JP2015095200A - Cluster system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To ensure that, even when a representative member for updating and delivering a member management table and other members go down at the same time in a cluster system, a new representative member can be selected at high speed.SOLUTION: The present invention is a cluster system in which a plurality of cluster members perform distributed processing. Each of the plurality of cluster members (representative member 100, non-representative member 200) is provided with: a storage unit 120 (220) for storing a monitored-member setting rule 123 (223) in which it is stipulated that, when determining the subject to be alive-monitored among the plurality of cluster members, a cluster member for performing alive-monitoring on a next-term representative member candidate be selected from other than the then representative member; and a processing unit 110 (210) for selecting, in accordance with the monitored-member setting rule 123 (223), the subject to be monitored by the member itself when determining the subject to be alive-monitored among the cluster members.

Description

  The present invention relates to a life and death monitoring technique between cluster members in a cluster system in which a plurality of cluster members (such as servers, hereinafter may be simply referred to as “members”) perform distributed processing.

  In recent Web systems that require large-capacity data retention, high-speed access, and high availability, cluster systems that improve the processing capacity of the entire system by coordinating a plurality of servers are often used. In the distributed processing by the cluster system, the cluster members constituting the cluster and the data in charge must be associated with each other. As a method of associating data with cluster members, for example, there is a method called a consistent hashing method (Non-patent Document 1).

  The consistent hashing method is a method of determining data assigned to a cluster member by mapping a hash value of data and an address assigned to the cluster member on the same ID (IDentifier) space. Even if the number of cluster members increases or decreases, this method requires only a change of 1 / N (N: the number of cluster members) of the correspondence between the cluster members and the data, thereby reducing the load of data relocation. There is.

  Further, in this method, in order to improve the availability of the cluster system, there is a case where a method of having data replication in addition to the server in charge is used in combination. In that case, each cluster member needs to hold a list of all members (member management table) and be able to specify the replication destination from the member management table under specific rules. As shown in FIG. 1, the member management table includes a member ID as a key for uniquely designating each member and additional information unique to the member such as an IP (Internet Protocol) address. At this time, in order to ensure the consistency of the member management table, it is desirable that the update authority is given only to specific servers inside and outside the cluster.

  As a method for realizing such consistency of the member management table, there is a method in which one of the cluster members is used as a representative member and the member management table is updated and distributed to other members (Non-patent Document 2). . In this method, the representative member is selected from among the cluster members in accordance with an arbitrarily determined representative member selection rule. Each cluster member grasps (stores) representative member selection rules, and can identify representative members including themselves based on information in the member management table. The change of the representative member due to the increase / decrease of the cluster member is also performed by the autonomous determination of the cluster member.

(Life monitoring between cluster members and replacement of representative members)
As a cluster member failure (abnormality) detection method, there is a method of monitoring between cluster members. The monitoring target is determined for each cluster member, and for example, failure detection within a predetermined time is performed by confirming the presence or absence of a response by periodic heartbeat transmission (every predetermined time). When a failure is detected, the cluster member who has performed monitoring is notified of the failure of the member to all cluster members. The member is removed from the cluster system. Therefore, the representative member updates the member management table and distributes the updated member management table to all members. At this time, if the failed member is a representative member, the next representative member appears autonomously in accordance with the representative member selection rule.

Functions (configuration of the processing unit) and information possessed by the representative member and other members will be described with reference to FIG.
As shown in FIG. 2A, the representative member includes a processing unit, a storage unit, and an input / output unit.
The processing unit includes a member management table distribution unit, a member management table update unit, a representative member specifying unit, and another member alive monitoring unit.
The storage unit stores a member management table, representative member selection rules, and old monitoring destination member setting rules.

As shown in FIG. 2B, the non-representative member includes a processing unit, a storage unit, and an input / output unit.
The processing unit includes a representative member specifying unit and another member alive monitoring unit.
The storage unit stores a member management table, representative member selection rules, and old monitoring destination member setting rules.

  Although not shown, the representative member and the non-representative member have a message processing function as a cluster member, a duplicate data transmission function, a management function, and the like.

(Implementation example and problems)
Consider an implementation example of member management based on the above-mentioned member management table (see also FIG. 1) and the alive monitoring system between cluster members. For example, according to the representative member selection rule, the first representative member is set as the first member in the description order of the member management table. At this time, the member management table is described in the order of participation in the cluster. Furthermore, the alive monitoring target of each cluster member is set as the next (one lower) member of itself in the member management table (the last member monitors the first member). As an example, FIG. 3 shows a case where the number of cluster members N = 4. In this case, the operation of the cluster member accompanying one member failure is as follows.

<(A) When a member other than the representative member fails (see FIG. 4)>
As an example, consider the case where member 3 fails. The member 2 who is alive monitoring the member 3 detects the failure of the member 3 (S41), and notifies all members of the failure of the member 3 (S42). After that, the representative member who has received the notification updates the member management table (S43), and distributes the updated member management table to all members (S44). The same applies when the members 2 and 4 other than the representative member fail.

<(B) When a representative member fails (see FIG. 5)>
The member 4 who is alive monitoring the member 1 (representative member) detects the failure of the member 1 (representative member) (S51), and notifies the failure of the member 1 (representative member) to all members (S52). Thereafter, the member 2 (next representative member candidate) who has received the notification recognizes itself as a new representative member (S53), updates the member management table (S54), and distributes the updated member management table to all members. (S55). Thereafter, the member 2 operates as a representative member.

  Next, consider the following case as an operation when two members fail simultaneously.

<(C) When a representative member fails and a member who is neither a member who monitors the next representative member candidate (second member from the top) nor a member who monitors the representative member fails (see FIG. 6)>
Consider a case where member 1 (representative member) and member 3 fail. First, the member 4 who is alive monitoring the member 1 (representative member) detects the failure of the member 1 (representative member) (S61), and notifies the failure of the member 1 (representative member) to all members (S62). In parallel, the member 2 who is alive monitoring the member 3 detects the failure of the member 3 (S63), and notifies the failure of the member 3 to all members (S64). Thereafter, the member 2 who is the next representative member candidate receives a failure notification of the member 1 (representative member) and the member 3, recognizes itself as a new representative member (S65), updates the member management table (S66), The updated member management table is distributed to all members (S67). Thereafter, the member 2 operates as a representative member.

<(D) When the representative member fails and the next representative member candidate (second member from the top) fails (see FIG. 7)>
Consider a case where member 1 (representative member) and member 2 (next representative member candidate) fail. First, the member 4 who is alive monitoring the member 1 (representative member) detects the failure of the member 1 (representative member) (S71), and notifies the failure of the member 1 (representative member) to all members (S72). At this time, the member 2 (next representative member candidate) also fails, but failure detection is not normally performed because the monitoring source member 1 (representative member) has failed. In this case, since member 1 (representative member) and member 2 (next representative member candidate) are at the same time, member 3 which is the next representative member candidate needs to be the representative member. Here, the member 2 is the next representative member candidate, but since it has failed, the member management table is not updated (first stage of S73). However, since the failure of the member 2 (next representative member candidate) is not notified, the member 3 recognizes the member 2 (next representative member candidate) as a new representative member (second stage of S73). Then, the member 3 continues to wait for the member 2 (next representative member candidate) to update and distribute the member management table as a new representative member, and the processing of the cluster system stops.

<(E) When a representative member fails and a member that monitors the representative member (last member) fails (see FIG. 8)>
Consider a case where member 1 (representative member) and member 4 (member monitoring the representative member) fail. First, the member 3 that is alive monitoring the member 4 (member that monitors the representative member) detects a failure of the member 4 (member that monitors the representative member) (S81), and the member 4 (member that monitors the representative member) Is reported to all members (S82). At this time, the member 1 (representative member) has also failed, but failure detection is not normally performed because the monitoring source member 4 (member that monitors the representative member) has failed. In this case, since member 1 (representative member) is out of order, member 2 as the next representative member candidate needs to be the representative member. Here, since member 1 (representative member) is out of order, the member management table is not updated (first stage of S83). However, the member 2 (next representative member candidate) recognizes the member 1 as a representative member because the failure of the member 1 (representative member) is not notified (second stage of S83). Then, since member 2 (representative member candidate) is not notified of the failure of member 1 (representative member), member 2 (representative member) continues to wait for member 1 (representative member) to update and distribute the member management table. Cluster system processing stops.

  When the representative member and the next representative member candidate shown in (d) (FIG. 7) fail simultaneously, and when the representative member shown in (e) (FIG. 8) and the member monitoring the representative member fail simultaneously However, there is a problem that the system stops because no new representative member appears and the member management table is not updated.

  One conventional solution to this problem is shown in FIG. As in the case of FIG. 7, consider a case where member 1 (representative member) and member 2 (next representative member candidate) fail. In this solution, when a failure notification of a cluster member occurs, if the member management table is not updated by the representative member for a predetermined time, a member different from the original alive monitoring member is designated as the representative member or The failure detection of the next representative member candidate is performed.

  First, the member 4 who is alive monitoring the member 1 (representative member) detects the failure of the member 1 (representative member) (S91), and notifies the failure of the member 1 (representative member) to all members (S92). Here, the member 2 is a candidate for the next representative member, but has failed, so the member management table is not updated (first stage of S93). However, since the failure of member 2 (next representative member candidate) is not notified, member 3 recognizes member 2 (next representative member candidate) as a new representative member (second stage of S93).

  Here, when the member management table is not updated for a predetermined time from the failure notification, the member 4 requests the member 3 to detect the failure of the member 2 (next representative member candidate) (S94). The member 3 detects the failure of the member 2 (next representative member candidate) (S95), and notifies all members of the failure of the member 2 (next representative member candidate) (S96). Thereafter, the member 3 recognizes itself as a new representative member (S97), updates the member management table (S98), and distributes the updated member management table to all members (S99).

Fujio Maruyama, Kazuyuki Shudo, “Technology of scale-out”, seizing beyond the cloud world Cloud Technology, ASCII Media Works, Inc., November 6, 2009, ISBN 978-4-04-886864-6 C3004, p . 88-99 Eriko Iwasa, 5 others, “Dynamic configuration change load reduction method in scale-out type session control server”, IEICE Technical Report, IEICE, 2012, NS2011-156 (2012-01), p. 65-70

  In this way, as shown in FIG. 9, when the failure notification of the cluster member occurs, if the member management table is not updated by the representative member for a predetermined time, it is different from the original alive monitoring member. A method in which a member detects a failure of a representative member or a next representative member candidate is effective. However, this method has a problem that it takes a relatively long time until the member management table is finally updated because it takes time to determine the failure of the representative member or the next representative member candidate. During the period when the member management table is not updated correctly, data replication is not managed properly, so the status of cluster members can be reflected more quickly in a highly available system where service is not allowed to stop. A method is required.

  Therefore, the present invention has been made in view of the above circumstances, and in a cluster system, even when a representative member that updates and distributes a member management table and another member fail at the same time, a new representative member is The task is to select the first.

  In order to solve the above problems, the present invention provides a cluster system in which a plurality of cluster members perform distributed processing, and each of the plurality of cluster members has a member management table including identifiers of the plurality of cluster members. A representative member of the plurality of cluster members updates the member management table, distributes the updated member management table to each of the plurality of cluster members, and among the plurality of cluster members, the representative member The next representative member candidate that is the next representative member when the member fails is selected in advance, and each of the plurality of cluster members determines the next life / death monitoring target to be performed between the plurality of cluster members. It was decided to select a cluster member to perform life and death monitoring for a representative member candidate from a member other than the representative member at that time. A storage unit for storing a viewing-destination member setting rule; and a processing unit for selecting a monitoring target according to the monitoring-destination member setting rule when determining a life / death monitoring target to be performed among the plurality of cluster members, It is characterized by providing.

  As a result, in the cluster system, by using the monitoring member setting rules described above, a new representative member can be selected at high speed even when the representative member that updates and distributes the member management table and another member fail at the same time. be able to.

  Also, in the present invention, the monitoring destination member setting rule is that the relationship of life and death monitoring performed between the plurality of cluster members starts from an arbitrary cluster member, and the cluster member monitoring destination cluster member and further the cluster It is desirable that the rule is set so that all the cluster members are covered and then the original starting cluster member is returned when the members are monitored in the order of the monitoring destination cluster members.

  Thus, by configuring the cluster members so that the monitoring destinations of all cluster members extend over all members, it is possible to further reduce the number of members that cannot be detected when a plurality of members fail simultaneously.

  In the present invention, when the number of cluster members (N) is an odd number, the monitoring destination member setting rule sets the monitoring target of each cluster member to one cluster member skipped in the descending order of the order in the member management table. The “N−1” -th cluster immediately before is monitored for the first cluster member, and the last N-th cluster member is a rule set to monitor the second cluster member from the top. desirable.

  As a result, when the number of cluster members (N) is an odd number, the monitoring destination of each cluster member can be easily configured so as to form a system across all members.

  In the present invention, when the number of cluster members (N) is an even number, the monitoring destination member setting rule specifies that the monitoring target of each cluster member is the member management for the “N-2” -th cluster member from the top. The cluster member is skipped by one in descending order in the table, the “N−1” -th cluster one previous from the last is monitored for the second cluster member from the top, and the last N-th cluster member is the first cluster member. The rules are preferably set to monitor cluster members.

  Thus, when the number of cluster members (N) is an even number, the monitoring destination of each cluster member can be easily configured so as to form a system across all members.

  According to the present invention, in a cluster system, a new representative member can be selected at high speed even when a representative member that updates and distributes a member management table and another member fail at the same time.

It is a data block diagram of the member management table | surface of a prior art. (A) is a block diagram of the representative member of a prior art. (B) is a block diagram of a non-representative member of the prior art. It is a figure which shows the alive monitoring relationship between cluster members in the case of the number N of cluster members in a prior art. In the prior art, when the member 3 fails in the example of FIG. 3, (a) is a schematic diagram showing the state of each member, and (b) is a flowchart showing processing by each member. In the prior art, when the member 1 fails in the example of FIG. 3, (a) is a schematic diagram showing the state of each member, and (b) is a flowchart showing processing by each member. FIG. 3A is a schematic diagram showing the state of each member and FIG. 5B is a flowchart showing processing by each member when the members 1 and 3 fail simultaneously in the example of FIG. FIG. 3A is a schematic diagram showing the state of each member and FIG. 4B is a flowchart showing processing by each member when the members 1 and 2 fail simultaneously in the example of FIG. FIG. 3A is a schematic diagram illustrating the state of each member and FIG. 3B is a flowchart illustrating processing by each member when the member 1 and the member 4 fail simultaneously in the example of FIG. FIG. 3A is a schematic diagram showing a state of each member, and FIG. 3B is a flowchart showing processing by each member when the members 1 and 2 fail simultaneously in the example of FIG. is there. (A) is a block diagram of the representative member of this embodiment. (B) is a block diagram of a non-representative member of the present embodiment. (A) is a figure which shows the monitoring relationship between members when the system of this embodiment is applied when the number of members (N) is an odd number. (B) is a figure which shows the monitoring relationship between members when the system of this embodiment is applied when the number of members (N) is an even number. In this embodiment, as in the case of FIG. 7, (a) is a schematic diagram showing the state of each member, and (b) shows the processing by each member when the members 1 and 2 fail simultaneously. It is a flowchart. (A) is a figure which shows the monitoring relationship between members in the case where it configures so that the monitoring destination may form the system over all members in this embodiment. (B) is a figure which shows the monitoring relationship between members in the case of the structure where the monitoring destination does not straddle all members as a comparative example. In this embodiment, it is explanatory drawing of the setting procedure of the monitoring destination of the cluster member for comprising so that a monitoring destination may form the system over all members, (a) is the arrangement | positioning condition of each member in a member management table | surface (B) is a flowchart showing the setting procedure.

Hereinafter, modes for carrying out the present invention (hereinafter referred to as embodiments) will be described with reference to the drawings (refer to drawings other than the referenced drawings as appropriate).
In this embodiment, a cluster system in which a plurality of cluster members perform distributed processing will be described. In the present embodiment, the consistent hashing method is used as a method for associating the cluster members with the data, but the present invention is not limited to the consistent hashing method.

First, functions (configuration of processing units) and information possessed by the representative member and other members of the present embodiment will be described with reference to FIG.
As shown in FIG. 10A, the representative member 100 includes a processing unit 110, a storage unit 120, and an input / output unit 130.
The processing unit 110 includes a member management table distribution unit 111, a member management table update unit 112, a representative member specifying unit 113, and another member alive monitoring unit 114.

The member management table distribution unit 111 distributes the member management table 121 to other members.
The member management table update unit 112 updates the member management table 121 at a predetermined opportunity (for example, when the number of members increases or decreases).
The representative member specifying unit 113 specifies a representative member using a representative member selection rule 122 at a predetermined opportunity (for example, when a representative member fails).
The other member alive monitoring unit 114 performs alive monitoring of other members defined by the monitoring destination member setting rule 123.

The storage unit 120 stores a member management table 121, a representative member selection rule 122, and a monitoring destination member setting rule 123.
The input / output unit 130 includes an information input interface, an output interface, and a communication interface.

  The representative member 100 has a technical feature in that the monitoring destination member setting rule 123 is different from the old monitoring destination member setting rule compared to the representative member shown in FIG. 2A (details will be described later). . The other configuration is the same as the corresponding configuration in the representative member shown in FIG.

  Although not shown in FIG. 10A, the representative member 100 also stores the selection rule for the next representative member candidate in the storage unit 120. The representative member 100 can newly select the next representative member candidate as necessary by using the selection rule for the next representative member candidate. Alternatively, the representative member 100 always stores the identifier of the next representative member candidate in a predetermined storage area. When the representative member is changed, the representative member 100 newly selects the next representative member candidate based on the selection rule for the next representative member candidate. The identifier of the next representative member candidate stored in the predetermined storage area may be updated. The next representative member candidate selection rule is, for example, a rule in which the next representative member in the member management table 121 is the next representative member.

As shown in FIG. 10B, the non-representative member 200 includes a processing unit 210, a storage unit 220, and an input / output unit 230.
The processing unit 210 includes a representative member specifying unit 211 and another member alive monitoring unit 212.
The representative member specifying unit 211 uses a representative member selection rule 222 to specify a representative member at a predetermined opportunity (for example, when a representative member fails).
The other member alive monitoring unit 212 performs alive monitoring of other members determined by the monitoring destination member setting rule 223.
The storage unit 220 stores a member management table 221, a representative member selection rule 222, and a monitoring destination member setting rule 223.
The input / output unit 230 includes an information input interface, an output interface, and a communication interface.

  The non-representative member 200 has a technical feature in that the monitoring destination member setting rule 223 is different from the old monitoring destination member setting rule as compared to the non-representative member shown in FIG. (Postscript). The other configuration is the same as the corresponding configuration in the non-representative member shown in FIG.

  Although not shown in FIG. 10B, the non-representative member 200 also stores a selection rule for the next representative member candidate (common to that of the representative member 100) in the storage unit 220.

  Although not shown, the representative member 100 and the non-representative member 200 have a message processing function as a cluster member, a duplicate data transmission function, a management function, and the like.

  In FIG. 10B, the processing unit 210 of the non-representative member 200 does not show a configuration corresponding to the member management table distribution unit 111 and the member management table update unit 112 in the processing unit 110 of the representative member 100. Needless to say, when the non-representative member 200 becomes the representative member 100, such a configuration is also provided.

  In the present embodiment, a monitoring method between cluster members that solves the problems (d) (FIG. 7) and (e) (FIG. 8) described above is proposed. For this purpose, it is effective to avoid the monitoring target of the representative member from becoming the next representative member candidate. In other words, the monitoring destination member setting rule 123 and the monitoring destination member setting rule 223 specify the cluster member that performs life and death monitoring for the next representative member candidate when determining the life and death monitoring target to be performed among a plurality of cluster members. It is a rule determined to select from members other than the representative member at the time, specifically, as follows.

  For example, when the number of cluster members (N) is an odd number, the monitoring target of each cluster member is one skipped (two lower) member of the member management table 121 (same as FIG. 1). In the member management table 121, the next column end (bottom row) is treated as returning to the top.

  That is, as shown in FIG. 11A, considering the case of N = 5, the member 1 (representative member) monitors the member 3. Member 2 monitors member 4. Member 3 monitors member 5. Then, the “N−1” -th member (member 4) immediately before the tail (member 5) monitors the head member (member 1). The last Nth member (member 5) monitors the second member (member 2) from the top. Thereby, it can be easily configured so that the monitoring destination of each cluster member forms a system across all members.

  Here, the “system in which the monitoring destination spans all members” means that the relationship of life and death monitoring performed between a plurality of cluster members starts from an arbitrary cluster member, the monitoring destination cluster member of the cluster member, When the cluster members are traced in order, the cluster members are monitored, and all cluster members are covered before returning to the original starting cluster member.

  As shown in FIG. 11B, when the number of members (N) is an even number, the Nth member (member 4) monitors the first member (member 1), and the (N-1) th member from the top. The second member will be monitored. Thereby, it can be easily configured so that the monitoring destination of each cluster member forms a system across all members. In addition, when the number of members N = 4, there is an effect that the members 2 and 4 monitor each other to prevent a situation in which failure detection is not performed at the time of the simultaneous failure.

  FIG. 12 shows the operation when member 1 (representative member) and member 2 (next representative member candidate) fail simultaneously when the above method is used (when N = 4). The operation subject is each unit in the processing unit 110 and each unit in the processing unit 210, but each member is an operation subject in order to simplify the description.

  First, the member 4 that is alive monitoring the member 1 (representative member) detects the failure of the member 1 (representative member) (S121), and notifies the failure of the member 1 (representative member) to all members (S122). In parallel, the member 3 who is alive monitoring the member 2 detects the failure of the member 2 (S123), and notifies the failure of the member 2 to all members (S124). Member 3 who is the next representative member candidate receives the failure notification of member 1 (representative member) and member 2 (next representative member candidate), recognizes himself as a new representative member (S125), and stores member management table 121. Update (S126), and distribute the updated member management table 121 to all members (S127). Thereafter, the member 3 operates as a representative member.

  As described above, by using this method, failure detection of the representative member and the next representative member is performed in parallel, and a new representative member can be selected at high speed. That is, in the cluster system, even when a representative member that updates and distributes the member management table 121 and another member fail at the same time, a new representative member can be selected at high speed.

  Here, this method is generalized and shown. In this method, as a premise, in order to minimize the number of members that are unable to detect a failure at the time of simultaneous failure of a plurality of members, in addition to not generating a pair to be mutually monitored, as shown in FIG. Monitoring-destination member setting rules 123 and 223 are used so that the monitoring destinations are configured to form a system across all members.

  As a comparative example, when the monitoring destination of each cluster member does not straddle all members, in the example of FIG. 13B, failure detection is impossible when members 1 to 3 fail simultaneously, Even if 4 to N fail simultaneously, failure detection is impossible.

  On the other hand, in this method employing the “system in which the monitoring destination spans all members”, when the member k is the representative member and the member j is the next representative member candidate, the monitoring destination of each cluster member is shown in FIG. Determine according to the procedure shown in. FIG. 14A shows that among the N members, the kth is the representative member and the jth is the next representative member candidate.

  First, the monitoring target member x of the member k (representative member) is determined according to the condition 1 (x ≠ k, x ≠ j) (S141). That is, the monitoring target member x of the member k (representative member) is selected from (N-2) members excluding the member j (next representative member candidate) and itself.

  Next, the monitoring target member y of the member x is determined according to the condition 2 (y ≠ k, y ≠ x) (S142). That is, the monitoring destination member y of the member x is selected from (N-2) members excluding the representative member k that is the starting point of the system and the member x that has already been determined by the monitoring source member.

  Next, it is determined whether the number of members whose monitoring destination is not set is “2 or more” or “1”. If “2 or more”, the process proceeds to S145. If “1”, the monitoring destination of the last member is member k. (Representative member) is set (S144), and the process ends.

  In S145, the monitoring target member z of the member y is determined according to the condition 3 (z ≠ k, z ≠ x, z ≠ y). That is, the monitoring target member z of the member y is selected from (N-3) members excluding the representative member k that is the starting point of the system and the members x and y that have already been determined by the monitoring source member. .

  Next, it is determined whether the number of members whose monitoring destinations are not set is “2 or more” or “1”. If the number is “2 or more”, the same processing (condition 4) as {S142, S143}, {S145, S146} , 5,... Is determined), and if “1”, the monitoring destination of the last member is set to member k (representative member) (S147), and the process ends.

  In this way, by preventing the representative member from monitoring the next representative member and configuring the cluster members to be monitored across all members, failure detection is not possible when multiple members fail simultaneously Can be reduced as much as possible. When N = 2, the representative member and the next representative member candidate are mutually monitored, so the necessary condition of N in the present embodiment is N ≧ 3.

  Since this method can be implemented without major changes to the software of the cluster system, it is effective for general use.

  Although description of this embodiment is finished above, the aspect of the present invention is not limited to these. The specific configuration and processing can be appropriately changed without departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 100 Representative member 110 Processing part 111 Member management table distribution part 112 Member management table update part 113 Representative member specific part 114 Other member alive monitoring part 120 Storage part 121 Member management table 122 Representative member selection rule 123 Monitoring destination member setting rule 130 Input / output Section 200 Non-representative member 210 Processing section 211 Representative member specifying section 212 Other member alive monitoring section 220 Storage section 222 Representative member selection rule 223 Monitoring destination member setting rule 230 Input / output section

Claims (4)

A cluster system in which multiple cluster members perform distributed processing,
Each of the plurality of cluster members has a member management table including identifiers of the plurality of cluster members,
The representative member of the plurality of cluster members updates the member management table, and distributes the updated member management table to each of the plurality of cluster members.
Of the plurality of cluster members, the next representative member candidate that is the next representative member when the representative member fails is selected in advance,
Each of the plurality of cluster members is
A monitoring destination that is determined to select a cluster member that performs life and death monitoring for the next representative member candidate from other than the representative member at the time of determining the life and death monitoring target to be performed among the plurality of cluster members. A storage unit for storing member setting rules;
When determining the life and death monitoring target to be performed among the plurality of cluster members, a processing unit that selects its own monitoring target according to the monitoring member setting rules;
A cluster system comprising: The monitoring member setting rules are:
When the relationship of life and death monitoring performed between the plurality of cluster members starts from an arbitrary cluster member, and follows the cluster member monitoring destination cluster member, and further the cluster member monitoring destination cluster member, The cluster system according to claim 1, wherein the rule is set so as to cover all of the plurality of cluster members and then return to the original starting cluster member. The monitoring member setting rules are:
When the number of cluster members (N) is an odd number, the monitoring target of each cluster member is one cluster member skipped in the descending order of the order in the member management table, and the “N−1” -th previous cluster from the last is The cluster according to claim 1 or 2, wherein the first cluster member is monitored, and the last Nth cluster member is a rule set to monitor the second cluster member from the top. system. The monitoring member setting rules are:
When the number of cluster members (N) is an even number, the monitoring target of each cluster member is
The “N-2” -th cluster member from the top of the order in the member management table is a cluster member that is skipped in descending order of the order in the member management table.
The “N-1” -th cluster immediately before the last is a rule set to monitor the second cluster member from the top, and the last N-th cluster member is a rule set to monitor the first cluster member. The cluster system according to claim 1 or 2, wherein:

Images