FR3061391A1 - COMPUTER NETWORK OF KNOTS COMMUNICATING BETWEEN THEM THROUGH PAIR MESSAGES AND INTERCONNECTION METHOD BETWEEN NODES ASSOCIATED THEREWITH - Google Patents

Abstract

The invention relates to a computer network of nodes (1, 2) communicating with each other by peer-to-peer messages, message generating nodes (1) sending their generated messages to message consumer nodes (2), a consumer node ( 2) can also be a producer node (1) at another level of the network (6), characterized in that each producer node (1) of message: chooses, among the nodes (1, 2) registered as active in the network , its consumer nodes (2) to interconnect with them, maintains the list of its interconnected consumer nodes (2), manages the load distribution of its product messages to its interconnected consumer nodes (2), as a function of time processing, by its interconnected consumer nodes (2), of its generated messages, by sending its messages produced especially to its interconnected consumers nodes (2) with a lower processing time, modifies its list of consuming nodes interconnected nodes (2) based on the processing time, by its interconnected consumer nodes (2), of its generated messages, by deleting from its list at least temporarily its interconnected nodes consumers (2) too slow.

Description

Holder (s):

WORLDLINE Public limited company.

O Extension request (s):

® Agent (s): CABINET PLASSERAUD.

® COMPUTER NETWORK OF NODES COMMUNICATING BETWEEN THEM BY PEER-TO-PEER MESSAGES AND ASSOCIATED INTERCONNECTION BETWEEN NODES.

FR 3,061,391 - A1 (57) The invention relates to a computer network of nodes (1, 2) communicating with each other by peer-to-peer messages, message producing nodes (1) sending their produced messages to consumer nodes ( 2) message, a consumer node (2) can also be a producer node (1) at another level of the network (6), characterized in that each message producer node (1): chooses, from among the nodes (1 , 2) registered as active in the network, its consumer nodes (2) to interconnect with them, keeps the list of its interconnected consumer nodes (2) up to date, manages the load distribution of its produced messages to its consumer nodes (2) interconnected, according to the processing time, by its consumer nodes (2) interconnected, of its produced messages, by sending its produced messages especially to its consumer nodes (2) interconnected with lower processing time, modifies its li ste of consumer nodes (2) interconnected as a function of the processing time, by its consumer nodes (2) interconnected, of its produced messages, by removing from its list at least temporarily its consumer nodes (2) interconnected too slow.

REGISTER

COMPUTER NETWORK OF NODES COMMUNICATING BETWEEN THEM BY PEER-TO-PEER MESSAGES AND METHOD OF INTERCONNECTING BETWEEN NODES

FIELD OF THE INVENTION

The invention relates to the field of node networks communicating with each other by peer-to-peer messages as well as the field of interconnection methods between nodes in a network of nodes communicating with each other by peer-to-peer messages.

The network of nodes considered is for example an architecture of micro-services, respectively distributed over potentially several machines, for example not having the same addresses between them, for example not being all on the same network, not having between them for example not the same computing capacity. In addition, each micro-service can have a potentially indefinite life time. The heterogeneous nature of this micro-services architecture makes it more fragile compared to unforeseeable events, of the localized breakdown type, and therefore difficult to manage.

BACKGROUND OF THE INVENTION

The network of nodes, so that its nodes communicate with each other, goes first in an initial phase to interconnect the nodes with each other, then then goes in an updating phase during the life of the network to maintain and adapt the interconnections between network nodes.

For example, the message distribution system will make microservices communicate with each other through a message passing system. For this, it can use an implementation centered on a message broker ("broker" in English).

According to a first prior art, there is known a network of nodes exchanging peer to peer messages (“peer to peer” in English) and an associated interconnection method, which are both managed by a central interconnection device. . A disadvantage of this first prior art is that this central interconnection device constitutes a weak link in the network, a single point of failure (SPF for "single point of failure" in English), including a failure, even temporary, will disrupt the operation of the network and significantly degrade its performance in terms of transmission efficiency and message processing.

For example, the message broker-based system can be centralized (such as “RabbitMQ” or “ActiveMQ”), but it has an important limit, because this centralized system will constitute a single point of failure which can disrupt operation of the network and degrade its capacities.

According to a second prior art, there is known a network of nodes exchanging messages in peer to peer ("peer to peer" in English) and an associated interconnection method, which are both completely distributed. A drawback of this second prior art is that the robustness linked to this totally distributed nature, in particular in a network of nodes exchanging messages in peer to peer, becomes a weakness as soon as the size of the messages exchanged within the network becomes large. , because a large volume of fully distributed data exchange, in a type of peer-to-peer communication, will tend to saturate the network and therefore also to degrade its performance in transmission and message processing efficiency.

For example, the message broker-based system can be distributed (such as "Kafka"), but it has an important limitation, as this distributed system will become operationally limited once the size and volume of messages that it will have to pass will become too important.

SUMMARY OF THE INVENTION

The object of the present invention is to provide a network of nodes and / or an interconnection method at least partially overcoming the aforementioned drawbacks.

More particularly, the invention aims to provide a network of nodes and / or an interconnection method, for which a good compromise between operational robustness of the network on the one hand and operational efficiency of this network on the other hand is achieved .

Indeed, on the one hand the robustness of the network will allow the network of nodes to continue to operate by allowing the nodes of this network to continue to exchange messages with each other even if a given component, for example a central component, breaks down, at least in the case of a temporary breakdown, the most frequent case.

On the other hand, the operating efficiency of this network will allow the network node to continue to optimize the transmission and processing time of the messages exchanged, even if a given component, for example a central component, breaks down, at least in the case of a temporary breakdown, the most frequent case.

In order to achieve this compromise between robustness of operation and efficiency of operation of the network, the invention proposes a network in which the producer nodes, on the one hand each manage their list of consumer nodes themselves, thus guaranteeing their autonomy with respect to a possible point of failure in the network, and on the other hand each optimize their list of consumer nodes in terms of processing time of their produced messages, by removing, at least temporarily from their list, the consumer nodes too slow to process the product messages received from the producer node corresponding to this list.

To this end, the present invention provides a computer network of nodes communicating with each other by peer-to-peer messages, message producer nodes sending their product messages to message consumer nodes, a consumer node can also be a producer node. another level of the network, characterized in that each message-producing node: chooses, from among the nodes registered as active in the network, its consumer nodes to interconnect with them, maintains the list of its interconnected consumer nodes, manages the load distribution of its produced messages to its interconnected consumer nodes, according to the processing time, by its interconnected consumer nodes, of its produced messages, by sending its produced messages especially to its interconnected consumer nodes with lower processing time, its list of consumer nodes interconnected in function n the processing time, by its interconnected consumer nodes, of its produced messages, by removing from its list at least temporarily its too slow interconnected consumer nodes.

To this end, the present invention also provides a method of interconnection and communication between nodes in a computer network according to the invention.

To this end, the present invention also provides a method of interconnection between nodes in a computer network of nodes communicating with each other by peer-to-peer messages, message producing nodes sending their produced messages to message consuming nodes, a node consumer can also be a producer node at another level of the network, characterized in that each message producer node: chooses, among the nodes registered as active in the network, its consumer nodes to interconnect with them, keeps the list of its interconnected consumer nodes, modifies its list of interconnected consumer nodes according to the processing time, by its consumer nodes, of its produced messages, by deleting from its list at least temporarily its too slow interconnected consumer nodes.

According to preferred embodiments of the invention, the very voluminous nature of the data exchanged and the critical real-time nature of the processing of this exchanged data make the use of the network of nodes according to the invention as well as the associated interconnection method particularly interesting.

According to preferred embodiments of the invention, the network will use a peer-to-peer message passing system which will be self-adaptive and highly resilient. This system will make it possible to establish communication between microservices through a communication channel functionally identical to a data bus, but the implementation of which is based on direct communication between the microservices themselves. In this architecture, each micro-service is responsible for maintaining the list of its peers in the network, as well as for the load distribution during the distribution of messages to these peers. This system makes it possible to create data flows between micro-services which will be able to resist even a significant change of scale, while keeping a strong resilience in the event of the disappearance of nodes or in the event of communication links being cut in the network.

According to preferred embodiments, the invention comprises one or more of the following characteristics which can be used separately or in partial combination with one another or in total combination with one or other of the abovementioned objects of the invention .

Preferably, the nodes register as active in the network with a centralized register in the network, which register distributes to at least all the active producing nodes of the network, preferably to all the active nodes of the network, the list of 'at least all the active consumer nodes of the network, preferably the list of all the active nodes of the network. Broadcasting is a different form of communication than peer to peer message exchange.

Preferably, said register does not establish for any particular producer node given the list of its interconnected consumer nodes, and said register does not send to any particular producer node given the list of its interconnected consumer nodes.

An advantage is the fact that once formed, the network of nodes becomes autonomous with respect to the registry service without which this network can then function normally temporarily during a limited duration outage of this registry.

On the other hand, the proper functioning of the register at least part of the time remains useful, insofar as only the register can signal to the various producing nodes the changes which have occurred in the network topology, thus allowing these producing nodes to update each their list of consumer nodes, thereby allowing each producer node to operate with a list of consumer nodes which is optimized in terms of efficiency compared to the instantaneous topology of the network of nodes. Alternatively, managing the evolution of the network topology in a distributed manner at each producer node remains possible, but this would reduce the operating efficiency of the network.

Thus, the distribution at the level of each producer node of the management of the list of its consumer nodes on the one hand associated with the centralized management of the only network topology with its permanent update on the other hand achieves a significant additional improvement of the compromise between robustness of operation and efficiency of operation of the network.

Preferably, the nodes periodically register as active in the network with the registry, as soon as they are able to process messages, the registration period preferably being less than or equal to 10 seconds, any active node missing at least two registration periods, preferably at least one registration period, is removed from the list of active nodes by the registry. Thus, even if this periodic recording is a bit tedious for each consumer node, it does however make it possible to further improve the compromise between robustness of operation and operating efficiency of the network, insofar as it has the effect of temporarily removing any node deficient consumer, and to delete it only for about the duration of his deficiency, for the duration of a registration period.

Preferably, each message producing node can modify its list of interconnected consumer nodes by optionally adding a newly registered consumer node if the latter is able to consume the messages produced by said producing node. Thus, each producer node can increase its list of consumer nodes in real time, and therefore improve its load management in terms of product messages sent in real time as well.

Preferably, each consumer node sends, at least for certain messages received from a producer node, its processing time for these messages or else a weight representative of its processing time for these messages, representative of the difference between the arrival of these messages at said consumer node and their departure from said consumer node possibly after transformation, and preferably each producer node adds a request for evaluation of processing time in all or part of the messages that it sends to each of its interconnected consumer nodes . In this way, each consumer node can update the performance of all the consumer nodes in its list and adapt the load management of its produced messages accordingly, as often as it wants, it just has to choose the proportion of its messages produced to which it adds a request for evaluation of processing time.

Preferably, each message producing node manages the load distribution of its produced messages to its interconnected consumer nodes, according to the processing times or else the weights sent by its interconnected consumer nodes, sending more produced messages to the interconnected consumer nodes whose treatment times are shorter or the weights correspond to shorter treatment times. Thus each producing node improves the load management of its produced messages, thereby improving the overall efficiency of transmission of produced messages and their processing in the network.

Preferably, each producer node calculates a sliding average weight over time, for each of its interconnected consumer nodes, which it uses to manage the load distribution of its produced messages to its interconnected consumer nodes. In this way, the calculated parameter is on the one hand well representative of the speed of processing of the messages produced received by a consumer node, while smoothing the punctual variations not very representative of this parameter, reducing the risks of instability and divergence d servo depending on the variations of this parameter.

Preferably, each producer node loads its interconnected consumer nodes in a random weighted manner, so that the probability for its interconnected consumer nodes of receiving a message produced by said producer node, depends on a random draw which is weighted by the weights of said nodes interconnected consumers. This weighted random load management achieves a good compromise between, on the one hand, the fair and distributed use of all the available consumer nodes, a factor for improving the robustness of network operation by avoiding overloading high-performance consumer nodes to avoid their collapse, factor of instability in the network, and on the other hand the network's operating efficiency by using, as a priority and more often, the fastest consumer nodes.

Preferably, each node is a micro-service which exchanges messages with at least one other node, each micro-service being either a processor which processes data received from a network node to then send this processed data to another node of the network, preferably said processing being reduced to a single processing task performed on this data, either a source which produces data to be injected into the network to at least one processor, or a sink which exports the data processed by at least one processor to a system third party on the network. This type of network is particularly suitable for the use of the invention, insofar as its nodes perform a lot of relatively simple operations on a lot of data.

Preferably, at least one of the processors, preferably the majority of the processors, even more preferably all the processors, filters the data received and then directs them to different types of nodes in the network. This is a particularly representative example of carrying out simple operations in large numbers on a large volume of data.

Preferably, a type of producer node is a filter which separates the messages received according to a criterion to be fulfilled, to direct them to three types of consumer nodes respectively depending on whether said criterion is fulfilled, or not fulfilled, or perhaps filled. Again this is a simple operation, but to be performed very quickly and a large number of times.

In a first preferred embodiment, the messages are representative of bank transactions, the criterion is whether or not a bank transaction is fraudulent. The very voluminous nature of the data exchanged and the critical real-time nature of the processing of this exchanged data make the use of the network of nodes according to the invention as well as the associated interconnection method particularly advantageous.

In a second preferred embodiment, the messages are representative of images filmed by video surveillance cameras, the criterion is the presence or absence of crime risk related to the content of a filmed image or a sequence of images filmed by these CCTV cameras. The very voluminous nature of the data exchanged and the critical real-time nature of the processing of this exchanged data make the reuse of the network of nodes according to the invention as well as the associated interconnection method particularly advantageous.

Other characteristics and advantages of the invention will appear on reading the following description of a preferred embodiment of the invention, given by way of example and with reference to the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 schematically represents an example of a network part according to an embodiment of the invention.

FIG. 2 schematically represents an example of a network part according to an embodiment of the invention.

FIG. 3 schematically represents an example of a network part according to an embodiment of the invention.

FIG. 4 schematically represents an example of a cumulative weight scale allowing management of the peas of the consumer nodes in a network according to an embodiment of the invention.

DETAILED DESCRIPTION OF THE INVENTION

FIG. 1 schematically represents an example of a network part according to an embodiment of the invention.

Network 6 includes producer nodes 1 and consumer nodes 2. In network 6 are represented nodes A, B, C and D. Here, node A is a producer node 1. Node B is a consumer node 2 at level of the link between nodes A and B in network 6, while this same node B is a producer node 1 at the level of the link between nodes B and C in network 6 as well as at the level of the link between nodes B and D in the network 6. The node B is a producer node 1 or a consumer node 2, depending on the level of the network 6 considered. On the other hand here, the node A is only a producer node 1, while the nodes C and D are only consumer nodes 2.

A register 3 broadcasts to all the nodes A, B, C and D of the network 6, initially the original topology of the network 6, then each update of this topology during the life of the network 6. At the level of the link between nodes A and B in network 6, the messages produced by producer node A go to consumer node B and circulate only in this direction, to be consumed by consumer node B. At the level of the link between nodes B and C in network 6, the messages produced by the producer node B go to the consumer node C and circulate only in this direction, to be consumed by the consumer node C. At the level of the link between the nodes B and D in network 6, the messages produced by the producer node B go to the consumer node D and circulate only in this direction, to be consumed by the consumer node D.

Each producer node 1 maintains a list of consumer nodes 2 to which it may have to transmit messages as part of the flow to which it belongs. In the flow in Figure 1, node A knows node B to which it transmits the messages it produces. Node B knows nodes C and D to which it transmits the messages it produces, each message being sent either to node C or to node D.

This list of consumer nodes 2 is initialized by a request to a register 3 to which each producer node 1 or consumer node 2 is registered when it is instantiated. This register 3 is only useful when a change occurs in the topology of the nodes. Thus, if the register 3 disappears, for example if it is damaged or if its link with the network 6 becomes inoperative, the operation of the flow shown in FIG. 1 is not impacted.

Each producer 1 or consumer 2 node registers, at its instantiation, in register 3. It will then send a new registration request at regular intervals, for example every 5 or 10 seconds, to register 3. From in this way, if register 3 becomes faulty, this node 1 or 2 will quickly receive the complete topology of the network 6 of nodes during the re-instantiation of this register 3.

FIG. 2 schematically represents an example of a network part according to an embodiment of the invention. The part of the network 6 is first shown at a time T0 and then at a time Tl.

At time T0, node A is interconnected with node B and node B is interconnected with node C. Node A sends messages to node B which sends messages to node C.

Then register 3 broadcasts the information that node D has registered with it. Neither node A, nor node C, interconnect with node D. On the other hand, node B interconnects with node D. At time T1, node A is interconnected with node B and node B is interconnected with nodes C and D. Node A sends messages to node B which sends messages to either node C or node D.

The role of register 3 is to transmit all the events it processes in broadcast to all the nodes 1 and 2 of the network 6 which it knows. Thus, when a new node 1 or 2 is declared, all the other nodes 1 or 2 are informed thereof and the producing nodes 1 choose to connect to it according to whether this new node is part of their flow or not.

In FIG. 2, for example at time T1, the node D is declared, the register 3 diffuses the information. Node A does not connect to node D, node B connects to node D because node D is part of the type of service included in its flow.

Each consumer node 2 maintains the median response time over a sliding window of N messages sent, which is called Tmed. When sending a message to a consumer node 2, if this target consumer node 2 has not responded after 2 * Tmed, the request is abandoned. This request will be reissued immediately under the same conditions. After a certain number of such retransmissions, for example after 2 or 3 retransmissions, the sending of the message is considered to have failed and the target consumer node 2 is removed from the list of target consumer nodes 2 of producer node 1 having tried unsuccessfully to send him messages. This operating mode makes it possible to quickly eliminate the consumer nodes 2 whose response time becomes abnormally high. These deleted consumer nodes 2 will then be able to re-register with register 3, when they become operational again and when they can again guarantee a reasonable processing time for messages received from producer nodes 1. Register 3 will broadcast this information which will in turn be intercepted by the producing node 1 which is the sender of messages. This mechanism therefore makes it possible to temporarily exclude faulty or suddenly too slow consumer nodes 2, before reintegrating them into the flow, once they have become operational again.

In a network 6 of micro-service architecture type organized in flow and communicating with each other so that each message-producing micro-service manages its own list of message-consuming micro-services, the loss, at least temporarily, of register 3 has no impact on the operation of the network 6. The register 3 is used to distribute knowledge of the topology of the micro-services, but once this has been acquired, the network 6 becomes autonomous for its operation, and does not needs register 3 again only when there is a change in network topology 6.

FIG. 3 schematically represents an example of a network part according to an embodiment of the invention.

The network 6 includes producer nodes 1 and consumer nodes 2. In the network 6 are represented the nodes PI, P2, Cl, C2 and C3. Here, the nodes PI and P2 are producer nodes 1, at the level of links considered. Here the nodes Cl, C2 and C3, are consumer nodes 2, at the level of links considered.

A register 3 broadcasts to all the nodes PI, P2, Cl, C2 and C3 of the network 6, initially the original topology of the network 6, then each update of this topology during the life of the network 6.

The messages produced by the producer node PI go to the consumer nodes Cl and C2, and circulate only in this direction, to be consumed by the consumer nodes Cl and C2. The messages produced by the producer node P2 go to the consumer nodes Cl, C2 and C3, and circulate only in this direction, to be consumed by the consumer nodes Cl, C2 and C3.

On such an architecture having N producer nodes 1, i.e. sending messages, as well as M consumer nodes 2, i.e. receiving these messages, it is useful to intelligently distribute the workload between the different nodes. Load balancing is a technique used by large networks to balance the demand on servers. The network 6 according to the invention has implemented a dynamic, intelligent and adaptive load distribution.

Each producer node 1, for example a producer micro-service 1, has a load distribution module which is grafted onto its output interface and which communicates with the input interface of each of the consumer nodes 2 in its list. , for example consumer micro-services 2. Each of the N producer nodes 1 knows its list of consumer nodes 2 from among the M consumer nodes 2, all the M consumer nodes 2 having at initialization an identical weight, for example of value 1 .

Every q seconds, for example every 5 seconds, each producer node 1 will place, within some of its produced and sent messages, a header specifying a request for evaluation of the processing time for the message considered. Thus the consumer nodes 2 will have weights which are dynamic because these are recalculated regularly.

Each consumer node 2 has a queue of messages to be processed, as well as a processing function which consumes these messages one after the other.

As soon as a message integrating a request for an evaluation of processing time is inserted in the queue, it is tagged with its arrival system time, and as soon as it leaves the processing function, it will be tagged with its system time departure, the difference between this starting system time and this arriving system time corresponding to the processing time of the message by this consumer node 2.

As an acknowledgment message, the consumer node 2 responds to the producer node 1 that sent the processed message, preferably returning to it a parameter representative of the difference between its departure time and its arrival time, for example a weight, or alternatively, the processing time itself.

Here, from this difference, a weight is generated proportional to the speed of execution of the consumer node 2.

Let the weight of a consumer node i:

Wj = ------------- ti.arrivee .departure which corresponds to formula (1).

This weight is updated in the table of consumer nodes of the producer node concerned. Each producer node independently collects the load information from each consumer node. Each producer node therefore maintains its own weight table of consumer nodes which can therefore be different from that of the other producer nodes.

In order to obtain a precise weight, the invention preferably uses the concept of sliding average weight, defined by the average of the weights over a time window δ, defining a number of iterations making it possible to calculate a sliding average, given for a node consumer in particular:

..... W _in } which corresponds to formula (2).

Let the sliding average weight of consumer i, noted Wigliss, for a set of measures of size n:

Wjgliss =

and Wjgliss = 1 for n = 0 which corresponds to formula (3).

This sliding weight is used to weight the random draw which is used to choose the consumer node receiving the next message.

The weighting is done by comparison of the value drawn with a scale on which the consumer nodes are positioned according to their weight.

This scale goes from 0 to the value of the sum of the Wigliss. The value of each terminal on the scale is calculated based on the cumulative values of the previous terminals.

With N the number of consumer nodes:

For i from 0 to N - 1: Terminal; = Terminal;.; + Wjgliss and Terminal ₀ = 0

For example, if we take the network part of Figure 3, we get the following calculation.

After initialization of the network, the producer nodes PI and P2 both have the same weight table:

TABLE 1

PI P2 Cl 1 1 C2 1 1 C3 1 1

During a first iteration, the weights are accumulated on a scale, and we obtain table 2.

TABLE 2

Cl C2 C3 PI 0 1 2 3 P2 0 1 2 3

For the sending of their next message, the producer nodes PI and P2 take a value between 0 and 3. If this value is between 0 and 1, the next message will be sent to the consumer node Cl, if the value is between 1 and 2 , the message will be sent to the consumer node C2, if the value is between 2 and 3, the message will be sent to the consumer node C3. By convention, if the value is exactly 1, it will be sent to the consumer node C1, and if the value is exactly 2, it will be sent to the consumer node C2.

Consider for example, as draw value 1.5 for the producer node PI and 2.2 for the producer node P2. The producer node PI therefore sends its message to the consumer node C2, while the producer node P2 sends its message to the consumer node C3.

The consumer node C2 responds in 5 seconds to the producer node PI and the consumer node C3 responds in 2 seconds to the producer node P2.

By applying formulas (1), (2) and (3) presented above, we obtain for the producer node PI:

1

W _c2 o = 7 ------—----- = 7 = 0.2

H. arrival H. departure

ÔC2 ^: {0.2}

W _c2 gliss = - * W _c20 = 0.2 and for the producer node P2:

1 w _c3 0 = 7 ------—---- = _? = 0.5

H. arrival H. departure ô _c3 : {0.5}

W _c3 gliss = - * W _c30 = 0.5

The producer nodes PI and P2 now have different weights for each of the consumer nodes Cl, C2 and C3, which gives the table

3.

TABLE 3

PI P2 Cl 1 1 C2 0.2 1 C3 1 0.5

In a second iteration, the weights are accumulated on a scale which is updated, and we obtain table 4.

TABLE 4

Cl C2 C3 PI 0 1 1.2 2.2 P2 0 1 2 2.5

The new random draw gives a value of 1.1 for the producer node PI and a value 0.3 for the producer node P2.

The producer node PI therefore sends its message to the consumer node C2 and the producer node P2 sends its message to the consumer node C1.

The consumer node C2 responds to the producer node PI in 2 seconds and the consumer node C1 responds to the producer node P2 in 10 seconds.

- = ° .s

By applying again the formulas (1), (2) and (3) for the producer node PI, we obtain: W _{c2 X} = ------- ί ---- ^L i.arrivee ^L i.depart ô _c2 : (0.2,0.5}

W _c2 gliss = | * (W _c2 o + W _c2 J = 0.35 and for the producer node P2, we obtain:

w _cl0 = --------- H. arrival H.departure - = ° -l

Ô _C1 : {0.1}

W _cl gliss = - * W _cl0 = 0.1

The producer nodes PI and P2 now have weights modified again for the consumer nodes Cl, C2 and C3, which gives the table

5.

TABLE 5

PI P2 Cl 1 0.1 C2 0.35 1 C3 1 0.5

In a third iteration, the weights are accumulated on a scale which is updated, and we obtain table 6.

TABLE 6

Cl C2 C3 PI 0 1 1.35 2.35 P2 0 0.1 1.1 1.6

Thus, the final weights for the consumer nodes C1 to C3 are dynamic sliding average weights.

FIG. 4 schematically represents an example of a cumulative weight scale allowing management of the peas of the consumer nodes in a network according to an embodiment of the invention.

The cumulative weight scale represents the three consumer nodes C1, C2 and C3 at time tO in a tree, each consumer node being represented by a sheet 4 comprising the name of the node and the value of its weight.

The cumulative weight scale represents the three consumer nodes C1, C2 and C3 at time tl in the same tree, each consumer node being represented by a sheet 5 comprising the name of the node and the value of its weight which has possibly been modified.

The cumulative weight scale is stored in the form of a cumulative weight table via a range tree, which brings average complexity with this technique in O (log (n)).

The leaves of the tree correspond to the boundaries of the intervals. The tree structure makes it possible to access the values in a given interval quickly and to recalculate the weights involves known algorithms.

One sees, for the nodes Cl, C2, C3, the limits of the intervals which are respectively 1/2/3, at time tO, and 1 / 1.2 / 2.2 at time tl.

This tree data structure storing this cumulative weight scale not only makes it possible to respond functionally to the problem of storing a scale of values, but also it has an additional useful advantage which is to present very good access performance. and modification to the values stored there.

Of course, the present invention is not limited to the examples and to the embodiment described and shown, but it is susceptible of numerous variants accessible to those skilled in the art.

Claims (16)

1. Computer network of nodes (1, 2) communicating with each other by peer-to-peer messages, message producing nodes (1) sending their produced messages to message consumer nodes (2), a consumer node (2) being able also be a producer node (1) at another level of the network (6), characterized in that each message producer node (1): > chooses, among the nodes (1, 2) registered as active in the network, its consumer nodes (2) to interconnect with them,> keeps the list of its interconnected consumer nodes (2) up to date,> manages the distribution load of its produced messages to its interconnected consumer nodes (2), depending on the processing time, by its interconnected consumer nodes (2), of its produced messages, by sending its produced messages mainly to its interconnected consumer nodes (2) with lower processing time,> modifies its list of interconnected consumer nodes (2) according to the processing time, by its interconnected consumer nodes (2), of its produced messages, by deleting its consumer nodes from its list at least temporarily (2) interconnected too slow. 2. Computer network according to claim 1, characterized in that the nodes (1,2) register as active in the network with a register (3) centralized in the network (6), which register broadcasts at least all active producer nodes (1) on the network (6), preferably all active network nodes (1, 2) (6), the list of at least all active consumer nodes (2) on the network (6) ), preferably the list of all active nodes (1,2) in the network (6). 3. Computer network according to claim 2, characterized in that said register (3) does not establish for any particular producer node (1) given the list of its interconnected consumer nodes (2), does not send to any producer node (1 ) particular given the list of its interconnected consumer nodes (2). 4. Computer network according to any one of claims 2 to 3, characterized in that the nodes (1, 2) periodically register as active in the network (6) with the register (3), since they are capable of processing messages, the recording period preferably being less than or equal to 10 seconds, any active node (1, 2) which misses at least two recording periods, preferably at least one recording period, is removed from the list of active nodes (1, 2) by the register (3). 5. Computer network according to any one of the preceding claims, characterized in that each message producing node (1) can modify its list of interconnected consumer nodes (2) by optionally adding a newly registered consumer node (2) if that -this is able to consume the messages produced by said producer node (1). 6. Computer network according to any one of the preceding claims, characterized in that each consumer node (2) sends, at least for certain messages received from a producer node (1), its processing time for these messages or else a weight representative of its processing time for these messages, representative of the difference between the arrival of these messages at said consumer node (2) and their departure from said consumer node (2) possibly after transformation, preferably each producing node (1 ) adds a request for evaluation of processing time in all or part of the messages which it sends to each of its interconnected consumer nodes (2). 7. Computer network according to claim 6, characterized in that each message producing node (1) manages the load distribution of its produced messages to its interconnected consumer nodes (2), as a function of the processing times or of the weights sent by its interconnected consumer nodes (2), sending more messages produced to the interconnected consumer nodes (2) whose processing times are shorter or the weights correspond to shorter processing times. 8. Computer network according to claim 7, characterized in that each producer node (1) calculates a sliding average weight over time, for each of its interconnected consumer nodes (2), which it uses to manage the load distribution of its messages produced to its interconnected consumer nodes (2). 9. Computer network according to claim 7 or 8, characterized in that each producer node (1) loads its interconnected consumer nodes (2) in a random weighted manner, so that the probability for its interconnected consumer nodes (2) of receiving a message produced by said producer node (1), depends on a random draw which is weighted by the weights of said interconnected consumer nodes (2). 10. Computer network according to any one of the preceding claims, characterized in that: > each node (1, 2) is a micro-service which exchanges messages with at least one other node (1,2),> each micro-service being: o either a processor which processes data received from a node (1,2) of the network (6) to then send this processed data to another node (1, 2) of the network (6), preferably said processing being reduced to a single processing task performed on these data, o either a source which produces data to be injected into the network (6) to at least one processor, o either a sink which exports the data processed by at least one processor to a system third party to the network (6). 11. Computer network according to claim 10, characterized in that at least one of the processors, preferably the majority of the processors, even more preferably all the processors, filters the data received and then directs them to different types of nodes (1,2) in the network (6). 12. Computer network according to any one of the preceding claims, characterized in that a type of producer node (1) is a filter which separates the messages received according to a criterion to be fulfilled, to direct them to three types of nodes consumers (2) respectively according to whether said criterion is: fulfilled, or not fulfilled, or perhaps fulfilled. 13. Computer network according to claim 12, characterized in that the messages are representative of banking transactions, the criterion is the fraudulent nature or not of a banking transaction. 14. Computer network according to claim 12, characterized in that the messages are representative of images filmed by video surveillance cameras, the criterion is the presence or absence of crime risk related to the content of a filmed image or of a sequence of images filmed by these video surveillance cameras. 15. A method of interconnection and communication between nodes (1, 2) in a computer network (6) according to any one of the preceding claims. 16. Method of interconnection between nodes (1, 2) in a computer network (6) of nodes (1,2) communicating with each other by peer-to-peer messages, message producing nodes (1) sending their produced messages to message consumer nodes (2), a consumer node (2) possibly also being a producer node (1) at another level of the network (6), characterized in that each message producer node 5 (1): > chooses, among the nodes (1, 2) registered as active in the network (6), its consumer nodes (2) to interconnect with them,> keeps the list of its interconnected consumer nodes (2) up to date, 10> modifies its list of interconnected consumer nodes (2) as a function of the processing time, by its consumer nodes (2), of its produced messages, by deleting from its list at least temporarily its interconnected consumer nodes (2) which are too slow. 6 < 1/2 f 6 < 1 1.2 2 ÏHD 1.2 / TO REGISTER 7 ------- k 2/2 REGISTER