client 1 for each connection.
The server 2 notifies the client 1 of the data size that can be held in the storage resource (buffer) at the next reception, and the client 1 sends the data of the size notified from the server 2 to the server 2 through the communication line 3. .

Therefore, when the server 2 has a high load, the data sent from the client 1 cannot be processed immediately, so that all or part of the data remains in the storage resource (buffer) of the server 2, resulting in As a result, the server 2 has no choice but to notify the client of a size smaller by the amount of residual data in the storage resource (buffer).

Here, since TCP is a protocol designed to exchange data of a size as large as possible, the size of data sent from the client 1 to the server 2 is the maximum before the server 2 is overloaded. After that, when the load on the server 2 increases, the data size transmitted on the communication line 3 also decreases. In this embodiment, as shown in FIG. 2, the high load state of the server is detected by paying attention to the fact that the data size becomes small.

In this embodiment, the maximum data size transmitted through the communication line 3 when the server 2 is in the high load state is stored in the database 12 as Lmax. Then, as shown in the equation (1), a value (threshold value) obtained by multiplying this Lmax by a constant α is compared with L, and when this L becomes equal to or less than the threshold value, the server 2 is under high load. It is determined to be in a state.

As described above, in the present embodiment, since the size per connection is checked, it is possible to prevent erroneous judgment due to the decrease in the total data size due to the decrease in the number of connections itself. Further, by using α, It is possible to prevent erroneous detection of high load due to fluctuations in L caused by disturbance.

The flow chart of FIG. 4 is almost the same as the flow chart described in FIG. 3, but the communication start packet SYN
2 shows a procedure for determining a high load without considering the communication end packet FIN.

[0055]

Second Embodiment The second embodiment is a high load detection method using a retransmission process from the client 1 to the server 2.

The structure of the apparatus used in the second embodiment is almost the same as that shown in FIG. 1 of the first embodiment, and the illustration thereof is omitted. In the second embodiment, information on each start packet SYN is recorded in the database 12 (FIG. 6 (a)-
(See (c)). Then, each start packet SYN
Information is identified by a set of a client address (IP), a client port number (sp), and a server port number (dp).

In TCP, when the server 2 receives the start packet SYN from the client 1, it returns a SYN reception confirmation packet to the client 1. Here, if the client 1 cannot receive the SYN reception confirmation packet from the server 2 even after a certain period of time, the client 1 retransmits the start packet SYN to the server 2 again.

FIG. 5 shows this concept. In FIG. 3A, first, a connection request (start packet SYN) is transmitted from the client 1a to the server 2. On the other hand, another client 1b also sends a connection request (start packet SYN) to the server 2. Where server 2
When the buffer 51 of the server 1 has a sufficient capacity, that is, when the load is low, the server 2 determines that
A response notification (reception confirmation packet) is transmitted to. However, when the buffer 51 of the server 2 has no space, as shown in FIG. 2B, it is not possible to respond to the connection request (start packet SYN) from the client 1. Therefore, in the present embodiment, the client 1 retransmits the connection request to the server 2 when the client 1 cannot receive the response notification (reception confirmation packet) from the server 2 within a certain time as shown in FIG. In 2, the number Cs of start packets SYN is counted by the connection number detector 6, the number of retransmissions of the start packet SYN from the client 1 is detected, and the ratio Rs of the number of retransmissions of the start packet SYN to Cs is calculated. The server load index value Crs is set.

Here, the retransmission of the start packet SYN can be determined to be a retransmission if the SYN information extracted from the start packet SYN is already recorded in the database 12. This is shown in FIG. In the same figure (a), in the database of the load detection device 4, SYN1 (IP1, sp1, dp1), SYN is stored as SYN information.
2 (IP2, sp2, dp2) and SYN3 (IP
3, sp3, dp3) is recorded. At this time, a connection request (start packet SYN4) is transmitted from the client 1 to the server through the communication line 3. If the connection request is not the connection request stored in the database 12 of the load detection apparatus 4, that is, the first connection request, the load detection device 4 outputs the connection request (SYN4: IP4, s4).
p4, dp4) are stored in the database 12 (FIG. 6 (b)). Then, when the server 2 does not notify the client 1 of the connection request (SYN4), the client 1 retransmits the connection request (SYN4) to the server 2. The load detection device 4 is
This connection request (SYN4) is captured by the communication data capture unit 5, and the load detection unit 10 searches the database 12 to find out that the connection request is already stored by itself, and as a result, the connection request is stored. It is determined that (SYN4) is the reconnection request.

A specific measuring method in the load detecting unit 10 is the number of connections C and the packet size S described in the first embodiment.
According to the counting and detection method of. Here, if the following expression (2) is satisfied by the obtained ratio Rs of the number of retransmissions of the start packet SYN, that is, Crs, it is determined that the server 1 has a high load. Crs> β However, 0 <β <1 (2) In the above equation (3), β is a preset constant.

The server 2 allocates the buffer 51 for holding the data from the client 1 for each connection, but when the allocated buffer 51 is exhausted, the server 2 does not connect and does not return the response notification (SYN reception confirmation packet) to the client 1. Therefore, in the client 1, the ratio of the number of retransmissions of the start packet SYN increases. Therefore, it becomes possible to detect the high load of the server from the equation (2). FIG. 6D is a graph showing the relationship between such a retransmission rate (retransmission count / communication count) and server load.

The constant β in the above equation (2) is for preventing erroneous detection due to disturbance or momentary high load condition.
An instantaneous high load condition has a low probability of occurrence and does not last long, so it can be ignored.

[0063]

[Third Embodiment] The third embodiment is a technique for discriminating a measurement object based on a communication data size when detecting a load. Since the device configuration of the third embodiment is the same as that of FIG. 1, it will be described with reference to FIG.

In the third embodiment, the load detection unit 10 detects the load without adding Si to the total packet size L when the following relationship is established between the packet size Si from the client 1 and Ds. . Si <γDs (3) where 0 <γ <1, Ds = f (S1, S
2 ,. ． ． ． Si-1).

Here, γ is a preset constant. Ds is a function for obtaining the distribution index of the measured packet size, and may be an average value, for example. If the Ds result value is a plurality of values, a single value may be used by weighted addition or selection.

In TCP, after connection, the client 1 sets the transmission data size to a size smaller than the data size notified from the server 2, starts transmission, and gradually increases to the notification data size. Therefore, the packet size from the client 1 shortly after the start of connection is small regardless of the load on the server 2.

Therefore, if the number of clients 1 shortly after the start of connection is large, L of the formula (1) may be underestimated due to a large amount of small transmission data, and the accuracy of load measurement and high load detection may be degraded. There is a nature.

FIG. 7 conceptually shows this. In the figure (a), the client 1a transmits the packet data A of a relatively large size to the server 2, but the client 1b has a relatively short time after the start of the communication, and therefore the command and the response signal are relatively large. The packet data B having a small size is transmitted. Such small packet data can be safely ignored when detecting the load on the server.

Therefore, in the third embodiment, by using the equation (3), the packet from the client 1 which is just after the start of the connection is detected and is excluded from the measurement target.
The accuracy of load measurement and high load detection is improved.

When the server has a high load, the data size from all connected clients becomes smaller, but the decrease in the buffer 51 for holding data is relatively slow, so the decrease in L above is also slow. Is. Further, since it is unlikely that all the clients start a new connection all at once, the formula (3) is sufficient.

In order to improve accuracy, the lower limit value Dsmin of Ds is set as an applicable condition in the equation (3), and Ds is Dsmi.
If n or less, the equation (3) is not applied, that is, Si is L
May be added to.

[0072]

[Fourth Embodiment] In the fourth embodiment, in the load detection described in the first embodiment, a technique for preventing a high server load from being erroneously detected due to a packet contradiction caused by congestion or the like on a communication line. Is.

The apparatus configuration of the fourth embodiment is the same as that of FIG. Here, in the packet from the client 1 to the server 2, a set of a client address (IP), a client port number (sp) and a server port number (dp) (packet identifier), and a sequence number are stored in the database 12 from the start to the end of the connection. Held in. The sequence number held at this time is the maximum value (the final value at that point).

When the load detection device 4 receives a packet from the client 1 to the server 2, the packet identifier and the sequence number Pi are obtained from the packet and compared with the sequence number Pj of the same packet identifier held in the database 12. To do.

Here, according to the judgment of the load detection unit 10, P
If i <Pj holds, it can be understood that the packet was overtaken on the communication line 3 or was retransmitted because the packet on the way was lost.

In any case, the data received by the server 2 in such a state will have a deficiency in the middle, and the server 2 cannot process the data after the deficiency and the data after the deficiency will not be processed. Will remain in the buffer 51. This reduces the size of data that the server 2 can receive, but the cause is not the server load but congestion in the route between the client and the server. FIG. 8 conceptually shows this. In the figure, the packet data “1 to
3 ”is transmitted, but this is the reason why only packet data“ 2 ”is lost due to factors such as route congestion. The client 2 stores the received packet data “1, 3” in the buffer 5
Store in 1. Here, a response notification (retransmission request for packet data “2”) is sent to the client 1, but since the packet data “2” is not received in its own buffer, the packet data “3” that has already arrived is sent. The subsequent processes cannot be processed.

The client 1 resends the packet data “2” when receiving duplicate response notifications regarding the packet data “2”. In this way, the packet data “2 to 5” is prepared, and the server 2 is ready to process these received packet data. However, since the process cannot be immediately started, the free size of the buffer notified to the client 1 Becomes n, which is much smaller than the original size N of the buffer.

Next, the client 1 transmits the packet data "6" of a size that can be stored in the size n notified from the server 2, but actually, at the stage of receiving this packet data "6", the packet data "6" is received. "1-5"
Is being processed, there is a large free space in the buffer and it is not in a high load state.

That is, in the fourth embodiment, such a structure as shown in FIG.
The state shown in is not judged as a high load. For the above reasons, the packet Pi for which Pi <Pj is satisfied is excluded from the measurement. Alternatively, a certain weighting may be performed for the accounting, and in the load detecting unit 10, Pj
-Pi may be accounted for in addition to the packet size.

Here, the calculation of Pj-Pi is performed from the server 2 to the client 1 when no data loss occurs by adding the predicted size of the data remaining in the buffer 51 in the server 2 to the packet size. This means that the receivable data size to be notified, that is, the size of the current packet is predicted.

[0081]

Fifth Embodiment In the fifth embodiment, the server 2 is the client 1
The load of the server 2 is determined by monitoring the packet data transmitted to the server.

The load detection device 4 of the fifth embodiment is the server 2
Monitors the total window size value Sw and the number of connections C in the packet sent to the client 1. The window size is the receivable data size that the server 2 notifies the client 1.

The value of C is obtained by incrementing 1 when the start packet SYN from the server 2 to the client 1 is detected and subtracting 1 when the end packet FIN is detected.
Here, the measurement of Sw and C is the same as that of the first embodiment.

The load index value L3 of the server 2 is obtained by the following equation. T is the same as T in Example 1, but is not always necessary. L3 = (Sw / C) / T (4) L3 means the window size per connection. L3
The method of detecting a high load on the server 2 using is as follows.

First, the processing capacity limit predicted value L3 of the server 2
Update max. The initial value of L3max is 0, and L3max
When the value exceeds L3max, the value of L3max is set to L3.

If the following relationship is established between L3 and L3max, the server 2 determines that the load is high. L3 <α3 · L3max where 0 <α3 <= 1 (5) In the above equation (5), α3 is a preset constant.

The server 2 notifies the client 1 of the free size of the buffer 51 that it can process, that is, the window size (FIG. 9A), but the load of the server 2 increases at this point. When the data sent from the client 1 cannot be completely processed, the server 2 notifies the client 1 of the window size n smaller than before (more specifically, next time reception), as shown in FIG. 9B. Possible data size). FIG. 9B is a graph showing the relationship between the window size and the time notified from the server 2 to the client.

Since the load of the server 2 affects all connected clients, L3 decreases as the server load increases. Therefore, the server load can be measured by the equation (4), and the high load can be detected by the equation (5).

[0089]

Sixth Embodiment The sixth embodiment is a case where the server allocation device of the present invention is realized as a device for relaying TCP packets between a client and a server.

In FIG. 10, when the packet 1010 received from the client 1 is the start packet SYN meaning a connection request, the destination conversion / packet relay means 1002 of the server allocation means 1001 determines the server to which the service is allocated. A server allocation instruction 1020 is output to the server selection means 1007, and a measurement instruction 1021 is given to the client side processing capacity measuring means 1008.

The server processing capacity measuring means 1004 calculates the processing capacity of each server and sends the result data 1013 to the server allocation probability calculating means 1006. The processing capacity of each server can be calculated from the response time to a ping sent to the server 2 or can be set in advance by the user. In addition, the first embodiment
It is also possible to use the server load detection device described in 5 above.

Client side processing capacity measuring means 1008
Upon receiving an instruction from the relay means 1002, calculates the processing capacity 1018 of the client 1 and the communication line 3, and reports it to the server allocation probability correction information generation means 1009. Here, the processing capacity on the client side can also be obtained from the response time by sending a ping or the like to the client, for example. In addition, a bandwidth measurement method such as Bprob is used, past records of the communication of the client 1, window size extracted from the packet, and TT.
It can also be obtained from L.

The server allocation correction information generating means 1009
A correction function 1022 for the server allocation probability distribution is generated from the client side processing capacity 1018. FIG. 11 shows an example of the server allocation probability distribution PsD, and the lower diagram of FIG. 12 shows an example of the correction function M (1022).

The server allocation probability calculating means 1006 applies the modification function M (1022) to the server allocation probability distribution PsD to obtain the server allocation probability distribution MPsD. Probability distribution P
The distribution of sD is such that the allocation probability increases with the server having higher processing capability at the present time. For example, the processing capacity values of each server at the present time (described later) are p1, p2 ,. ． ． pn
If (n is the number of servers), the allocation probability Pi to the server Si can be calculated by the following equation. Pi = pi / (p1 + p2 + ... + pn) (6) As shown in FIG. 12, the probability correction function M is a function for correcting PsD so that it becomes closer to a uniform distribution as the processing capacity on the client side is lower. Becomes For example, if the response time Tping due to ping is the client processing capability,
The modified Pi ′ may be obtained from the following formula for each server processing capacity Pi. Pi ′ = Pi + (Pav−Pi) * 2 / π * arc_tan (α * Tping) (7) where Pav is the average value of Pi, and α is a preset number greater than 0. is there. Also, arc_t
an (x) means tan-1 (x).

A modified probability distribution MPsD is obtained from this Pi '. The server allocation probability calculation means 1006 determines the calculated MP
The sD is sent to the server selection means (1007). The server selection means (1007) generates the table shown in FIG. 13 from MPsD and implements it using a uniform random number value that takes an arbitrary value of 0 to 1. The table shown in the figure is realized by, for example, an array having the number of servers, and each element has a range Pi from 0 to 1.
A server address of an element having a range including a uniform random number value may be set as a service allocation server address by setting a pair of the maximum value and the minimum value of and the server address. However, the range of each element should not overlap with the range of other elements.

Regarding the probability distribution of PsD and MPsD, the server processing capacity values Pi and Pi ′ may be realized as a frequency distribution. In this case, the uniform random number is from 0 to all P
Try to take a range up to the total value of i.

The server selecting means 1007 sends the server address 1012 to the connection managing means 1003 after determining the assigned server. The connection management means 1003 converts the destination
The group information of the client address (IP), the client port number (sp), and the destination port number (dp) is extracted from the start packet SYN received from the packet relay unit or a part thereof, and is received from the group information and the server allocation unit 1001. Record the pair of server addresses here,
A hash table that uses the group information as a key may be used for recording. The connection management means 1003 sends the server address 1012 to the destination conversion / packet relay means 1002.

The destination conversion / packet relay means 1002
The destination of the received packet from the client 1 is the server address 1012 received from the connection management means 1003.
Converted to and transmitted to the server 2.

During the service, the destination conversion / packet relay means 1002 sends the packet 1014 to the connection management means 1003.
The connection management means 1003 sends the packet 101
The assignment server address 1012 is obtained from the group information obtained from No. 4 and is sent to the destination conversion / packet relay means 1002. Similar to the start packet SYN, the destination translation / packet relay unit 1002 translates the destination of the received packet from the client 1 into the server address 1012 received from the connection management unit 1003, and sends it to the server 2.

At the end of the service, that is, the end packet F
When IN is received, it is the same as during service, but the connection management means 1003 receiving this discards the group information corresponding to the packet.

In the present embodiment, by determining service allocation using the probability distribution, it becomes easier to allocate a server having a higher processing capacity to a client having a higher processing capacity. Therefore, the influence of the server processing capacity on the service quality such as response time is affected. Services can be assigned according to the strength of the service.

The server allocation probability correction information generating means 1009 of the server allocating means 1001 obtains the past client side processing capacity distribution (FIG. 14) and obtains the deviation δ from the client processing capacity value distribution of the newly connected client, The distance δc from the distribution of the client side processing capability value of the newly connected client is calculated. Then, the probability distribution is corrected by adding δc to the correction function M (FIG. 15). Here, for example, δc is calculated by the following formula. δc = Pca−pci Here, Pca is the past average processing capacity on the client side, and pci is the processing capacity value on the client side of the newly connected client. The correction function M is determined such that the smaller δc is, the closer the server processing capacity value pi is to the average value of all pi, and the larger the difference is, the larger the deviation from the average value is. However, when the distance from the average value is increased, the modified value pi 'of pi should not be a negative number. For example, the equation (7) may be as follows. pi '= pi + (Pav-pi) * β * 2 / π * arc_tan (α * δc + γ) (7 ′) where Pav is the average value of pi, and α and γ are preset to 0. It is a larger number. Also, arc_
tan (x) means tan-1 (x). β is -1
, Δc <0, and when −dpj / pj, δ> =
It is 0. pj is the minimum value of pi, and dpj = Pav
-Pj.

In this way, by allocating the server according to the gap between the client-side processing capacity value of the newly connected client and the past distribution of the client-side processing capacity value, it becomes possible to allocate the server according to the client at each time point. . For example, it becomes possible to automatically cope with a case where the ratio of clients from a remote place and from a neighboring place changes depending on the time of day.

Furthermore, in the present embodiment, a plurality of server allocation means (1001) may be arranged and each may be selected according to the client address, client port number, service port number and the like.

It becomes possible to selectively use the allocation target server group for each service or each client and switch the service distribution policy, and various devices can be allocated by one device. According to the present invention, the load measurement and the high load detection of the server are performed by monitoring the communication between the client and the server. Therefore, it is not necessary to modify the server and packets other than services are not output. Therefore, there is an effect that any server can be dealt with, the introduction cost is low, and there is no interference with the load. In addition, since the load is measured and the high load is detected using an index that does not depend on the protocol, it has the effect of being able to handle any service, and the effect of disturbance is small and the accuracy is high because the communication status during service is monitored. is there. In addition, when the services provided by a server are shared by multiple servers, the load of each server is shared according to the size of the impact of the server processing capacity on the quality of service seen by the client in response to changes in the server configuration and changes in the server status. Is automatically and efficiently allocated, so that there is an effect that the client can be promptly provided with a service. (Others) The above embodiments disclose the following inventions. (Supplementary Note 1) In a device that transfers data from a client to a plurality of servers via a network, a relay unit that transfers the data transmitted from the client to one of the servers by changing the destination, and the correspondence relationship between the data and the server Connection management means for holding and instructing the destination to the relay means, and measuring and obtaining the processing capacity of the server, client and route, and determining the correspondence between the data and the server by using the function according to the service distribution rate based on it A network server allocating device comprising a server allocating means for notifying to management means. (Supplementary Note 2) In the server allocating means of Supplementary Note 1, a correction probability distribution obtained by correcting the probability distribution according to the processing capacity of the server so as to be closer to a uniform distribution as the processing capacity of the client and the route is lower. Network server allocating device for distribution rate. (Supplementary Note 3) In the server allocating means of Supplementary Note 1, the distribution of the processing capacities of the clients and the routes of the clients currently in service is obtained. A network server allocating device that approximates a probability distribution according to capacity to a uniform distribution, and conversely corrects the processing capacity of each server as it gets higher to obtain a correction probability distribution, which is used as a distribution rate. (Supplementary Note 4) The network server allocating apparatus according to Supplementary Note 1, comprising a plurality of server allocating means, the server allocating means being selected for each client or service.

[0106]

As described above, according to the present invention, the service is distributed according to the method obtained by measuring the performance / load of the server and the performance / load of the client side. It is possible to automatically respond to changes in the above, and moreover, it is possible to allocate as many servers as necessary to maintain the quality of service seen by the client, and there is an effect that the utilization efficiency of the servers is maximized.
Furthermore, since the server allocation is determined using the function, there is an effect that the service distribution rate can be accurately reflected in the distribution.

Further, since the proportional relationship between the high processing capacity of the client side and the effect of the processing capacity of the server on the service quality is reflected in the service distribution ratio, the client having a large effect on the service quality by the influence of the server processing capacity. There is an effect that a server with high processing capacity can be preferentially assigned. Further, since the service distribution ratio is adjusted in relation to the distribution of the processing capacity of the clients currently in service, there is an effect that it is possible to automatically cope with the case where the ratio of the clients from the distant place and from the near place changes.

[Brief description of drawings]

FIG. 1 is a diagram showing a connection configuration of a load detection device according to an embodiment of the present invention.

FIG. 2 is a graph showing the relationship between the data size and time for making a high load determination of the server in the embodiment.

FIG. 3 is a flowchart showing a packet monitoring method according to the first embodiment (1).

FIG. 4 is a flowchart (2) showing a packet monitoring method according to the first embodiment.

FIG. 5 is a diagram for explaining a connection request from the client to the server and a response process according to the state of the buffer.

FIG. 6 is a diagram for explaining retransmission of a connection request from a client to a server.

FIG. 7 is an explanatory diagram showing an example of discriminating whether or not to be target data according to a data size in load detection.

FIG. 8 is a diagram for explaining processing of a server based on a sequence number.

FIG. 9 is an explanatory diagram for monitoring communication from the server to the client.

FIG. 10 is a block diagram showing a configuration of a server allocation device according to an embodiment.

FIG. 11 is a graph diagram for explaining a server allocation probability distribution PsD.

FIG. 12 is a diagram (1) for explaining a correction function.

FIG. 13 is a diagram showing an example of a table generated by a server selection unit in the embodiment.

FIG. 14 is a graph showing an example of distribution of past processing capacity values on the client side.

FIG. 15 is a diagram (2) for explaining a correction function.

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Shinji Kikuchi Shinji Kikuchi 4-1-1 Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Fujitsu Limited (56) Reference JP-A-9-282287 (JP, A) Hei 10-91549 (JP, A) Teiyuki Kiyosue, Architecture of Interspace, NTT R & D, Denki Denshi Co., Ltd., April 10, 1998, Volume 47, No. 4, p. 459-464 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06F 15/177 G06F 13/00

Claims (3)

(57) [Claims] 1. In a device for transferring data from a client to a plurality of servers via a network, a relay means for transferring the data transmitted from the client to one of the servers by changing the destination, and the correspondence relationship between the data and the server. The connection management means that holds the information and determines the destination to the relay means, and the processing capacity of the server and the processing capacity on the client side are measured and obtained, and the correspondence between the data and the server is determined using a function according to the service distribution rate based on that. Then, the server allocating means for notifying the connection managing means is sought, and in the server allocating means, the probability distribution according to the processing capacity of the server is corrected by making it closer to a uniform distribution as the client side processing capacity is lower. A network server allocation device having a distribution ratio of a modified probability distribution. 2. In a device for transferring data from a client to a plurality of servers via a network, a relay means for transferring the data transmitted from the client to one of the servers by changing the destination, and the correspondence relationship between the data and the server. The connection management means that holds the information and determines the destination to the relay means, and the processing capacity of the server and the processing capacity on the client side are measured and obtained, and the correspondence between the data and the server is determined using a function according to the service distribution rate based on that. And a server allocating unit for notifying to the connection managing unit, wherein the server allocating unit obtains a distribution of the client-side processing capacity of the clients currently in service, and the lower the new-connection client-side processing capacity is, the server The probability distribution according to the processing capacity of the server is approximated to a uniform distribution. A network server allocating device that makes corrections that highlight the ability and obtains a correction probability distribution, which is used as the distribution rate. 3. In a device for transferring data from a client to a plurality of servers via a network, a relay means for transferring the data transmitted from the client to one of the servers by changing the destination, and the correspondence relationship between the data and the server. The connection management means that holds the information and determines the destination to the relay means, and the processing capacity of the server and the processing capacity on the client side are measured and obtained, and the correspondence between the data and the server is determined using a function according to the service distribution rate based on that. and a server allocation unit for transmitting to the connected management means, in said server allocation unit, Cry currently in service
Find the distribution of client-side processing power for Ants
Therefore, new connection client side processing capacity is low against distribution
The uniform probability distribution according to the processing capacity of the server
, And conversely, the higher it is, the more the processing capacity of each server is emphasized.
And a correction probability distribution is obtained, and the correction probability distribution is used as a distribution rate. The server allocation means has a plurality of the server allocation means, and the server allocation means can change the service distribution policy for each service or each client so that the clients and the services can be switched. Network server allocation device selected for each.