JP2011095976A - Device, method and program for distributed data management - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a high fault tolerance without data loss against a fault in entire applications using a distributed data management program, which handles data of which loss may cause trouble, and to improve use efficiency of each node. <P>SOLUTION: In a certain physical node, a node eigenvalue common to the entire plurality of virtual node ID obtained by a hash function is set, while a data decision method managed by each node (the node having a node ID close to a data ID manages an object data) is maintained in the distributed data management program. Also, at the time of data ID calculation being duplicated from certain data, the probability of nodes which manages the original and the duplication of data becoming physically different is increased by the change of a point corresponding to the node eigenvalue of the data ID. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a distributed data management apparatus, method, and program, and more particularly, to a distributed data management apparatus, method, and program for managing data distributed on a plurality of nodes (such as servers) on a network.

  The distributed data management program that is the subject of the present invention is used to manage data across a plurality of nodes (computers, etc.) in various networks using IPs typified by the Internet. This is called a hash table in computer science. For some data, data management technology that holds key information and data contents as a pair is retained in a distributed environment with multiple nodes. It is distributed and managed at each node. As a result, while the actual processing is distributed to each node, the data commonly used among the nodes can be shared by the distributed data management program. Therefore, the field of application covers the entire distributed computing such as parallel / distributed processing.

  In the computer science, the term distributed data management program used in this specification is consistent hashing (see, for example, Non-Patent Document 1), distributed hash table (see, for example, Non-Patent Document 2), distributed key, value, This is a general term for a method called a store (see, for example, Non-Patent Document 3) and programs that implement the method. A feature common to these is that a data management method called a hash table is divided and managed on a plurality of physically distributed nodes.

  A hash table has information called a key for identifying certain data and a key as a pair, and uses a value (hash value) obtained by giving this key to a function called a hash function as an index. A hash function is a function that returns an output with no law for the input. As a result, even when similar keys are consecutive, the hash value used as the key index becomes a value irrelevant to the continuity of the queue, preventing data from being concentrated in a specific area.

  In consistent hashing, this hash table is not held by one physical node (unit such as a computer) but is divided and held by a plurality of nodes. Which node is responsible for which area of the divided hash table is determined by the hash value (this specification) obtained by applying the ID (set as node ID in this specification) set to the node and the data key to the hash function. In this case, it is determined by the distance to “data ID”.

  In a typical example, a node having a certain ID is responsible for data having a data ID located in an area between its own ID and the ID of the nearest node among nodes having an ID smaller than its own ID. It is a mechanism to do.

  The distributed hash table is an extension of the consistent hashing concept. In consistent hashing, a node manages the IDs and addresses of all other nodes in a table in order to communicate with other nodes. On the other hand, the management of this address table becomes more expensive and difficult as the total number of nodes increases. Therefore, in the distributed hash table, management is simplified by managing not only the hash table but also a part of the address table, not the whole. On the other hand, since many functions of the data management method are the same as those of consistent hashing, in the present invention, both are similarly identified as a distributed data management program in one word.

  For the distributed key / value store, for example, data replication, caching, and transaction management are expanded to a data management method distributed in the same way as consistent hashing. Since the fundamental data management method is the same as that of consistent hashing, in the present invention, these three technologies are collectively referred to as a distributed data management program.

  In the distributed data management program, since data is generally managed by a node having a node ID close to the data ID, the deviation of the node ID appears as the deviation of the data management amount between the nodes as it is. The range of values that can be taken by the node ID and the data ID is referred to as “ID space”, and the above-mentioned bias can be expressed as the bias of the node ID in the ID space.

  In the distributed data management program, a probabilistic measure is generally taken to assign a plurality of virtual node IDs to one physical node in general for such a bias of node IDs in the ID space. That is, the bias is averaged and mitigated by the fact that a large number of virtual node IDs of a physical node are scattered in the ID space according to the law of large numbers. In proportion to the number of virtual node IDs assigned to one node, the distribution bias in the ID space is reduced.

  Here, one form of usage of the distributed data management program will be described. In a distributed data management program, one hash table is divided and managed by multiple nodes, so when a particular node finishes its role as a hash table component due to a failure or program termination, that node manages it. The data that was being used is generally lost. There is no problem if the distributed data management program is used as a cache and the original data is managed by another device, but the original data is managed by the distributed data management program and the data is lost. If this is not possible, the loss of data associated with the removal of this component is fatal.

  For this problem, for data that you don't want to lose in general, create a copy of the data and give it another data ID, so that it is registered in the hash table as new data completely different from the original data. A usage form is used. Here, the duplicate data ID is generated so as to be inferred from the original data ID. For example, when a hash function for calculating the data ID from the data key is set as f (x), the copy is calculated by applying the original data key to another hash function as in the function g (x). A calculation method can be considered. In this form, even when the node that manages the original of certain data leaves, the node that has acquired the data calculates the data ID of the duplicate of the data according to the calculation logic as described above. A copy of the data can be obtained from the node that manages the data.

  If a node that manages the original and duplicate of a data is disconnected at the same time, for example, due to a failure, both the original and duplicate data are lost, and the data is completely lost. In general, a plurality of copies are generated for a certain data against such a simultaneous departure, thereby avoiding the probability that the original and all copies are lost at the same time. In other words, in proportion to the number of replicas generated for one data, the probability that either the original or the replica of certain data exists in the ID space increases.

D. Karger, E. Lehman, T. Leighton, M. Levine, D. Lewin, and R. Panigrahy, "Consistent Hashing and Random Trees: Distributed Caching Protocols for Relieving Hot Spots on the World Wide Web," in Proceedings of the 29th ACM Symposium on Theory of Computing (STOC'97), pp.654-663, May 1997. I. Stoica, R. Morris, D. Karger, MF Kaashoek, and H. Balakrishnan, "Chord: A Scalable Peertopeer Lookup Service for Internet Applications," in Proceedings of the Annual Conference of the Special Interest Group on Data Communication (SIGCOMM ' 01), pp.149-160, August 2001. G. DeCandia, D. Hastorun, M. Jampani, G. Kakulapati, A. Lakshman, A. Pilchin, S. Sivasubramanian, P. Vosshall and W. Vogels, "Dynamo: Amazon's Highly Available Key-value Store," in Proceedings of the 21st ACM Symposium on Operating Systems Principles (SOSP'07), pp.205-220, October 2007.

  In order to avoid the bias of node IDs in the ID space described above, when considering using a plurality of virtual node IDs for physical nodes and using data replication simultaneously, the original data and There is a possibility that the physical nodes that manage the replicated data will be the same. Since the hash function for generating the ID of the node or data has no law with respect to the input, the original and duplicate data IDs of certain data are mapped to different virtual node IDs of a certain physical node. there is a possibility.

  When such a mapping is made, the distributed data management system in which the physical node that manages both the original and the copy with different virtual node IDs is composed of nodes of the distributed data management program due to a failure or the like. In the event of withdrawal, the node that manages the target data does not exist on the system and the data will be lost with a certain probability even though data replication is used. End up.

  In addition, even if data is not lost, the probability that data will be lost when the node that manages the remaining replicas leaves because a certain amount of the original and several replicas are lost from the system at the same time. There is no change in raising.

  This problem is limited to ad hoc measures such as recalculating the data ID if it is likely to be managed from the same physical node when calculating the data ID of the data copy. This is because there is a possibility that a node always joins and leaves, and a node that manages data with a data ID that accompanies joining or leaving keeps changing.

  There are two types of conventional solutions to this problem: probabilistic means and those that give a certain rule to the ID generation rule.

  The former (probabilistic means) increases the number of duplicates generated so that if the same physical node manages several originals or duplicates simultaneously, some remaining duplicates are It increases the probability that different nodes manage.

  Although this approach reduces the probability that certain data will be completely lost from the system, the amount of data managed by each node is proportional to the number of replicas, so the amount of data managed by the entire system is the maximum amount of data. On the other hand, if it is large, this replication cost cannot be ignored.

  The latter (giving a certain rule to the ID generation rule) uses a node-specific value for the most significant bits for the virtual node ID of a physical node, and between the original and each copy of the data. Then, by changing the value of the most significant bit corresponding to a certain rule, the probability that a certain physical node simultaneously manages the original and duplicated data is reduced.

  For example, if the node-specific value of a physical node is “0101” (in binary notation), the upper 4 bits of all virtual node IDs of that node are all “0101”. On the other hand, when the original data ID of certain data starts with “0101”, this data is managed by the node having the node specific value “0101” with high probability. At this time, the probability that data replication is managed in another physical node having another node specific value by adding a mechanism that always changes the upper 4 bits to the data ID calculation rule for data replication. Becomes higher.

  This approach has the advantage that the amount of data managed throughout the system does not change and the probability of data being lost from the system is reduced because the number of replicas does not need to be increased over typical usage.

  On the other hand, since the rule of generating node IDs is given a rule of unique node eigenvalues for each physical node, the deviation of node IDs in the ID space is all caused by this deviation of node eigenvalues. It will be. Since there is one node-specific value for each physical node, no matter how many virtual node IDs are assigned to each node, if the node-specific values are biased, avoid the bias of node IDs in the ID space. It has a problem that it cannot be done.

  The present invention has been made in view of the above points. When managing data distributed on a plurality of nodes (such as servers) on a network, the present invention suppresses the bias of node IDs in the ID space, and An object of the present invention is to provide a distributed data management apparatus, method, and program capable of reducing the probability that the same node manages both the original and the replica of data.

  In the present invention, when this distributed data management program is used, the problem that occurs in the method of enhancing the fault tolerance of data at the time of a node failure by placing a copy of the data is to generate the node ID and the ID of the data copy. The calculation method is subjected to processing more than a simple hash function to solve it probabilistically.

  FIG. 1 is a principle configuration diagram of the present invention.

According to the present invention (claim 1), a node having a node ID close to a data ID that is a hash value obtained using a hash function that is a function that returns an output having no law unrelated to the law of the input value is subject data. A distributed data management device for managing
Hash function storage means 900 for holding a hash function;
A virtual node ID generation unit 1000 that generates a plurality of virtual node IDs set for a physical node;
The virtual node ID generation unit 1000
A node unique value calculated using the hash function of the hash function storage unit 900 is set to a unique value for each physical node between the upper random value and the lower random value calculated by an arbitrary means. Thus, node specific value insertion means 1010 for generating a virtual node ID is provided.

Further, the present invention (claim 2) is a communication means for acquiring the data ID, key, and data of the data to be handled by the own device from another node;
A copy data ID calculating means for calculating a copy ID indicating data copy;
The duplicate data ID calculation means
Eigenvalue partial conversion means for converting a portion corresponding to the node specific value of the original data ID acquired by the communication means by combining a value of a node ID different from the node ID holding the original data with a hash value Have.

According to the present invention (Claim 3), a node having a node ID close to a data ID that is a hash value obtained by using a hash function that is a function that returns an output having no law unrelated to the law of the input value is subject data. A distributed data management method in a device for managing
The virtual node ID generation means of the device is
The virtual node ID is set by setting a node unique value calculated by using a hash function to a unique value for each physical node between the upper random value and the lower random value calculated by an arbitrary means. Is generated.

According to the present invention (claim 4), the duplicate data ID calculation means of the device
Obtain the data ID, key, and data of the data that the device should be in charge of from other nodes via communication means,
A portion corresponding to the node specific value of the original data ID acquired by the communication means is converted.

  The present invention (Claim 5) is a distributed data management program for causing a computer to function as each means of the distributed data management apparatus according to Claim 1 or 2.

  As described above, the present invention has a unique value for each physical node in the remaining several bits between the upper few bits and the lower few bits in the middle of the node ID, and all the virtual node IDs of the node have a unique value. Has a common value (node-specific value). Further, the data ID of the duplicate of the data is calculated by changing the portion corresponding to the node specific value from the original data ID. As a result, it is possible to achieve both improvement in data redundancy by data replication in the distributed data management program and avoidance of ID bias in the ID space by generation of virtual node IDs. In all applications that handle data that is difficult for us, it is possible to perform data management with high fault tolerance and excellent use efficiency of each node without data loss even in the event of a failure.

It is a principle block diagram of this invention. It is a figure which shows the structure of node ID. 1 is a system configuration diagram according to an embodiment of the present invention. It is a block diagram of the node in one embodiment of this invention. It is an example of generation of virtual node ID in one embodiment of the present invention. It is a generation example (part 1) of the duplicate data ID in the distributed data management program according to the embodiment of the present invention. It is a generation example (the 2) of replication data ID in the distributed data management program in one embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 2 shows the configuration of the node ID. FIG. 2A shows a conventional node ID 200, and FIG. 2B shows the configuration of the node ID 100 of the present invention.

  In the present invention, in the present invention, a unique node unique value 120 is set for each physical node in the remaining few bits between the upper few bits 110 and the lower few bits 130 in the middle of the node ID 100 (FIG. 2). (B)). This node unique value has a unique value for each physical node as well as the node unique value set in the most significant bit in the existing solution described above, and has a value common to all virtual node IDs of the node.

  For example, when the node unique value of a certain node is “0010” and this corresponds to the 9th to 12th bits of the node ID, the 9th to 12th bits of all the virtual node IDs of the node are “0010” which is the node unique value. Become.

  Also, the method for calculating the data ID of the data copy is calculated by changing the part corresponding to the node specific value from the original data ID.

  For example, if the node unique value in the node ID is present in the 9th to 12th bits, the data ID of the data copy takes a different value in the 9th to 12th bits. If the 9th to 12th bits of the data ID of the original data are “1011”, the 9th to 12th bits of the duplicate data ID are “1011” such as “1100”, “1101”, “1110”. Take a value other than At this time, regarding the method of taking a value other than “1011” and the legality, it is necessary to be “separated” from the corresponding part in the original data ID. Therefore, the method of changing the value is not limited in this specification. A typical example will be described later.

  The distributed data management program of the present invention is a program on a widely used computer (node) such as a personal computer or a server, and the programs executed on a plurality of nodes cooperate to manage data. Is.

  From an application that uses a distributed data management program for data management, it appears to be a data management system that looks like a single hash table, and can be used for data management without being conscious of being actually distributed. it can.

  The distributed data management program is assumed to be used mainly in a local area network based on the Internet or Internet Protocol (IP).

  FIG. 3 is a system configuration diagram according to an embodiment of the present invention, and shows the relationship of nodes of the distributed data management program connected to the network 1200. There is no master-slave relationship such as which is the master and which is the slave between the nodes (300A to 300F), and the relationship of all the nodes is equal. Depending on the bias of the ID space, there is a possibility that the amount of data handled between the nodes may vary, but the role of each node is equal.

  FIG. 4 is a configuration diagram of a node according to an embodiment of the present invention. Hereinafter, each function with which the node 300 is provided is demonstrated. The node 300 includes a local node ID management unit 400, another node management unit 500, a hash table management unit 600, a communication unit 700, a search unit 800, a hash function 900, a virtual node ID generation unit 1000, and a duplicate data ID calculation unit 1100. . Among these components, the own node ID management unit 400, the other node management unit 500, the hash table management unit 600, the communication unit 700, the search unit 800, and the hash function 900 are common to the conventional configuration. The parts specific to the present invention are a node unique value insertion unit 1010 in the virtual node ID generation unit 1000 and a unique value part conversion unit 1110 in the duplicate data ID calculation unit 1100.

  The own node ID management unit 400 stores the own node ID 410, the other node management unit 500 stores the node ID 510 and the address 520, the hash table management unit 600 stores the data ID 610, the key 620, and the data value 630, respectively. Is stored. The hash function 900 is assumed to be stored in a storage unit such as a memory.

  When the node 300 calculates its own node ID at the time of activation, the node 300 is connected to the distributed data management system configured by another node using the communication unit 700, and the current state of the system and information on other nodes Etc.

  The communication unit 700 is used when each node executing the distributed data management program communicates with other nodes to exchange information.

  The own node ID is generally calculated using a hash function 900 by inputting a value unique to the node such as an IP address or other unique identifier. As described above, the hash function referred to here is a function that returns an output having no law that is not related to the law of the input value. Typical hash functions include SHA-1 and MD5. However, in the method of the present invention, the hash function 900 is not limited to a specific hash function, and the hash value obtained by the hash function 900 is large. It is not limited to.

  The own node ID management unit 400 has a function of managing the node ID 410A of the node itself calculated by such a hash function 900, and manages its own virtual node ID, which will be described later, as node IDs 410B to 410C.

  After starting the distributed data management program, the node 300 calculates its own node ID, and from the other nodes, the current state of the system (information on other participating nodes and data to be managed by the participating nodes). Information) is received using the communication unit 700.

  Information on other participating nodes (node ID for each node and communication destination information (address) of the node used in the communication unit 700) is stored as a pair of the node ID 510 and the address 520 in the other node management unit 500 of the node 300. The other nodes are managed as node IDs 510A to 510C and addresses 520A to 520C.

  Here, when a mechanism such as a virtual node ID is used, a form in which the virtual node ID / address pair is managed for each physical node in the other node management unit 500 is also conceivable. Since management for each node is not a necessary condition, it is simplified and described as a pair of node ID and address. In this case, a plurality of virtual node IDs of other physical nodes appear discontinuously in the other node management function, and in some cases are stored in a form indicating the same address.

  Note that the node ID formats in the own node ID management unit 400 and the other node management unit 500 are omitted here, but have a structure corresponding to the node ID 100 in FIG.

  The node 300 receives, from the other nodes via the communication unit 700, three pieces of information, that is, data ID, key, and data value, for data to be handled by the node 300.

  This is managed in the hash table management unit 600 as data IDs 610A to 610C, keys 620A to 620C, and values 630A to 630C as combinations of the data ID 610, the key 620, and the data value 630 for each data. The form of the data keys 610A to 610C is basically the same as that of the node ID 100, and the location corresponding to the node unique value 120 is copied by the unique value partial conversion unit 1110 of the duplicate data ID calculation unit 1100 described later for each data copy. It becomes a form to convert.

  After the initialization process performed at the time of starting these distributed data management programs is completed, the communication unit 700 receives a data search request or storage request from the application, and performs a corresponding process.

  The search request is sent to the search unit 800 of the node 300. The search request is to obtain data corresponding to the data key 620, and the search unit 800 of the node 300 searches the other node management unit 500 for a node that manages the data ID 610 obtained by applying the key 620 to the hash function. If the node itself manages the data specified by the key 610, the corresponding data is extracted from its own hash table management unit 600 and transmitted to the request source. If the data is managed by a node other than itself, the address of the node is acquired by referring to the other node management unit 500, and the request is transferred to this node.

  At this time, if the node that should manage the target data has left the system due to a failure or the like and the target data cannot be acquired, the replication data ID calculation unit 1100 described later sets the data ID 610 of the replication. The request is calculated from the key 620, and the request is transferred to the node that manages the duplicate data ID 610. In addition to obtaining such duplicate data only when such a request fails, the request is transferred to the original and a certain number of duplicates for speedup, and the data returned most quickly as the result of the request. The form of adoption is conceivable.

  The data storage request from the application starts by calculating the data ID by applying the key of the data passed from the application to the hash function. The node that should manage the data ID obtained here is acquired by referring to the other node management unit 500, and the data ID and the key and value passed from the application are sent to the address of the node. The node that has received this data manages this data in its own hash table management unit 600.

  When data replication is used, the replication data ID calculation unit 1100 calculates the replication data ID when a storage request is received. A specific example of the calculation method in the eigenvalue partial converter 1110 will be described later.

  Thereafter, similarly to the operation at the time of storing the original data, the node that manages the generated duplicate data ID is acquired by the other node management unit 500, and the duplicate data ID, key, and value are sent to the target node. . The node that received the duplicate data manages this duplicate data in its own hash table management unit 600.

  In order to maintain consistency between the original and duplicate of data, it is generally possible to apply some kind of transaction management mechanism. Therefore, in the present invention, the order of writing the original and duplicated data is determined in this embodiment. It is not limited to the order.

  Hereinafter, examples of operations of the node eigenvalue insertion unit 1010 and the eigenvalue partial conversion unit 1110 according to the present invention will be described.

  FIG. 5 shows a generation example of the virtual node ID according to the embodiment of the present invention, and shows a generation example of the virtual node ID by the virtual node ID generation unit 1000 and the node unique value insertion unit 1010. Here, as an example, the node ID is a 32-bit value. In this example, the upper 8 bits are the upper random value 110, the subsequent 4 bits are the node specific value 120, and the lower 20 bits are the lower random value 130. Here, when the node IDa is the original node ID, the node unique value (0110) in this node IDa is fixed as the node unique value in this physical node. Thereafter, when calculating the virtual node ID (node IDb, node IDc,...), The node specific value portion (the upper 9th to 12th bits) is always fixed to this node specific value (0110).

  The upper random value and the lower random value at the time of calculating the virtual node ID are random values calculated by any means. In general, a return value or the like obtained by multiplying a hash function with an input of a virtual node ID order at the end of a node IP address is used.

  Next, a method for calculating a duplicate data ID by the eigenvalue partial converter 1110 in the duplicate data ID calculator 1100 will be described with reference to the examples of FIGS.

  FIG. 6 shows a generation example of a duplicate data ID in the distributed data management program according to an embodiment of the present invention (when ID is treated as a generation and the difference is set as a distance), and FIG. An example of generation of duplicate data ID in the embodiment (when the value of the exclusive OR between ID bit strings is set as the distance) is shown. Here, the duplicate is determined according to the criteria for determining which node stores which data. The calculation method of the data ID changes.

  FIG. 6 shows the most common node ID and data ID as integers, and the data at a node ID closest to the data ID, or a node having a closest node ID whose integer value is smaller than the data ID. This is an example of generating a duplicate data ID in a distributed data management program having a determination criterion of storing.

  A portion (9th to 12th bits) corresponding to the node unique value in the node ID of the original data ID (original data ID) is (0010). The duplicate data ID is generated by converting a portion corresponding to the node eigenvalue by the eigenvalue partial conversion unit 1110. In this example, the eigenvalue part (0010) in the data ID is multiplied by the exclusive OR (XOR) of the bit masks (1000, 0100, 1100, 0010) for each duplicated data ID, so that the eigenvalue part is (1010). , 0110, 1110, 0000).

  As a result, the duplicate data ID originally has a value that is very similar to the upper 8 bits of the data ID, but the 4 bits corresponding to the unique value take values that are far from each other in terms of the distance due to the difference between the data ID and the integer. Thus, the probability that a physical node that originally has a data ID as one of the virtual node IDs also manages duplicated data is reduced.

  FIG. 7 considers a value (integer) of an exclusive OR of a node ID and a data ID as a distance between IDs, and has a determination criterion that data is stored in a node having a node ID with the smallest distance. It is an example of generation of duplicate data ID in the data management program.

  The portion (9th to 12th bits) corresponding to the node unique value in the node ID of the original data ID (original data ID) is (0010) as in the above example. The duplicate data ID is generated by converting a portion corresponding to the node eigenvalue by the eigenvalue partial conversion unit 1110. In this example, the eigenvalue part (0010) in the data ID is multiplied by the exclusive OR (XOR) of the bit masks (1111, 0111, 1011, 0011) for each duplicated data ID, so that the eigenvalue part is (1101). , 0101, 1001, 0001).

  As a result, the duplicate data ID originally has a value that is very similar to the upper 8 bits of the data ID, but takes a value that is far from the original data ID and the concept of exclusive OR for the 4 bits corresponding to the unique value. In other words, the probability that a physical node that originally has a data ID as one of the virtual node IDs also manages duplicated data is reduced.

  By the method shown in the above embodiment, the bias of the node ID on the ID space in the above-mentioned problem is suppressed, and the probability that the same node manages both the original and the replica of the data is reduced. These two features are in a trade-off relationship, and if the size of a random value (a value without a law property obtained by a hash function) located at the top of the node eigenvalue is small, the bias of the ID space becomes large, On the other hand, when the size is increased, the probability that the same node manages both the original and the copy increases.

  In the present invention, since only the node ID generation method and the data ID generation / calculation method are changed, the specifications of the other distributed data management programs can be inherited as they are. It can be applied to all tables and distributed key / value stores.

  Further, the function of each component of the node 300 shown in FIG. 4 can be constructed as a program, installed in a computer used as a node and executed, or distributed via a network.

  Further, the constructed program can be stored in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, and can be installed or distributed in a computer.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

100 Node ID
110 Upper random value 120 Node specific value 130 Lower random value 200 Node ID
210 Random value 300 Node 400 Own node ID management unit 410 Node ID
500 Other node management unit 510 Node ID
520 Address 600 Hash table management unit 610 Data ID
620 Key 630 Data value 700 Communication unit 800 Search unit 900 Hash function 1000 Virtual node ID generation unit 1010 Node eigenvalue insertion unit 1100 Duplicate data ID calculation unit 1110 Eigenvalue partial conversion unit 1200 Network

Claims (5)

A distributed data management apparatus in which a node having a node ID close to a data ID that is a hash value obtained using a hash function that returns an output having no law unrelated to the input value law manages target data. And
A hash function storage means for holding a hash function;
A virtual node ID generating means for generating a plurality of virtual node IDs to be set in a certain physical node;
The virtual node ID generation means includes
A node unique value calculated using the hash function of the hash function storage means is set to a unique value for each physical node between the upper random value and the lower random value calculated by any means. A distributed data management apparatus comprising node unique value insertion means for generating the virtual node ID. A communication means for acquiring the data ID, key, and data of the data to be handled by the own device from another node;
Copy data ID calculating means for calculating a copy ID indicating a copy of the data,
The duplicate data ID calculation means
Eigen value partial conversion means for converting a portion corresponding to the node unique value of the original data ID acquired by the communication means by combining a value of a node ID different from the node ID holding the original data with a hash value 2. The distributed data management apparatus according to claim 1, further comprising: A distributed data management method in a device in which a node having a node ID close to a data ID that is a hash value obtained using a hash function that is a function that returns an output having no law unrelated to the input value law manages target data Because
The virtual node ID generation means of the device is
By setting a node unique value calculated using the hash function to a unique value for each physical node between the upper random value and the lower random value calculated by any means, the virtual node A distributed data management method characterized by generating an ID. The duplicate data ID calculation means of the device is
Obtain the data ID, key, and data of the data that the device should be in charge of from other nodes via communication means,
4. The distributed data management method according to claim 3, wherein a portion corresponding to the node specific value of the original data ID acquired by the communication means is converted.   A distributed data management program for causing a computer to function as each means of the distributed data management device according to claim 1.

Images