WO2020032824A1

WO2020032824A1 - Method for consensus decision making in a distributed computer system

Info

Publication number: WO2020032824A1
Application number: PCT/RU2019/000510
Authority: WO
Inventors: Максим Михайлович МИХАЙЛЕНКО; Игорь Олегович БЕЛОУСОВ
Original assignee: Максим Михайлович МИХАЙЛЕНКО; Игорь Олегович БЕЛОУСОВ
Priority date: 2018-08-08
Filing date: 2019-07-16
Publication date: 2020-02-13
Also published as: RU2706459C1

Abstract

The invention relates to the field of processing digital data using electronic devices, more particularly to the processes of computer systems that are based on specific computation models intended for specific functions, and specifically to self-adaptive control systems, i.e. systems that automatically select an optimal operating mode in order to meet a given criterion. The technical result of the present invention is that of optimizing the procedure for agreeing a common decision between independent equal automated (computerized) centres which is necessary in order to achieve an agreed outcome of a round in a distributed network. A method for consensus decision making in a distributed computer system consists in each node sending a message to all nodes about the start of a round, then each node calculates a new value and sends it to all nodes; if a new value is not calculated, the nodes analyze a common matrix of messages and exclude the messages from nodes identified as exhibiting Byzantine behaviour, whereupon a new value is calculated and a decision is made. The method is applicable in any distributed system where agreed joint action is required, such as cloud systems, blockchain platforms, smart home systems, auctions, and others.

Description

The way to make a single agreed decision in

distributed computer system

FIELD OF THE INVENTION

The invention relates to the field of digital data processing using electrical devices, in particular to methods of computer systems based on specific computational models designed for specific functions, namely self-tuning control systems, i.e. systems that automatically select the optimal mode of operation to achieve a given criterion.

Glossary

In order to ensure the sufficiency of the disclosure of the invention and to enable the information search in relation to the claimed technical solution, the following is a list of terms used in the present description.

A hash is a sequence of characters formed by using a mathematical function that consists in converting input data of an arbitrary length into an (output) line of a fixed, shorter length. With any change in the input data, the hash of this data will be different from the hash of unchanged input data.

Node - a computer that has communications with other computers on which the distributed system software runs.

MTP - confirmation of the Merkle Tree Proof. The Merkle tree is a complete binary tree, in the leaf vertices of which hashes from data blocks are placed, and the inner vertices contain hashes from adding values in child vertices. Merkle tree confirmation is one branch of the full Merkle tree of the block, in which there is information about only one transaction (the confirmation of which is formed), and all other vertices contain only hashes from neighboring branches. When MTP is provided, if all hashes of the nodes are true and the root hashes of the MTP and the block coincide, we can clearly state that the MTP is part of the block.

Valid MTP - an information block that contains MTP digital signatures of more than 50% + 1 system nodes.

A digital signature is a small sequence of characters which, using cryptographic transformations, unambiguously allows you to determine that it was used for signing information and the user's private key signing this information.

The Byzantine task (the task of the Byzantine generals) - (English Byzantine fault tolerance (BFT), Byzantine agreement problem, Byzantine generals problem, Byzantine failure) - in cryptology the task of the interaction of several remote subscribers who received orders from one center. Some subscribers, including the center, may be intruders. It is necessary to develop a unified action strategy that will be beneficial for subscribers.

Byzantine behavior - to coordinate the result of a new round of the system, it is important that all nodes work according to the same rules. The behavior of the node, in which it participates in the work of reconciling a new meaning, sending requests to other nodes and responding to their requests, but at the same time systematically or occasionally allows a departure from the established rules, is usually called Byzantine behavior. This behavior is much more difficult to determine than the usual disconnection of a node from the system. If the host is disconnected from the network, the message simply does not arrive on time. In the case of the Byzantine behavior of the node, the message from such a node comes to other nodes, but the message is completely or partially false. Therefore, additional verification is necessary in order to identify such node behavior.

N, F (50% + 1) - a simplified designation of the ratio of the number of idle nodes and all nodes in a distributed system. N indicate the total number of nodes in a distributed system, F is the number of idle nodes. N = 2F + 1 and 50% + 1 denote such a system, where F more idle nodes account for another F + l (50% + 1) workers. Asynchronous systems are most often described by N = 3F + 1, which means that for normal operation of the system, there must be at least 2F + 1 working nodes on F idle nodes.

System Values — The internal state of the system, expressed as a collection of individual subsets. For example, in distributed ledger systems, this is the aggregate of current registry balances.

Proposed subsets of the new value of the system — a request to change the value of the system in part of a subset of it. For example, in distributed ledger systems, this is a transaction from a client for a change in a balance sheet.

Distributed system (system) - a set of nodes physically located at a distance from each other, having a connection between each other, performing one common function (remoteness strongly depends on the nature of the distributed system, both distribution within the country and inside the submarine are possible)

System round - a time interval measured from the beginning of the coordination of a new value by the system until the end of the coordination of a new value by the system.

State of the art

Determining the optimal from the standpoint of saving time and computing power order of coordination of a common solution between independent peer (computerized) centers is an important modern challenge. Various methods for solving this problem are now used in many industries where it is necessary to minimize random errors of electronic equipment, for example, in space or underwater vehicles. Also, these methods have found applications for building noise-protected distributed computer systems.

Most modern methods for solving this problem are based on the formulated problem and the ways to solve it, which were described in the article “The Byzantine Generals Problem” published in the journal TOPLAS, Vol 4, No 3, in July 1982.

Later in an article in Practical Byzantine Fault Tolerance, published in February 1999. [Practical Byzantine Fault Tolerance, Appears in the Proceedings of the Third Symposium on Operating Systems Design and Implementation, New Orleans, USA, Laboratory for Computer Science, Massachusetts Institute of Technology, February 1999, Miguel Castro, Barbara Liskov], a new, faster way has been described to find solutions to this problem. Based on this method, a lot of applied implementations of decision coordination methods in distributed systems have been created. This method involves the temporary separation of the peers of the system into Primary (Master) and replicas (nodes that store a copy of the state of the system). Also in the method the concept of “View” is introduced: the state of the system in which of all the available nodes one is selected Primary. Obviously, the Primary node may fail, which leads to the creation of a new View with another Primary. All interactions between nodes are assumed to be asynchronous.

The method described in the article, provided that the View does not change in the process, is described in the following steps:

1. The client sends a message to the Primary node. 2. Primary assigns the next N to this message and sends a broadcast message PRE-PREP ARE (pre-preparation) to all replicas.

3. Upon receipt of a PRE-PREPARE message, each node from replicas checks it for compliance with the current View and the N set at the last stage, and if the verification is successful, sends a PREPARE message (preparation) to all nodes.

4. When receiving a PREPARE message, the node checks that it matches the current View and N and the PREPREPARE message it has. If the node receives the corresponding PREPARE message from% of the system nodes, it manages the COMMIT (commit) message for all nodes.

5. When receiving a COMMIT message, the node checks that it matches the current View and N and the PREPARE message it has. If the node receives the corresponding COMMIT message from% of the system nodes, it executes the client message, changing the state of the system and sends the result to the client.

6. The client expects to receive the result of its message from the nodes of the system. When he receives the same messages from the Uz nodes of the system, this will be the final result of processing his request.

The disadvantages of this method:

1. The separation of the system nodes into ordinary nodes and the Primary node necessitated the introduction of additional steps PRE-PREPARE and PREPARE in order to guarantee the operability of this node.

2. Due to the fact that all interactions between nodes are carried out asynchronously, it is not known the final time the client receives the result of the system, which he receives in step 6. This is due to the fact that in step 6, the client must wait 1 to get the result / 3 replies. But if at stages 5 inclusive you can still to set the time periods in which this or that stage should be completed, then in stage 6 there is no need to set such periods, since the system itself does not depend on this. Thus, before the start of the method, it is impossible to determine at what time the client will receive data that his request has been completed.

3. Clients send requests only to the current Primary node. This means that if this node is disconnected, the system will not be operational until a new Primary is assigned.

4. Separation of system nodes into ordinary nodes and the Primary node also lead to the Primary node being able to censor requests from clients, accepting these requests from clients but not transmitting them further.

5. The system can work only if the system has no more than Uz nodes with Byzantine behavior.

6. The system must have an additional subsystem responsible for collecting and delivering to Primary a change in the value of the system from clients of the system, which introduces additional time costs.

On September 11, 2017, the article “Efficient Synchronous Byzantine Consensus” [Efficient Synchronous Byzantine Consensus, Ittai Abraham, Srinivas Devadas, Danny Dolev, Kartik Nayak, Ling Ren, 2017.] was published, in which one more method can be used to solve the problem coordination of decisions in distributed systems.

The method described in the article is based on a four-stage main protocol for each round of interaction:

1. <status> each node sends a message to the leader of this round, which contains the current value accepted by the node, the round number and a confirmation certificate from the end of the previous round (at least F + 1 <notification> or <fix> messages from the previous round from other nodes system) At the end of the stage, the leader collects at least f + I valid messages <status> to form a status confirmation.

2. <position> The leader of this round further forms a new value, which is based on the confirmed state and sends to all nodes of the system, including himself.

At the end of the stage, when receiving a proposal of a new value from the leader, any node checks it and verifies the confirmation of the leader’s state. If the test passes, then the new value is proposed as the new state of this node in this round.

3. <confirmation> If the leader’s <sentence> has successfully passed the test at the end of stage 2, the node sends its decision to all nodes of the system to accept the proposal for a new value in this <commitment> round, accompanied by a copy of the <position> message.

At the end of the stage, if the node receives two messages valid <fix> but with different values from the leader, then the leader exhibits a dual behavior. If the node received f + I messages from other nodes <fixation> with the same new value and does not receive information about the dual behavior of the leader, then it takes this value.

4. <notification> After, at the end of stage 3, the node receives a new value from the leader, he creates and sends out a <notification> about this to all nodes of the system.

At the end of the stage, if the node receives a <notification> from another node, it generates a certificate of a new value (at least f + I <notification> or <fix> messages from the last round from other nodes of the system). If the node receives several valid <notification> messages, it can select any or none of them for the certificate. As you can see from the description, this method is similar to that presented in the article “Practical Byzantine Fault Tolerance”. However, in this article, special attention is drawn to the fact that all the actions described in the stages of the method occur synchronously. This reduces the minimum requirements for the number of workable system nodes from ² / s + 1 for an asynchronous system (article “Practical Byzantine Fault Tolerance”) to! + l for a synchronous system (article "Efficient Synchronous Byzantine Consensus").

The method proposed in the article "Efficient Synchronous Byzantine Consensus" eliminates many of the disadvantages that were typical of the method described in the article "Practical Byzantine Fault Tolerance". To solve problems with the possible appointment of a leader from a part of damaged nodes, it is proposed to fight by constant search of all nodes. Thus, on average, any decision will be taken in 2 iterations of the round (8 stages). It is also suggested that if the leader complies with all the rules accepted for the system (honest leader) not to change it, then stage 4 and 1 can be excluded by receiving 3 stages for one round.

The disadvantages of this method:

1. In the method, if there are F Byzantine nodes, it is proposed to change the leader of the node every round so that all existing nodes pass through the leadership in a row. As a result, the average number of stages to reach an agreement increases to 8.

2. Also, the method proposed another optimization option - when finding an honest leader, do not change it. However, the article does not describe the method of finding such a leader, the number of rounds for which it is proposed to stop choosing an honest leader (freezing the leader) and the method of changing leadership. All of these actions are suggested. develop yourself. What does not allow to directly use this optimization.

3. Highlighting the leader who forms the new value of the system, and especially the proposed optimization with freezing the leader, allows him to censor requests from customers, accepting these requests from customers, but not passing on. This leads to the fact that no matter how much an individual client sends requests, they are not executed. While for most, the work is going in normal mode. Such a system operation is a significant drawback. So, for example, in practice, if such a system provides interaction between stores and suppliers, this flaw in the system can lead to the fact that requests from store A for the delivery of rare goods will be fulfilled, but not from Store B. Currently, this vulnerability is also used at online auctions operating on this system: one bid of one client is missed and the client wins, another bid of the other client is not missed and the client, in principle, does not participate in the auction, even if his bid is actually winning.

4. The system must have an additional subsystem responsible for collecting and delivering to the leader the change in the value of the system from customers of the system, which introduces additional time costs.

The present invention is intended to solve the technical problem of accelerating the process of reliable coordination of the adoption by nodes of a single decision in a distributed system to form a new common value.

The technical result of the present invention is to optimize the procedure for agreeing on a common solution between independent peer-to-peer automatic (computerized) centers, necessary to get a consistent round result in a distributed network.

The specified technical result when using the proposed method is achieved through the use of fewer stages in the round, relative to existing methods used to obtain a consistent round result in a distributed network, which reduces the time spent each round.

Disclosure of Invention

The initial conditions are the same for all nodes, necessary so that a system consisting of these distributed nodes can use this method:

- All interactions for transmitting and receiving information between nodes are set as synchronous.

- Each node in the system has the ability to exchange information over data channels with each other node in the system.

- Each node of the system has information about each other node of the system (address in the communication system between nodes providing data transfer between nodes, a public key for identifying a node in the system), which allows each node to communicate through any data channel with any other nodes in the system .

- Each node of the system has information about each other node of the system, which allows each node to use cryptography to uniquely identify whether a digital signature of the information message belongs to any of the nodes of the system.

- A system consisting of a set of nodes, depending on the method of coordinating the result, can work with a different indicator of the number of failed nodes or nodes with Byzantine behavior. If the number of such nodes is less than the threshold, the system will not be able to work. Such a threshold for the implementation of the claimed method is determined by the formula N> = 2F + 1.

The essence of the proposed method consists in the sequential execution of the round of approval of a new value of the system. Each round, the nodes of the system agree on a new value, passing through the following stages:

1. The first stage <beginning of the round>. Each node of the system sends to each node of the system, including itself, through the transmission channel, a message about the start of the round. This message conveys the node's proposed subsets of the new system value, the number of the new round, and the valid MTP with the number of the previous round and the value of the round repeat counter, if the round has not been completed at least once.

The nodes for processing in the next stage select messages with the maximum number of the previous round. If there are message options with the same maximum round number but different valid MTP with the number of the previous round, the node selects those messages that indicate the maximum value of the round repeat counter.

As a result of the stage, if there are no nodes with Byzantine behavior, each working node receives the same number of messages about the beginning of the round with a new round number from other nodes.

If the system contains nodes with Byzantine behavior, then from such a node other nodes can receive a message with a round number less than the number of the new round. In this case, the node that receives a message from another node with a round number less than the new round number from it, sends a valid MTP with the number of the previous round stored in the database of this node in response to this node.

2. The second stage <end of the round>. Each node, from messages selected at the previous stage, extracts a subset of the new value systems recorded in them. Then each such node sorts these subsets in the same way (the sorting method does not matter, it is only important that the sorting method is the same on all nodes). After that, the nodes check the obtained sequence of subsets for the possibility of calculating a new value. When checking the subsets for the possibility of calculating a new value, it is checked whether it is possible to carry out the calculation based on the given initial conditions and the given operations with them. For example, if there is an initial condition that there is a variable equal to 5, it should be strictly greater than zero, and in the operation it is proposed to subtract 10 (the usual expenditure operation, both financial and any accounting), therefore, it is impossible to calculate a new value, - the operation is not must be performed. There are simpler examples, for example, the operation must be X / Y, and Y = 0, therefore, it is impossible to calculate a new value. Further, a part of the subsets that cannot be executed (a repeated subset, a deliberately false subset, a subset which cannot be calculated based on the current value), are removed from the list of subsets of the value. Then the node calculates the new value and sends to all the nodes, including itself, a message containing the hash of the new value, the round number and the highest value of the round repeat counter from the selected messages and verifying the digital signature of this node.

At the end of the stage, the node accumulates messages from all nodes of the system, containing the hash of the new value, the round number and the value of the round repeat counter. As a result of the accumulation of messages, depending on the number of idle nodes, nodes with Byzantine behavior in the system and the operation of the communication system between nodes, a node may be in the following situations: <round failure> The node received less than 50% of the hash signatures with the same round number and the value of the round repeat counter. If the node finds itself in such a situation, it increases the value of the round repeat counter and begins to negotiate the new value of the round from the very beginning.

<round success> The node received more than 50% of the hash signatures with the same round number and the value of the round repeat counter, and 50% + 1 identical hashes were accumulated. Round coordination is considered successful, and therefore, in a distributed system, a single agreed decision is made. The node is ready to start the execution of a new round of decision-making, if necessary, the value of the counter of repetitions of the round of the node is set to 0.

<round failure, correction required> The node received more than 50% of the hash signatures with the same round number and the value of the round repeat counter, but 50% + 1 identical hashes were not accumulated. This situation signals that nodes with Byzantine behavior have appeared, it is necessary to identify such nodes and end the round. To do this, we use the additional stages of correction below.

If the system contains at least one node with Byzantine behavior, it can at the first stage send messages to the rest with different subsets of a new meaning. This will lead to the fact that no node in the system will collect 50% + 1 identical signed hashes of a new value, which means that until the round ends with an agreed decision on the new value. In order that the system in this case could develop a consistent value, it is proposed to add the correction stages:

1. The first stage <beginning of correction>. A node that did not receive 50% + 1 identical answers at the stage <end of round> creates a message containing the number of the current round and messages received from other nodes at the stage <beginning of a round> containing the subsets of the new system value proposed by the node, the number of the new round, and the valid MTP with the number of the previous round. Next, the node sends such a message to each node.

2. If there are one or several nodes in the system that collected 50% + 1 of the same digital signatures at the stage <end of round>, then such nodes transmit a response message to nodes that did not receive 50% + 1 of the same answers at stage < end of round> and these nodes begin to synchronize with the new value of the system.

3. In the event that none of the nodes in the system collected 50% + 1 identical digital signatures at the stage <end of the round>, then on each working node a common message matrix is collected containing information about all sent from each node to each of the message nodes . Each node, analyzing the general matrix of messages, determines the nodes that sent messages with various subsets of values in this round (nodes with Byzantine behavior).

4. <end of correction> Each node excludes from consideration at this stage messages <beginning of a round> received from nodes identified as (nodes with Byzantine behavior) and then repeats the stage <end of round>.

When using the proposed method, it is possible to achieve coordination of a new value of the system with idle 50% + 1 nodes in 2 stages, if there are nodes in the round with Byzantine behavior in 4 stages. Due to the fact that the client can offer his change in the value of the system to any of the nodes of the system, while not limiting the number of such offers in one round, the impossibility of censorship is achieved. Also in this regard, there is no need to create an additional subsystem for storing and transferring applications from clients to the system. An example of the benefits achieved by using the proposed method is the use of the proposed method for an auction. Suppose you want to hold an auction in which participants who are physically located in any country can participate. At the same time, the auction process should be as automated as possible, fair (no bidder can gain an advantage in bidding by influencing the auction process) and safe (creating significant obstacles to an unfair stopping bidding).

An auction using the proposed method is as follows:

The computer of each participant is a node of the general distributed (decentralized) system. All bidders are equal and are identified using cryptographic keys. The announcement of the lot is accompanied by the announcement of its price, as well as auction rules (type of auction, auction step, number of auction rounds before winning).

Auction begins with the fact that one or more bidders offer to put up a lot for auction. In the case of simultaneous bidding, the first item to be auctioned is a lot with a lower initial price.

Auction work takes place in rounds. Each round, the participant can either offer his action for this round and not offer it.

At the beginning of the round, all participants through their nodes send the remaining participants a signed message with their proposal, if there is no proposal, a message is sent with the proposal “no proposal”.

Participants accumulate a message from all nodes, check signatures for these messages. Then retrieved from messages all proposals, sort them, and begin to evaluate the result of this round. The result of the round may be:

1. there is a winner of the round - who offered the best price in this round

2. no one has offered the best price and the round counter to win is greater than 0

3. no one has offered the best price and the round counter to win is 0

After finding out the result of the round, the node sends it in a signed message to all other participants. If the node received the same result from more than 50% of the participants, the participant considers this round completed and proceeds to work on the next round.

As for the logic of work in determining the winner of the lot, it is presented in Figure 1.

In general, the inventive method can be applied in any distributed systems where an agreed joint solution is needed. For example, when processing information from sensors to obtain stable information about them, even taking into account the failure of a piece of equipment. In turn, this use of the method can be successfully applied to create reliable systems of “smart homes”.

In addition, this method can be used to create conventional Internet services with increased reliability. For example, cloud systems, blockchain platforms or distributed registry platforms.

Claims

Claim

A method for making a single coordinated decision in a distributed computer system (nodes), in which all interactions for transmitting and receiving information between nodes are set as synchronous, with each node in the system having the ability to exchange information over data channels with each other node in the system, each node in the system has information about every other node in the system (address in the communication system between nodes providing data transfer between nodes, a public key for identifying a node in the system), allowing each a node to communicate through a data transmission channel with each other node that is part of the system, each node of the system has information about each other node of the system, which allows each node to use cryptography to uniquely identify whether a digital signature of an information message belongs to any of the nodes of the system, and the threshold indicator of quantity failed nodes or nodes with Byzantine behavior, in which the system cannot make a single coordinated decision using the proposed method, corresponds to rmule N> = 2F + 1, where N is the total number of nodes in a distributed system, a F denotes the number of idle nodes, characterized in that it comprises:

stage of the first stage <beginning of a round>, at which each node of the system sends to each node of the system, including itself, a transmission message about the beginning of the round containing the subsets of the new system value proposed by the node, the number of the new round, valid MTP with the number of the previous round and the value of the counter for repeating the round, if the round has not been completed at least once, while the nodes for processing in the next stage select messages with the maximum number of the previous round, and if there are message options with the same max the round number, but different valid MTP with the number of the previous round, the nodes select those messages that indicate the maximum value of the counter of round repeats, as a result of which, if there are no nodes with Byzantine behavior, each working node receives the same number of messages about the beginning of the round with a new round number from other nodes, and if the system contains nodes with Byzantine behavior, then from such a node other nodes receive a message with a round number less than the number of a new round from him, and in this case more often, a node that receives a message from another node with a round number less than the number of a new round from it, sends a valid MTP with the number of the previous round stored in the database of this node in response to this node;

stage of the second stage <end of round>, in which each node, from messages received from other nodes and selected at the previous stage, extracts subsets of the new system value recorded in them, and then selects from the subsets proposed by other nodes the new value of the subset system with the maximum round number, then each such node sorts these subsets in the same way, after which the nodes check the resulting sequence of subsets for the possibility of calculating a new value, after which part of the subset which cannot be executed are removed from the list of subsets of values, and then each node calculates a new value and sends to all nodes, including itself, a message containing a hash of the new value, round number, the largest value of the round repeat counter from the selected messages and verifying the digital signature of this node, after which, as a result of this stage, each node accumulates messages from all nodes of the system containing the hash of the new value, the round number and the value of the round repeat counter, after which, if any of the nodes received less than 50% of hashes with the same round number and the value of the round repeat counter, then such a node increases the value of the repeat counter of this round by one unit and starts reconciling the new value of the round from the very beginning, if any of the nodes received more than 50% of the hash signatures with with the same round number and the value of the round repeat counter, and 50% + 1 of the same hashes were accumulated, then the round coordination is considered successful, and therefore, in a distributed system, a single agreed decision is made, and in that case, if any of the nodes received more than 50% of the hash signatures with the same round number and the value of the round repeat counter, but did not accumulate 50% + 1 identical hashes, then there are nodes with Byzantine behavior in the system, and in order to identify such nodes and finish round, go to the additional stage of the first stage <start correction>;

an additional stage of the first stage <correction start>, which occurs if at least one of the nodes has not received 50% + 1 of the same answers at the stage <end of the round>, at which the node that has not received 50% + 1 of the same answers other nodes at the stage <end of a round>, creates a message containing the number of the current round and messages received by it from other nodes at the stage <beginning of a round> containing the subsets of the new system value proposed by the node, the number of the new round, and a valid MTP with the number of the previous round, then such a The e-mail sends such a message to each node, after which if there are one or more nodes in the system that have collected 50% + 1 identical digital signatures at the <end of round> stage, such nodes transmit a response message to the nodes that have not received 50% + 1 identical answers at the stage <end of a round> and these nodes begin synchronization with the new value of the system, and if none of the nodes in the system collected 50% + 1 identical digital of signatures at the stage <end of round>, then on each working node a common message matrix is collected containing information about all sent by each of the nodes to each of the message nodes and each node, analyzing the general message matrix, determines the nodes with Byzantine behavior, sending other message nodes with different subsets values at the stage <beginning of the round>;

an additional stage of the second stage <end of correction>, in which each node excludes from consideration at this stage messages received at the stage <beginning of round> from nodes identified in the previous stage with Byzantine behavior, and then repeats stage <end of round>.