JPH0391337A - Network control system - Google Patents

Abstract

PURPOSE:To make the network communication processing efficient and to allow the system to cope flexibly with other communication pattern by providing a switch box or the like having a distributer and a concentrator to eliminate the need for the control by a centralizing controller. CONSTITUTION:Each node 50 is provided with plural switch boxes 51 (one box only is shown in the figure) and a hand-shake section 52 synchronizing message transmission reception. Moreover, an input section of each box 51 is provided with distributers 53-1-53-3, and an output section is provided with concentrators 54-1-54-3. Moreover, each distributer 53 and each concentrator 54 are provided respectively with a message passing detection section and in the case of message passing, a message end tag at the tail end of a message detects the message passing. Then the connection state of a 2-branch distribution switch in each distributer 53 or a 2-branch merger in each concentrator 54 is switched. Thus, the control by the centralized type controller is not required, the communication processing of the network is made efficient and the system copes flexibly with other communication pattern.

Description

[Detailed description of the invention] (Table of contents) Overview Industrial field of application Prior art (Fig. 10) Means for solving the problems to be solved by the invention (Fig. 1) Working examples (Fig. 2) - Figure 9) Effects of the invention [Summary] Regarding the network control method, the present invention enables efficient execution of network communication processing without requiring control by a centralized controller, and also enables storage-and-forward type control methods, etc. , with the aim of being able to respond flexibly to other communication patterns, 26 nodes are interconnected by links of each dMi such that each node forms each vertex of a d-dimensional hypercube, In a network that constitutes a communication path between each node and devices connected by human output links, each node is provided with a plurality of switch boxes, and these switch boxes are interconnected via a handshake unit. The box has a distributor and aggregator,
They are provided at the input and output sections of the message, respectively, and when the message passes through each node, a connection pattern is automatically set, and the set connection pattern is configured to enable all-point to all-point communication.

[Industrial application field]

The present invention relates to a network control method, and more specifically, the present invention is used to control a hypercube-shaped network in which a plurality of nodes are connected by links such that each node is a vertex of the hypercube, and in particular, it uses a centralized controller. The present invention relates to a network control method that can efficiently execute communication processing without any problems and can flexibly respond to various communication patterns.

[Conventional technology]

Conventionally, 2'' nodes are connected to each other by d sets of links so that each node forms each vertex of a d-dimensional hypercube, and each node is connected to each other by a predetermined input/output link. A network (hypercubic network) in which communication channels are configured is known (for example, see Japanese Patent Application Laid-Open No. 149350).

The conventional network control method described above is as shown in FIG.

In the figure, l is a node, 2 is a global controller,
3 is a switch circuit, 4 is a buffer, 5 is a register, 6 is a control section, 7 is a connection pattern setting section, 8 is an inverter, and 9 is a bit string generator.

Each node 1 (there are many nodes with the same configuration) has a link 10 connecting it to a predetermined node, and is also connected to a processor by an input link 12 and an output link 11.

The global controller 2 uses the bit string generator 9 to generate different bit examples at fixed intervals set as the time required to transmit one or two messages, and sends them as a phase signal to all nodes with even node names. .

At the same time, each bit of the same bit string is inverted by the inverter 8 and the bit string is sent to all nodes with odd node names.

The signal from the global controller 2 is received by the connection pattern setting section 7 at each node, and is passed to the control section 6 as a connection pattern.

The control unit 6 decodes the received connection pattern and sends a signal to the switch circuit 3 to switch the switch circuit. This establishes connections between links and between links and input/output links.

[Problem to be solved by the invention]

The above-mentioned conventional devices had the following drawbacks.

(11) The conventional method described above assumes the existence of a centralized controller such as a global controller.

This global controller generates a phase signal at the beginning of each phase and distributes it to each node. Each node knows the start of each phase from this phase signal and processes each phase in synchronization as a whole, making it possible to efficiently execute extremely heavy communication processing such as all-point to all-point communication.

However, as the scale of the network increases, the length of the signal line that distributes the phase signal from the centralized controller to each node becomes longer, and the variation in length also increases.

As a result, variations occur in the times when phase signals arrive at each node, and in order to absorb this variation, it is necessary to allow for a margin at the start and end times of each phase, which degrades network performance. let

(2) Although the conventional method described above efficiently processes all-point to all-point communication, it is difficult to efficiently process other communication patterns.

In order to flexibly respond to various communication patterns, it is better to adopt a store-and-forward type network control method, but with the conventional method in which switching networks are controlled by a centralized controller, it is not recommended to incorporate the above control method. Yes.

The present invention eliminates these conventional drawbacks, makes it possible to efficiently execute network communication processing without the need for control by a centralized controller, and makes it possible to efficiently execute network communication processing without requiring control by a centralized controller. The aim is to be able to respond flexibly to various communication patterns.

[Means to solve the problem]

FIG. 1 is a diagram showing the principle of the present invention, where FIG. A is a block diagram of a node, FIG. B is a block diagram of a distributor, and FIG. 0 is a block diagram of an aggregator.

In the figure, 50 is a node, 51 is a switch box,
52 is a handshake section, 53 (531 to 53-3)
is a distributor, 54 (54-1 to 543) is a concentrator, 55 is a message passage detection unit, 56 is a two-way distribution switch, 57
indicates a forked Marsha.

The hypercubic network targeted by the present invention is similar to the conventional example described above, that is, 2d (d is an integer of 1 or more).
are connected to each other by d sets of links connecting each node so that each node forms each vertex of a d-dimensional hypercube, and connected to each node by a predetermined input/output link. This is a network configured to constitute a communication path.

In the network control method of the present invention, the switch box control signal, which was conventionally generated by the phase signal received from the global controller, is now generated by detecting that a message has passed through the switch box. This is what I did.

This eliminates the need for the conventionally used global controller.

Each node 50 includes a plurality of switch boxes 51 (1
(Although only one switch box is shown in the figure, in reality, a plurality of switch boxes are provided) and a handshake section 52 for synchronizing message reception and passing.

Furthermore, in each switch box 51, distributors 53-1 to 53-3 are provided at the input section, and aggregators 54-1 to 54-3 are provided at the output section.
54-3 is provided.

Each distributor 53 is provided with a message passage detection section 55 and a bifurcated distribution switch 56, and each concentrator 54 is provided with a message passage detection section 55 and a bifurcated marshaller 57.

In the network control method of the present invention, a message end tag is provided at the rear end of the message to notify the message passage detection section 55 of the passage of the message, and furthermore, a message in a certain phase is configured to have a message end tag at the end of the message to notify the passage of the message to the message passage detection unit 55. Therefore, it is necessary to perform a handshake between the nodes to receive and pass the message. Therefore, a handshake section 52 is provided between each switch box.

In each distributor 53 or aggregator 54, a message passage detection section 55 detects the passage of messages, and when a preset number of messages is detected, a two-way distribution switch 55
6. Or switch the connection status of the two-pronged Marsha.

Now, the distributors 53-1 to 53- in the switch box 51
3, and the aggregators 54-1 to 54-3 are connected as shown by the solid line in the figure, the message input to the distributor 53-1 is output to the aggregator 54-1, and is input to the distributor 53-2. The received message is output to the aggregator 54-3, and the message input to the distributor 53-3 is output to the aggregator 54-2.

Moreover, if the connection is made as shown by the dotted line in the figure, the distributor 53-
Messages input to the distributor 53-2 are output to the aggregator 542, messages input manually to the distributor 53-1 are output to the aggregator 54-3, and messages input to the distributor 53-3 are output to the aggregator 541. do.

By connecting and combining a plurality of switch boxes having the above-mentioned switching patterns within each node, efficient control of a hypercubic network can be realized.

[Effect]

As described above, the present invention detects the passage of messages in the distributor and concentrator in each switch box, and when it is detected that a preset number of messages have passed, switches the two-way distribution switch or the two-way marshaller. .

This changes the connection pattern within the switch box and changes the message passage route.

In this way, all-point to all-point communication can be processed efficiently without the need for a centralized controller like in the conventional example, and other communication patterns can also be flexibly handled.

〔Example〕

Embodiments of the present invention will be described below based on the drawings.

2 to 8 are diagrams showing one embodiment of the present invention, in which FIG. 2 is a block diagram of a node, FIG. 3 is a block diagram of a shake hand section, and FIG. 4 is a block diagram of a switch box. Figure 5 is a block diagram of the distributor, Figure 6 is a block diagram of the aggregator, Figure 7 is an overall diagram for 4 nodes (two-dimensional), and Figure 8 is an explanatory diagram of the switching state of the switch box. It is.

In the figure, 50 is a node, 51 is a switch box,
52 is a handshake unit, 53 is a distributor, 54 is a concentrator, 56 is a two-way distribution switch, 57 is a two-way marshaller, 58
is a manual link, 59 is an output link, 60 is a link, 61 is a buffer, 62 is a receiving section,
63 is a transmitter, 64 is a cycle register, 65 is a FIFO memory, 66 is a message end determination unit, 67 is a counter,
68 indicates a comparator.

For example, as shown in FIG. 2, each node 50 has links 60-1 to 60-4 connecting to a predetermined node, and is also connected to a processor through a human power link 58 and an output link 59, and forms a hypercube as a whole. Bottom the network of molds.

For example, four switch boxes 51-1 to 51-4 are provided inside the node 50, and are connected to each other via a handshake unit 52.

The handshake unit 52, as shown in FIG.
, buffer 61, and transmitter 63.

The illustrated solid arrows indicate the flow of data, and the dotted arrows indicate the flow of ready signals (synchronization signals).

The receiving unit 62 receives and receives data and stores it in the buffer 61.
write to. When the transmitter 63 receives the ready signal,
If there is data in the buffer 61, it is read out and output.

The buffer 61 from which the data has been read outputs a ready signal to indicate that it is now ready to receive new data.

When the receiving section 62 receives and receives data and the buffer 61 is filled, the buffer 61 stops outputting the ready signal.

In the initial state, the buffer 61 is empty and outputs a ready signal.

For example, as shown in FIG. 4, the switch box 51 has three distributors 53-1 to 53-3 and three concentrators 54-1.
54-3 and the period register 64.

The period register 64 has a value of 2 to the (K-1) power, where k (1≦k<the number of dimensions of the hypercube) is the dimension that the switch box is responsible for.

Each distributor and aggregator proceeds with processing while referring to this value.

Therefore, the period register 64 is connected to each distributor 5 l 3-1 to 53-3 and each aggregator 54-1 to 543 so as to send a predetermined value.

The distributor 53, as shown in FIG.
6, a counter 67, a comparator 68, etc.

The message (data) input to the distributor 53 flows in the direction of the solid arrow shown in the figure, first passes through the FIFO memory 65, is input to the two-way distribution switch 56, and is distributed.

The connection state of the two-way distribution switch 56 is such that the value of the counter 67 and the value sent from the external period register 64 are
A comparator 68 compares the values, and when the two values match, switching is performed. Also, at the same time as this switching, the counter 6
7 is reset.

The counter 67 is incremented each time a message ends by a message end determination section 66 that monitors the message passing through the FIFO memory 65 and detects when the message ends.

5 To detect the passage of a message, the message end tag provided in the message is detected by the message end determination unit 6.
6 is detected.

The switching of the two-way distribution switch 56 is determined by the number of times a message passes through the FIF◯ memory 65, and a value corresponding to the number of times is given to the comparator 68 from the cycle register.

Therefore, the number of FIFs given from the period register is
Each time the message passes through the O memory 65, the bifurcated distribution switch 56 is alternately switched.

Note that the synchronization signal (for example, the ready signal) flows in the direction of the dotted arrow shown in the figure (in the opposite direction to the data), and predetermined switching is performed.

The concentrator 54 is shown in FIG. 6, and its function is similar to that of a distributor.

The difference is that a distributor uses a two-way distribution switch 56 to distribute messages (data) and ready signals, whereas a concentrator uses a two-way marshaller 57 to distribute messages (data) and ready signals.
These signals are aggregated.

The control of this bifurcated marshaller 57 is the same as that of the distributor, and when the number of messages passing through the FIFO memory 65 matches the value from the period register, the bifurcated marshrer
7.

FIG. 7 is an overall diagram of the network in the case of four nodes (two-dimensional).

Four nodes, node 0 to node 3, are each configured with two switch boxes 51-1 and 51-2. Connections between the switch boxes 51-1 and 51-2 and each node 50 are made through a handshake unit 52 as shown in the figure.

FIG. 8 shows the switching states of the switch boxes in the above network.

(A) The figure shows the initial state of the switch box 51, and the connection pattern of the switch box 51 is the connection pattern set before data passing (the pattern set in another circuit).
It becomes.

That is, the distributors 53-l to 53-3 and the concentrators 54-1 to 54-5
4-3 has been switched to the position indicated by the bold line in the diagram,
Connections are made between the distributor 53-1 and the concentrator 54-1, between the distributor 53-2 and the concentrator 54-3, and between the distributor 53-3 and the concentrator 54-2.

Assume that in this connected state, messages "E-phase O" and "E-phase 3" are sent.

After a predetermined period of time has elapsed from this initial state, the state shown in (B) is reached.

In state (B), "-Z0" in "E phase 0"
The portion marked "-Z3" of "E phase 3" passes through the switch box 51.

After that, when the state (C) is reached, "E phase O" and "E
All messages for "Phase 3" are sent to switch box 51.
The subsequently sent messages "E-phase 1" and "E-phase 2" are manually input to the shake hand section 52.

At this time, for example, if the connection state in the switch 51 is changed by passing the message once, the distributor 53-1.53-2 and the concentrator 541.54-2.54
-3 is switched, but the distributor 53-3 remains in its original connection state because no message passes through it.

Next, in state (D), since the distributor 53-1 and the concentrator 54-3 are connected, the "-z1" part of the message "E phase 1" passes, but the distribution Vessel 5
3-2 and the concentrator 54-2 are not connected, so they are in a waiting state.

After that, "E phase 1" is divided into the distributor 53-1 and the concentrator 54.
-3, and "Efe" passes through the distributor 53-3.
When the signal passes between the concentrator 54-2 and the concentrator 54-2, the connection state is switched to state (E).

In state (E), the state between the distributor 53-1 and the aggregator 54-1 is in a waiting state.

After that, in the state (F), the distributor 53-2 and the concentrator 5
4-2 is connected, the message "E phase 2" is in a passing state.

Subsequently, if a message passes between the distributor 53-3 and the concentrator 54-1, the above waiting state is resolved.

In this way, it is possible to switch the connection state of the switch box by passing a message.

Note that in the above example, the connection state of the switch is changed when the message passes once, but the connection state of the switch can be changed when the message passes N times.

FIG. 9 is an explanatory diagram when combined with a conventional (store-and-forward) network control method.

In the figure, 100 is an all-points-to-all-points processing unit (
101 is a conventional store and forward (S&F) processing unit (network), 102 is a determination unit, and 103 is a queue.

The message shall have a tag at its head indicating whether it is a message for all-point to all-point communication or a message for S&F communication.

The determination unit 102 looks at the tags of the messages and either passes each message to the all-points-to-all-points processing unit 100 or sends it to S&F.
It is determined whether to pass it to the processing unit 101.

The output from each processing unit is sent to a queue 1030. merged and output externally.

In this way, the network control method of the present invention,
In combination with the S&F type network control system, all point-to-all point communications can be processed efficiently, and conventional point-to-point communication processing can also be executed flexibly.

〔Effect of the invention〕

As explained above, the present invention has the following effects.

(11) Since connection patterns can be set without using a conventional centralized controller (global controller), communication processing can be executed efficiently even when the scale of the network increases.

(2) All point-to-all communication can be efficiently processed without a centralized controller, so when this method is combined with, for example, a store-and-forward type network control method, all-point to all Point communication can be processed efficiently, and point-to-point communication, etc., which was conventionally used in the first embodiment, can also be processed flexibly.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the principle of a network control system according to the present invention, FIGS. 2 to 8 are diagrams showing one embodiment of the present invention, FIG. 2 is a block diagram of a node, and FIG. 3 is a shake diagram. A block diagram of the hand section, Fig. 4 is a block diagram of the switch box, Fig. 5 is a block diagram of the distributor, Fig. 6 is a block diagram of the concentrator, Fig. 7 is an overall diagram for the case of 4 nodes, Fig. 8 FIG. 9 is an explanatory diagram of the switching state of the switch box, FIG. 9 is an explanatory diagram when combined with a conventional network control method, and FIG. 10 is an explanatory diagram of the conventional network control method. 2 0 - Node 51 - Switch box 2 - Handshake unit 3 - Distributor 54 - Aggregator 5 - Message passage detection unit 6 - Bifurcated distribution switch 7 - Bifurcated marsha

Claims (1)

[Claims] 2^d nodes (d is an integer of 1 or more) are connected to each other by d sets of links connecting each node so that each node forms each vertex of a d-dimensional hypercube. In a network that configures a communication path between devices that are connected to each other and connected to each node via a predetermined input/output link, each node (50) has a link that connects to other nodes, and a link that connects the input/output links to each other. A plurality of switch boxes (51) are provided to switch the connection state of the switch boxes (51), and these switch boxes are connected to each other via a handshake unit (52) that synchronizes message exchange. A distributor (53) and an aggregator (54) that detect that a predetermined number of nodes have passed and switch the internal connection state are provided at the message input and output sections, respectively, so that the message passes through each of the nodes. A network control method characterized by automatically setting a connection pattern and enabling all-point to all-point communication according to the set connection pattern.