JPH09231188A - Multicluster information processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To eliminate the overhead at data transfer time and improve the performance of the information processing system, constituted by connecting plural clusters including plural arithmetic processors by a crossbar network, by a broadcasting function which enables an arithmetic processor in one cluster to transfer data to all other clusters with a single instruction. SOLUTION: If one of arithmetic processors 21-2N issues a broadcast instruction, it is received by a data transfer device 5 through a local network 4. When an instruction decoding means 51 knows that the received instruction is the broadcast instruction by deciphering, a common memory access means 53 reads transfer data out of a common memory 3 through the local network 4 in response to the indication of a transfer control means 52 and transfers it to a data transfer means 54. The data transfer means 54 holds the data in a data holding means 541 and transfers the data to all other clusters in order.

Description

Detailed Description of the Invention

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information processing system including a plurality of clusters, and more particularly to an information processing system including means for controlling data transfer between clusters.

2. Description of the Related Art An example of a conventional technology of an information processing system including a plurality of clusters is described in Kai Hwang, "Advanced Computer Archit" by Kai Wang, 1993, published by McGraw-Hill.
ecture ", 1993, McGraw-Hill Inc, New York) No. 336
It is disclosed in FIG. 7.4 on page (hereinafter, Prior Art 1). In this conventional technique, one cluster includes a plurality of processors P, one main memory MM, and a cluster bus connecting them. One information processing system is
Contains multiple clusters. The cluster buses of each cluster are connected by a global bus.

In such an information processing system, 1
Data transfer from one cluster to other clusters is achieved by the following steps. In this example, the information processing system includes cluster # 0 to cluster # (N-
1). Cluster # 0 is main memory MM #
0, and processors # 0-0 to 0-n. At the request of the processors # 0-0, the data D in the main memory MM # 0 is changed to cluster # 1 to cluster #
(N-1).

In the first step, processor # 0
-0 reads the data D from the main memory MM # 0. The time required for the first step is Ta.

In the second step, processor # 0
-0 is via cluster bus and global bus
Send data D to cluster # 1. Processor # 0-
0 waits for a reception end notification from the cluster # 1. The time required for the second step is Tb.

Referring to FIG. 1, by repeating the first and second steps, cluster # 2 to cluster #
The data D is transferred to (N-1). The total time T required for data transfer is T = (Ta + Tb) × (N−1)
It is.

In addition to the above-mentioned conventional technique, the following technique is known.

"Iepleeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee (IEE) 1696 92-Ieeeeeeeeeee Lee e
Standard for Scalable Coherent
Interface (IEEE Standard fo
r Scalable Coherent Inter
face, "pages 23, 24, and 71, disclose a broadcast technique in an information processing system having a ringlet type transmission path (hereinafter," prior art 2 "). Even with this technology, N
The time required to transfer data to one cluster is (Ta + Tb) * (N-1).

Japanese Unexamined Patent Publication No. 4-287153 discloses broadcasting separately from the inter-cluster network.
A technique for providing a type communication bus is described (hereinafter, "prior art 3").

In the above-mentioned prior art,
There were the following problems.

In the prior art 1, since the transfer data is repeatedly read from the main memory, it takes a long time to transfer the data.

In the prior art 2, since the other cluster is present in the data transfer path to the transfer destination cluster, the transfer speed is slow.

In the prior art 3, since a broadcast communication bus is separately provided, the amount of hardware increases.

Further, in the system in which the processor directly receives data, there is a problem that the processing of the processor is interrupted each time data is transferred.

In view of these problems, one of the objects of the present invention is
One is to broadcast the data to multiple clusters in a short time.

Another object of the present invention is to reduce the number of times of reading data from the main memory due to broadcasting.

Another object of the present invention is to transfer data to a plurality of clusters simultaneously.

Another object of the present invention is to transfer data to a transfer destination cluster without passing through another cluster.

Another object of the present invention is to achieve the above object without increasing the amount of hardware.

In view of the above problems, a multi-cluster information processing system of the present invention is a multi-cluster information processing system including a plurality of clusters and an inter-cluster connecting means for connecting the plurality of clusters.
Each of the plurality of clusters includes a plurality of processors, a shared memory accessed from the plurality of processors,
Intra-cluster connecting means for connecting the plurality of processors and the shared memory, data holding means for holding transfer data to be transferred to another cluster, and at least one of the plurality of processors to a plurality of transfer destination clusters Transfer data reading means for reading the transfer data from the shared memory and storing it in the data holding means when requesting the data transfer of the data, and the data stored in the data holding means by the plurality of data via the crossbar network. Data transfer means for sequentially transferring to each of the transfer destination clusters.

In another embodiment, the multi-cluster information processing system of the present invention is a multi-cluster information processing system including a plurality of clusters and inter-cluster connecting means for connecting the plurality of clusters. Data holding means for storing transfer data received from one of the plurality of clusters, and data transfer means for simultaneously transmitting the transfer data to a plurality of transfer destination clusters.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION Next, a first embodiment of the present invention will be described with reference to the drawings.

Referring to FIG. 2, the information processing system according to this embodiment includes clusters # 1 to # 3 and an inter-cluster crossbar network 60 connecting these clusters. The cluster # 1 has N processing units 2
1 to 2N, a shared memory 3 shared by all of these arithmetic processing devices 21 to 2N, and a data transfer control device 5 for controlling data transfer between clusters. The arithmetic processing devices 21 to 2N, the shared memory 3, and the data transfer control device 5 are connected by a local network 4.

Referring to FIG. 2, the data transfer control device 5
Includes instruction decoding means 51, transfer control means 52, shared memory access means 53, and data transfer control means 54. The shared memory access unit 53 executes reading and writing to the shared memory 3 via the local network 4. The instruction decoding means 51 decodes the instructions issued by the arithmetic processing units 21 to 2N. When this command is a broadcast command, the command decoding means 51
It instructs the transfer control means 52 to execute broadcast transfer. The data transfer control means 54 executes data transfer to another cluster. Data transfer control means 54
Includes a data holding unit 541. Data holding means 54
1 temporarily holds the data read from the shared memory 3. Since the data holding unit 541 temporarily holds the data, the reading of the data from the shared memory 3 is
It only needs to be done once.

The transfer control means 52 includes an instruction decoding means 51,
It controls the shared memory access means 53 and the data transfer control means 54.

Referring to FIG. 3, the data transfer control means 5
4 is a data holding unit 541 and a parameter holding unit 5
42, a control means 543, a data transfer means 544,
A broadcast end monitoring unit 545, a shared memory access end monitoring unit 546, an instruction decoding unit 547,
And data storage means 548. The shared memory access end monitoring means 546 detects whether or not the access to the shared memory 3 has ended. The parameter holding unit 542
Stores parameters required for broadcasting. The data transfer means 544 is the data holding means 541.
And the information stored in the parameter holding unit 542,
Transfer to another cluster via the crossbar network 60. The data storage unit 548 stores the information fetched from the crossbar network 60. Instruction decoding means 5
47 determines whether or not the information stored in the data storage means 548 indicates a broadcast command.
The broadcast end monitoring unit 545 determines whether or not end reports have been received from all the clusters.

Referring to FIG. 4, the parameters are V,
It includes seven fields of F, R / W, LC #, RC #, ADD, and LENGTH. The V field is a valid bit. The R / W field stores information indicating the read / write designation of the transfer destination cluster. The LC # field stores the transfer source cluster number. The RC # field stores the transfer destination cluster number. F field is
Stores information indicating that the corresponding data is the last data. The ADD field stores the virtual address of the transfer destination cluster. Transfer data is stored from the address indicated by the ADD field. The LENGTH field stores the length of data to be transferred and the length of data that can be transferred at one time.

Referring to FIG. 5, the parameter holding means 5
42 includes a selector 5420, a register 5421, a parameter buffer 5422, an updating unit 5423, and an address pointer 5424. The selector 5420 is
The parameters received from the control means 543 and the updating means 5
One of the parameters output by 423 is selected.
The selected parameter is stored in the register 5421. The parameters stored in the register 5421 are stored in the parameter buffer 5422. The parameter buffer 5422 stores a plurality of parameters. Among the parameters in the parameter buffer 5422, the one designated by the address pointer 5424 is output from the parameter buffer 5422. Parameter buffer 542
The parameters output from 2 are data holding means 541.
And to the updating means 5423. When data is sent to a plurality of clusters, a plurality of parameters corresponding to each of the destination clusters are generated.

The updating means 5423 updates the contents of the parameters. The updated parameter is the selector 542.
It is again stored in the parameter buffer 5422 via 0 and the register 5421. Updating means 5423
Generates a parameter for each data block when the transfer data is divided into a plurality of data blocks. When a long data is transferred, it is divided into a plurality of data blocks having a predetermined data length. This predetermined data length is stored in the LENGTH field of the parameter. The data blocks are sent to the crossbar network 60 one by one. Update means 5423,
By adding this predetermined data length to the ADD field of the parameter, the parameter corresponding to the next data block is generated. "1" is set in the F field of the parameter corresponding to the last data block.

Referring to FIG. 6, data holding means 541
Includes a register 5411, a data buffer 5412, a selector 5413, a register 5414, and an address pointer 5415. The data holding unit 541 may include a plurality of data buffers 5412. By preparing a plurality of data buffers 5412, the storage of data in the data holding means 541 and the data holding means 541 are performed.
The data can be read from the memory simultaneously. In this embodiment, the data holding unit 541 includes three data buffers 5412.

The data received from the control means 543 is stored in the data buffer 5412 via the register 5411. Of the data in the data buffer 5412, the data designated by the address pointer 5415 is the register 5
It is stored in 414. The data stored in the register 5414 is output to the data transfer means 544. The parameters received from the parameter holding unit 542 are also output to the data transfer unit 544.

Next, the operation of this embodiment will be described with reference to the drawings. In addition, among the following operations, steps 1 to
10 and 19 to 24 are cluster # 1 and step 11
Is executed by the crossbar network 60, and steps 12 to 18 are executed by the cluster # 2.

Referring to FIGS. 7 and 8, step 1
At, the arithmetic processing unit 21 of the cluster # 1 issues a broadcast command.

In step 2, the instruction decoding means 51 determines whether or not this instruction is a broadcast instruction.
When the issued command is a broadcast command, step 3 is executed.

In step 3, the transfer control means 52 instructs the shared memory access means 53 to acquire the parameters. The shared memory access unit 53 acquires information necessary for generating parameters from the shared memory 3 and the arithmetic processing unit 21. Shared memory access means 53
Generates a parameter based on the acquired information.

In step 4, the shared memory access means 53 stores the generated parameter in the parameter holding means 542.

In step 5, transfer control means 52
Determines whether all the parameters have been set in the parameter holding unit 542. Parameter holding means 542
Must be set with a parameter corresponding to each of the transfer destination clusters. In this example, the cluster #
2 and two parameters corresponding to cluster # 3 are set. When all parameters have been set,
Step 6 is executed. Otherwise, steps 3-4 are performed again.

In step 6, transfer control means 52
Instructs the shared memory access means 53 to acquire the data to be transferred. In response to this instruction, the shared memory access unit 53 acquires the data to be transferred from the shared memory 3.

In step 7, the shared memory access means 53 stores the acquired data in the data holding means 541.

In step 8, transfer control means 52
Determines whether storage of all data has been completed.
When the storage of all data is completed, step 9 is executed. Otherwise, steps 6 and 7 are executed again.

Steps 6-8 are only performed once for each broadcast.

In step 9, the control means 543
The parameters and data to be transmitted are read from the parameter holding unit 542 and the parameter holding unit 542, respectively. The read data is sent to the data transfer means 544.

Referring to FIGS. 7 and 9, step 1
At 0, the data transfer means 544 sends the parameters and data to the crossbar network 60. Further, when the data is divided into a plurality of data blocks, the updating means 5423 generates the parameters of the data block to be transmitted next according to the procedure described above. The generated parameters are stored in the parameter buffer 54.
It is set to 22.

In step 11, the crossbar network 60 sends the parameter and data to the cluster designated by the RC # field in the parameter. Here, it is assumed that the parameters and data are first sent to the cluster # 2. The parameters and data are then sent to cluster # 3.

In step 12, the data storage means 548 of cluster # 2 receives and stores the parameters and data therein.

In step 13, the instruction decoding means 54
7 determines the contents of the R / W field of the received parameter. When the R / W field indicates storage, step 14 is executed.

In step 14, the control means 543
Instructs the transfer control means 52 to store the data. In response to this instruction, the transfer control means 52 stores the data in the data storage means 548 in the shared memory 3. The storage position is the address indicated by the ADD field of the parameter.

In step 15, the shared memory access end monitoring means 546 holds the contents of the F field and LC # field of the parameter.

In step 16, the transfer control means 52
Notifies the shared memory access end monitoring means 546 of the end of data storage.

In step 17, the shared memory access end monitoring means 546 determines the end of data transfer. When the F field of the parameter corresponding to the data block stored in steps 13 and 14 is set to "1", step 18 is executed. Otherwise, steps 12-16 are performed on subsequent data blocks.

In step 18, the shared memory access end monitoring means 546 reports the end of data transfer to the transfer source cluster. The transfer source cluster is indicated by the LC # field of the parameter. In this case, the transfer source cluster is cluster # 1.

In step 19, the end notification information transmitted from the cluster # 2 is the crossbar network 60.
Through the data storage means 548 of the cluster # 1.

Referring to FIGS. 7 and 10, in step 20, the instruction decoding means 547 determines whether or not the information in the data storage means 548 is end notification information. When this is the end notification information, step 21 is executed.

In step 21, the instruction decoding means 54
7 notifies the broadcast end monitoring means 545 that the data transfer to the cluster # 2 has ended.

At step 22, the parameter corresponding to the cluster for which the transfer is completed is stored in the parameter holding means 54
2 is deleted.

In step 23, the broadcast end monitoring means 545 determines whether or not the transfer to all clusters has been completed. Specifically, it is determined whether the end reports have been received from all the clusters. When the transfer to all clusters is complete, step 24 is executed.
At other times, steps 9 to 23 are repeatedly executed for the cluster to which the data has not been transferred yet.
At this time, steps 3 to 8 are not repeated. This means that the data is read from the shared memory 3 only once.

At step 24, the control means 543
Releases the data holding unit 541. Further, the control unit 543 notifies the arithmetic processing unit 21 of the end of the broadcast.

As described above, in the first embodiment, the data to be broadcast is read from the shared memory 3 and stored in the data holding means 541. When sending this data to multiple clusters, the shared memory 3
Instead of, the data is read from the data holding unit 541. Data can be read from the shared memory 3 only once.

Referring to FIG. 11, the total time required to execute a broadcast command to N-1 transfer destination clusters is T1, and the time required to read data from the shared memory 3 is Ta1. If the time required for writing the transferred data to a shared memory in one cluster is Tb1, then T1 = Ta1 + Tb
It is 1 * (N-1). The time Ta1 is required only once when transferring to the cluster # 1. In the case of transfer to another cluster, the data stored in the data holding means 541 is sent out, so that the time Ta1 does not take. Thus, according to the first embodiment, data can be broadcast in a shorter time than the conventional information processing system.

In the above embodiment, the parameter may be provided with a field indicating the type of transfer. The transfer types include at least point-to-point transfer and broadcast transfer.

Next, a second embodiment of the present invention will be described with reference to the drawings.

Referring to FIG. 12, in the second embodiment of the present invention, a data transfer control device 7 is provided in each cluster instead of the data transfer control device 5 of the first embodiment. Further, the broadcasting means 8 is provided in the crossbar network 60.

Referring to FIG. 13, the data transfer control device 7 includes an instruction decoding means 71, a transfer control means 72, a shared memory access means 73, and a data transfer control means 74. The shared memory access unit 73 executes reading and writing to the supply memory 3 via the local network 4 and the local network 4. The instruction decoding means 71 receives the instructions issued by the arithmetic processing units 21 to 2N via the local network 4,
Decode this. When this command is a broadcast command, the command decoding unit 71 instructs the transfer control unit 72 to execute the broadcast. The data transfer control means 74 transfers data to another cluster. The transfer control means 72 controls the operations of the instruction decoding means 71, the shared memory access means 73, and the data transfer control means 74.

The data transfer control means 74 is a control means 74.
3, a data transfer unit 744, a shared memory access end monitoring unit 746, an instruction decoding unit 747, and a data storage unit 748. The shared memory access end monitoring means 746 detects that the access to the shared memory 3 has been completed. The instruction decoding unit 747 is the data storage unit 7.
The type of information stored in 48 is determined. The data storage unit 748 acquires data from the crossbar network 60 and stores it. The data transfer means 744 transfers the data read from the shared memory 3 to the crossbar network 60.
To send to.

Referring to FIG. 14, the broadcasting means 8 includes data transfer control devices 81-83. The data transfer control devices 81 to 83 have clusters # 1 to #
3 respectively. The broadcasting means 8 is
Further, it includes arbitration means 84 and transfer means 85. The arbitration unit 84 arbitrates a plurality of competing transfer requests from the cluster # 1 to the cluster # 3. The transfer means 85 sends the data to the transfer destination cluster in response to the transfer request selected by the arbitration means 84.

The data transfer control device 81 has a data storage means 811, an instruction decoding means 812, and a control means 813.
And data transfer means 814. Data storage means 8
11 temporarily holds the data to be broadcast. This data is received from the corresponding cluster.
The instruction decoding unit 812 determines the type of information received from the corresponding cluster. There are at least parameters and data as types of information. Data transfer means 81
4 transfers the data received from the corresponding cluster to the destination cluster. The control unit 813 controls the operations of the data storage unit 811, the instruction decoding unit 812, and the data transfer unit 814.

The data transfer means 814 includes the broadcast control means 8141 and the broadcast end monitoring means 8
142, a data holding unit 8143, and a parameter holding unit 8144. The data holding means 8143
Holds the data received from the corresponding cluster. The parameter holding unit 8144 holds the parameter received from the corresponding cluster. The broadcast end monitoring means 8142 monitors the end of the broadcast, and when the broadcast ends, sends an end notification to the corresponding cluster.

Referring to FIG. 15, the parameter holding means 8144 has the parameter holding means 8145 and 814.
6 inclusive. Parameter holding means 8145 and 8146
Stores parameters corresponding to cluster # 2 and cluster # 3. The structure of the parameter holding means 8145 and the parameter holding means 8146 is substantially the same as that of the parameter holding means 542 shown in FIG.
The structure of the data holding means 8143 is substantially the same as that of the data holding means 541 shown in FIG.

Referring to FIG. 16, the arbitration means 84 includes a first arbitration means 841, a second arbitration means 842, a third arbitration means 843, and a permission signal generating means 844. The first arbitration means 841 communicates between two points (point
Arbitration between two-point communication).

The first arbitration means 841 uses the clusters # 1 to # 1.
A request for point-to-point communication and a transfer destination are received from cluster # 3. The first arbitration means 841 selects one of these requests according to a predetermined priority. The selected request is notified to the permission signal generation means 844.

The second arbitration means 842 arbitrates between the point-to-point communication and the broadcast communication request. The second arbitration unit 842 receives the point-to-point communication request and the broadcast request from the clusters # 1 to # 3. The second arbitration means 842 selects one of these requests according to a predetermined priority. The designer is free to prioritize either point-to-point communication or broadcast communication. The result of the arbitration is notified to the permission signal generating means 844.

The third arbitration means 843 arbitrates between broadcast communication requests. The third arbitration unit 843 receives the broadcast communication request from the cluster # 1 to the cluster # 3. The third arbitration means 843 selects one of these requests according to a predetermined priority. The result of the arbitration is notified to the permission signal generating means 844.

The permission signal generation means 844 sends a permission signal to the clusters to which data transfer is permitted, based on the information notified from the first to third arbitration means 841 to 843.

Next, the operation of the second embodiment will be described with reference to the drawings. In this operation example, the arithmetic processing unit 21 of the cluster # 1 requests the broadcast to the cluster # 2 and the cluster # 3.

Referring to FIGS. 17 and 21, in step 1, the arithmetic processing unit 21 of the cluster # 1 issues a broadcast command.

In step 2, the instruction decoding means 71
Decodes the instruction issued by the arithmetic processing unit 21. When this command is a broadcast command, step 3 is executed.

In step 3, transfer control means 72
Instructs the shared memory access means 73 to acquire the parameters of the data to be broadcast. In response to this instruction, the shared memory access unit 73 acquires information from the shared memory 3 and the arithmetic processing unit 21. Based on this information, the shared memory access means 73 creates a parameter. The generated parameters are sent to the crossbar network 60 by the data transfer means 744. The parameter is generated for each transfer destination cluster. In this operation example, two parameters corresponding to cluster # 2 and cluster # 3 are generated.

Referring to FIG. 18, the parameters sent in step 3 are V, CMD, LC #, RC.
It contains six fields: #, ADD, LENGTH.
The V field is a valid bit. The CMD field stores information indicating the type of transfer. The transfer types include at least point-to-point transfer and broadcast transfer. The LC # field stores information indicating the transfer source cluster number. The RC # field stores information indicating the transfer destination cluster number. The ADD field is
Stores the virtual address of the transfer destination cluster. Transfer data is stored from the address indicated by the ADD field. The LENGH field stores the length of transfer data and the length of data that can be transferred at one time.

Referring again to FIGS. 17 and 21, step 3 is repeated until all parameters have been sent in step 4.

In step 5, the data transfer control device 81 receives the parameter from the cluster # 1. The received data is stored in the data storage unit 811.

In step 6, instruction decoding means 812
Determines whether the information stored in the data storage unit 811 is a parameter. When this information is a parameter, step 7 is executed, in step 7,
The parameters in the data storage means 811 are stored in the parameter holding means. At this time, the parameter corresponding to cluster # 2 is stored in the parameter holding unit 8145, and the parameter corresponding to cluster # 3 is stored in the parameter holding unit 81.
46, respectively.

Referring to FIG. 20, in step 7, the F field is added to the parameter.

In step 8, transfer control means 72
Instructs the shared memory access unit 73 to acquire the data to be broadcast. In response to this instruction, the shared memory access unit 73 reads the data from the shared memory 3. The read data is sent to the crossbar network 60.

In step 9, step 8 is repeated until transmission of all data is completed.

Referring to FIGS. 17 and 22, in step 10, the data transmitted in step 9 is received by the data transfer control device 81. The received data is stored in the data storage unit 811.

In step 11, the instruction decoding means 81
2 determines whether the information stored in the data storage unit 811 is data. If the information is data, step 12
Is executed.

In step 12, the data storage means 8
The data in 11 is transferred to the data holding means 8143.

In step 13, the broadcast control means 8141 sends a broadcast transfer request to the arbitration means 84.

In step 14, the data transfer control device 81 waits for a permission signal from the arbitration means 84. When the permission signal is received, the data transfer control device 81 executes step 15.

In step 15, the broadcast control means 8141 acquires parameters and data from the parameter holding means and data holding means 8143. The acquired parameters and data are sent to the transfer means 85. When the data is divided into a plurality of blocks, after the transfer of the data block, the parameter of the data block to be transferred next is generated. This process is substantially the same as that of the first embodiment. The data blocks are sequentially sent to the transfer means 85 until the F field of the parameter becomes "1".

At step 16, the transfer means 85
The data and the parameter are sent to the transfer destination cluster indicated by the #RC field of the parameter. Transfer means 85
Includes a crossbar network, so that data is transferred to multiple clusters simultaneously. In this case, cluster # 2
And the data is simultaneously transferred to the cluster # 3.

Referring to FIG. 20, the format of the parameters sent in step 16 is the same as that stored in the parameter holding means.

Referring again to FIGS. 17 and 22, in step 17, the destination cluster receives the parameters and data. The operation of cluster # 2 after receiving these parameters and data will be described below. The operation of cluster # 3 is substantially the same as that of cluster # 2. Data storage means 74 of cluster # 2
8 stores the data received from the transfer means 85.

In step 18, the instruction decoding means 74
7 decodes the contents of the parameters stored in the data storage means 748. When the parameter directs the storage of data in cluster # 2, step 19 is executed.

At step 19, the control means 743
Instructs the shared memory access means 73 to store the data in the data storage means 748. In response to this instruction, the shared memory access unit 73 stores the data in the data storage unit 748 in the shared memory 3. The data is stored at the address indicated by the ADD field of the parameter.

Referring to FIG. 23, in step 20, the contents of the CMD, F and LC # fields of the parameters are retained.

At step 21, the shared memory access means 73 causes the shared memory access end monitoring means 746.
Is notified of the completion of data storage.

In step 22, the shared memory access end monitoring means 746 determines the end of data transfer. This determination is made based on the contents of the F field of the transferred data. When the transfer of all the data blocks is completed, step 23 is executed. Otherwise, steps 17-21 are repeated.

Referring to FIGS. 17 and 23, in step 23, the shared memory access end monitoring means 74.
6 notifies the cluster # 1 of the end of the data transfer via the data transfer unit 744 and the crossbar network 60.

In step 24, the end notification information sent by the cluster # 2 is received by the data transfer control device 82 corresponding to the cluster # 2. The received end notification information is the data storage means 811 of the data transfer control device 82.
Stored in.

At step 25, the instruction decoding means 812 of the data transfer control device 82 causes the data storage means 811.
Determine the content of the information stored in. When the information is the end notification information, step 26 is executed.

In step 26, when the end notification information has notified the end of the broadcast transfer, the control means 813 causes the arbitration means 84 to notify the broadcast end monitoring means 8142 of the data transfer control device 81.
Sends end notification information. When the end notification information notifies the end of the point-to-point transfer, the control unit 813 sends the end notification information to the transfer source cluster via the transfer unit 85.

At step 27, the broadcast end monitoring means 8142 of the data transfer control device 81 receives the end notification information. The parameter corresponding to the end notification information in the parameter holding unit 8144 is deleted.

In step 28, the broadcast end monitoring means 8142 determines whether the end notification information has been received from all the transfer destination clusters. When the end notification information is received from all the transfer destination clusters, step 29 is executed.

In step 29, the broadcast end monitoring means 8142 notifies the cluster # 1 of the end of broadcast.

As described above, in the second embodiment, data can be simultaneously transferred to a plurality of clusters.

Referring to FIG. 24, in the second embodiment, the time required for broadcasting is T2, the shared memory 3 is
When the time from reading to reading data is Ta2, and the time required to write the read transfer data to the shared memory 3 in one cluster is Tb2, respectively, T2 = Ta2 + Tb2.

As described above, in the information processing system of the present invention, when executing one broadcast,
Data is read from the main memory only once. Therefore, the data can be broadcast in a short time.

Further, in the information processing system of the present invention, since other clusters do not exist in the transfer path to the transfer destination cluster, data can be transferred at high speed.

Furthermore, in the information processing system of the present invention,
Since a dedicated bus for broadcasting is unnecessary, the amount of hardware does not increase.

Furthermore, in the information processing system of the present invention,
The data transfer control device 5 transfers data in place of the arithmetic processing device. Therefore, even if data transfer is performed, the processing of the arithmetic processing device is not interrupted.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a time required for conventional broadcast transfer.

FIG. 2 is a block diagram showing the structure of the first embodiment of the present invention.

FIG. 3 is a block diagram showing a detailed structure of a data transfer unit 54.

FIG. 4 is a diagram showing a parameter format of the first embodiment of the present invention.

FIG. 5 is a block diagram showing a detailed structure of a parameter holding unit 542.

FIG. 6 is a block diagram showing a detailed structure of a data holding unit 541.

FIG. 7 is a diagram illustrating the operation of the first exemplary embodiment of the present invention.

FIG. 8 is a flowchart illustrating the operation of the first exemplary embodiment of the present invention.

FIG. 9 is a flowchart illustrating the operation of the first exemplary embodiment of the present invention.

FIG. 10 is a flowchart illustrating the operation of the first exemplary embodiment of the present invention.

FIG. 11 is a diagram illustrating a time required for broadcast transfer in the first embodiment of the present invention.

FIG. 12 is a block diagram showing the structure of a second embodiment of the present invention.

FIG. 13 is a block diagram showing a detailed structure of data transfer control means 74.

FIG. 14 is a block diagram showing a detailed structure of broadcasting means 8.

FIG. 15 is a block diagram showing a detailed structure of a parameter holding unit.

16 is a block diagram showing a detailed structure of arbitration means 84. FIG.

FIG. 17 is a diagram for explaining the operation of the second embodiment of the present invention.

FIG. 18 is a diagram showing a parameter format in the second embodiment of the present invention.

FIG. 19 is a diagram showing a parameter format in the second embodiment of the present invention.

FIG. 20 is a diagram showing a parameter format in the second embodiment of the present invention.

FIG. 21 is a flowchart for explaining the operation of the second embodiment of the present invention.

FIG. 22 is a flowchart illustrating the operation of the second embodiment of the present invention.

FIG. 23 is a flowchart illustrating the operation of the second embodiment of the present invention.

FIG. 24 illustrates the time required for broadcast transfer in the second embodiment of the present invention.

[Explanation of symbols]

 21 to 2N arithmetic processing unit 3 shared memory 4 local network 5 data transfer control unit 51 instruction decoding unit 52 transfer control unit 53 shared memory access unit 54 data transfer unit 541 data holding unit 542 parameter holding unit 543 control unit 544 data transfer unit 545 Broadcast end monitoring means 546 Shared memory access end monitoring means 547 Instruction decoding means 548 Data storage means 5420 Selector 5421 Register 5422 Parameter buffer 5423 Update means 5424 Address pointer 5411 Register 5412 Data buffer 5413 Selector 5414 Register 5415 Address pointer 60 Crossbar network 7 Data transfer Control device 71 Instruction decoding means 72 Transfer control means 73 Shared memory access Means 74 data transfer means 743 control means 744 data transfer means 746 shared memory access end monitoring means 747 instruction decoding means 748 data storage means 8 broadcasting means 81-83 data transfer control device 811 data storage means 812 instruction decoding means 813 control means 814 Data transfer means 8141 Broadcast control means 8142 Broadcast end monitoring means 8143 Data holding means 8144 Parameter holding means 8145 Parameter holding means 8146 Parameter holding means 84 Arbitration means 841 First arbitration means 842 Second arbitration means 843 Third arbitration means 844 Permission signal generating means 85 Transfer means

Claims (10)

[Claims] 1. A multi-cluster information processing system including a plurality of clusters and an inter-cluster connecting means for connecting the plurality of clusters, wherein each of the plurality of clusters is accessed by a plurality of processors and the plurality of processors. Shared memory, an intra-cluster connection unit that connects the plurality of processors to the shared memory, a data holding unit that holds transfer data to be transferred to another cluster, and at least one of the plurality of processors is a plurality. A transfer data reading unit that reads transfer data from the shared memory and stores the transfer data in the data holding unit when requesting data transfer to the transfer destination cluster; Sequential transfer to each of the plurality of transfer destination clusters via the crossbar network Multi-cluster processing system which comprises a data transfer unit. 2. The multi-cluster according to claim 1, wherein each of the plurality of clusters includes transfer data storage means for storing transfer data received via the inter-cluster connection means in the shared memory. Information processing system. 3. The data transfer means, wherein each of the plurality of clusters includes parameter storage means for storing a parameter including information indicating a transfer destination cluster and an address at which transfer data is to be stored, for each transfer destination cluster. 2. The multi-cluster information processing system according to claim 1, wherein the parameter stores the parameter stored in the parameter storage unit together with the transfer data to the transfer destination cluster. 4. The data transfer means divides the transfer data into a plurality of data blocks and transfers the data, and each of the plurality of clusters sequentially adds the data length of the data blocks to the address in the parameter. The multi-cluster information processing system according to claim 1, further comprising parameter updating means for sequentially generating parameters corresponding to the plurality of data blocks. 5. The transfer data storage unit sends an end notification to a transfer source cluster when receiving the transfer data is completed, and each of the plurality of clusters receives the end notification from all transfer destination clusters. The multi-cluster information processing system according to claim 1, further comprising a termination monitoring unit that determines whether or not the multi-cluster information processing is performed. 6. A multi-cluster information processing system including a plurality of clusters and an inter-cluster connection means for connecting the plurality of clusters, wherein the inter-cluster connection means receives the transfer data received from one of the plurality of clusters. A multi-cluster information processing system comprising: a data holding / storage means for storing the data; and a data transfer means for simultaneously sending the transfer data to a plurality of transfer destination clusters. 7. The inter-cluster connecting means includes a parameter holding means for storing a parameter received from a transfer source cluster, and the parameter includes information indicating a transfer destination cluster and an address at which the transfer data is to be stored. 7. The multi-cluster information processing system according to claim 6, wherein the data transfer means sends the parameter together with the transfer data. 8. The data transfer means divides transfer data into a plurality of data blocks for transfer, and the inter-cluster connection means sequentially adds the data length of the data blocks to the address in the parameter. 8. The multi-cluster information processing system according to claim 7, further comprising parameter updating means for sequentially generating parameters corresponding to the plurality of data blocks. 9. The inter-cluster connecting means includes an end monitoring means for judging that the reception end notifications have been received from all the transfer destination clusters and for notifying the reception end notifications to the transfer source clusters. The described multi-cluster information processing system. 10. The multi-cluster information processing system according to claim 6, wherein said inter-cluster connection means has arbitration means for arbitrating data transfer between clusters.