JPH10336175A - Multi-node information processing system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce load on a processor when an identical data is transferred to a shared memory for plural nodes and to perform multicast to an optical node when data is transferred to a shared memory for plural nodes. SOLUTION: A multicast instruction to be used when data transfer is instructed from a self-node to plural other nodes is provided in an inter-node data transfer instruction which is given as a signal from processors 61 and 6N in a node 1 to a transfer controller in the same node in accordance with the data transfer from the node 1 to a node 2 and plural nodes. Also, multicast control parts 6 and 7 or a multicast control part 8, which hold operation parameters necessary for data and transfer, supervise the ends of multicast instructions and control multicast to plural nodes, are provided in transfer controller 3 and 4 or an inter-node crossbar network 5.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a multi-node information processing system, and more particularly to a multi-node information processing system having a node configuration.

[0002]

2. Description of the Related Art A conventional multi-node information processing system is used for the purpose of performing one-to-one data transfer between nodes. That is, in a system in which a plurality of nodes are connected by an inter-node network, when data transfer is performed between nodes (actually, between memories), a transfer instruction is first issued from an arithmetic processing unit of a transfer source node, and this operation is performed. Data is transferred from the memory of the node to which the processing device belongs to the memory of the transfer destination node. If the same data is to be transferred to a plurality of nodes (hereinafter referred to as multicast), the arithmetic processing unit of the transfer source node issues the same number of transfer instructions as each of the transfer destination nodes, and Necessary information on transfer and transfer data are read from the memory of the node to which the node belongs, and are transferred serially to the transfer destination node. FIG. 17 shows this state in a time chart. For the configuration of the system used for the description, FIG. 1 which is the principle diagram of the present invention is applied mutatis mutandis. However, the multicast control unit is not used. Here, a time chart in the case where transfer is performed five times to two nodes 1a and 1b is shown. First, the time required from the issuance of the multicast instruction of the arithmetic processing unit 61 to the instruction identification of the transfer control device 3 is Ti, the time for reading the parameters from the shared memory and transferring to the transfer control device 3 via the local network is Tp, and the data is The time for reading from the shared memory and transferring to the transfer control device 3 via the local network is Td0, the time for transferring parameters and data from the transfer source node 1 to the crossbar network 5 between nodes is Td1, and the time for transferring parameters and data is 5 between the nodes. Is Td2, the time required for the transfer destination node 2 to write data to the shared memory is Td3, and the time required for reporting the completion to the transfer source node is Te. And

[0003] As an improvement of this, for example, in Japanese Patent Application Laid-Open No. 08-088627, transfer is performed in a communication network between processors in a parallel computer system. A technique is described in which all nodes are grouped and communicated via nodes in the group within the group, thereby achieving a purpose by a method of hierarchically assigning addresses to a plurality of groups.

[0004]

The above-mentioned conventional multi-node information processing system issues the same number of transfer instructions as each transfer destination node when performing multicasting between nodes, and relates to the same number of transfers as this instruction. It is necessary to read out necessary information and transfer data from the memory of the node to which it belongs, and a large amount of overhead is required for the transfer source arithmetic processing unit, the memory, and the network connecting them. There is a problem that there is a possibility. Also, JP-A-08-08
In the device disclosed in Japanese Patent No. 8627, nodes connected to a network are restricted by a group, and addresses are assigned hierarchically to each group. Therefore, an arbitrary node is accessed by communication between nodes. In this case, there is a case where the target node passes through a group to which the node does not belong, which leads to an increase in access time and an increase in network load, and there is also a problem that these may lead to a decrease in system performance.

It is an object of the present invention to provide a multi-node information processing system in which the same data is transferred to a shared memory of a plurality of nodes with a small load on an arithmetic processing unit.

Another object of the present invention is to enable multicasting to an arbitrary node when transferring the same data to a shared memory of a plurality of nodes.

[0007]

A multi-node information processing system according to the present invention comprises a plurality of processing units, a shared memory shared and used by the processing units, and data transfer in accordance with a data transfer request of the processing units. A plurality of nodes each comprising a transfer control device for controlling operation and a local network for interconnecting the arithmetic processing device, the shared memory, and the transfer control device, are connected to each other by an inter-node crossbar network. In a multi-node information processing system in which data is transferred between nodes, an inter-node data transfer instruction, which is a transfer instruction sent from an arithmetic processing unit in the arbitrary node to a transfer control unit in the same node, includes A multicast command for transferring data from a node to a plurality of other nodes; Is configured to include a multicast control unit for controlling multicast to a plurality of nodes receives the.

[0008] In the multi-node information processing system according to the present invention, the multicast control unit may be provided in the transfer control device.

In the multi-node information processing system according to the present invention, the multicast control unit may be provided in the inter-node crossbar network.

[0010] The multi-node information processing system according to the present invention includes an instruction decoding unit for decoding the transfer instruction sent from the arithmetic processing unit in the transfer control device, and a logical address for accessing the shared memory. A shared memory access unit that includes an address conversion unit that performs physical address conversion from and accesses the shared memory; and a data holding unit that holds data for single-cast and multicast transfer control transmitted from the shared memory access unit. Parameter holding means for holding the operation parameters for the number of nodes required for the transfer sent from the shared memory access section and generating an address, and multicast end monitoring means for monitoring the end of the multicast command issued by the node to which the node belongs, and The data holding means and the parameter A multicast control unit for controlling multicast for a plurality of nodes including a multicast control unit for controlling the multicasting unit and the multicast end monitoring unit, and a logical node number to a physical node when performing single-cast and multicast inter-node data transfer. Node number conversion means for converting the number into a number, shared memory access end monitoring means for monitoring the end of access to the shared memory in the own node from another node, transfer means for sending data to the crossbar network between the nodes, and the node A data transfer control unit that includes a data transfer control unit that controls a number conversion unit, the shared memory access end monitoring unit, and the transfer unit, and that controls data transfer between the own node and the crossbar network between the nodes; Between nodes A temporary save unit for temporarily saving data transfer information, a transfer control unit for controlling the command decoding unit, the shared memory access unit, the multicast unit, the data transfer control unit, and the temporary save unit. A crossbar control unit for controlling data transfer to a transfer destination node may be provided in the internode crossbar network.

In the multi-node information processing system according to the present invention, the parameter holding means may include a one-word parameter register for holding a parameter and a circuit for generating an address.

[0012] The multi-node information processing system according to the present invention includes an instruction decoding unit for decoding the transfer instruction sent from the arithmetic processing unit in the transfer control device, and a logical address when accessing the shared memory. A shared memory access unit that includes an address translation unit that performs a physical address translation from and accesses the shared memory, and a shared memory access end monitoring unit that monitors the end of access to the shared memory in the own node from another node. A transfer unit for sending data to the inter-node crossbar network and a data transfer control unit for controlling the shared memory access end monitoring unit and the transfer unit, and controls data transfer between the own node and the crossbar network between nodes. Data transfer control unit and data transfer between nodes from other nodes A temporary evacuation unit for temporarily evacuation of said information, a transfer control unit for controlling said instruction decoding unit, said shared memory access unit, said data transfer control unit and said evacuation means, A plurality of inter-node transfer controllers having an interface with each node corresponding to each node on a one-to-one basis, and connecting all the inter-node transfer controllers with a crossbar to perform data transfer between the inter-node transfer controllers. An inter-node data transfer unit that performs the operation, and a network control unit that controls the inter-node transfer control unit and the inter-node data transfer unit may be provided.

[0013] The multi-node information processing system according to the present invention, in the inter-node transfer control device, a decoding unit for decoding a data transfer instruction between the nodes sent from the node, and a single-cast and multicast inter-node A node number conversion unit that receives a logical node number and converts it to a physical node number when performing data transfer, and a transfer unit that sends transfer data in the inter-node transfer control device to a corresponding node or the inter-node data transfer unit, A data holding unit for holding data sent from a corresponding node, and a parameter holding unit for holding addresses for generating the address by holding parameters corresponding to the number of partner nodes required for transfer data sent from the corresponding node; Means for monitoring the end of the multicast issued by the node; A multicast control unit for controlling multicast to a plurality of nodes, the multicast control unit including a multicast control unit for controlling the data holding unit, the parameter holding unit, and the multicast end monitoring unit; the decoding unit; the node number conversion unit; And a transfer control unit that controls the unit and the multicast control unit.

In the multi-node information processing system according to the present invention, the parameter holding means may include a one-word parameter register for holding a parameter and a circuit for generating an address.

[Operation] As shown in the principle diagram of the multi-node information processing system of FIG. 1, the multi-node information processing system of the present invention transfers data from node 1 to node 2 and a plurality of nodes (not shown). Is performed, the arithmetic processing units 61 and 6 in the node 1 are used as signals.
A multicast command used when instructing data transfer from its own node to a plurality of other nodes is provided in the inter-node data transfer command given from N to the transfer control device 3 in the same node. Multicast control units 6 and 7 or multicast control units which hold data and operation parameters necessary for transfer in the crossbar network 5 between nodes, monitor the end of a multicast command, and control multicast to a plurality of nodes. Providing 8 makes it possible to issue the multicast command only once when transferring the same data to a plurality of nodes.

[0016]

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

FIG. 2 is a block diagram showing first and second embodiments of the present invention.

The multi-node information processing system 1 of the present invention
0 is composed of nodes 11 and 12 of the same type, a plurality of nodes (not shown), and an inter-node crossbar network 13 connecting these nodes. Each node (described as the node 11 as a representative) has N operations. Processing equipment 61-6N
And the shared memory 14 shared by the arithmetic processing units 61 to 6N
A local network 15 having an interface between the arithmetic processing units 61 to 6N and the shared memory 14, a local network 15 and a crossbar network 1 between nodes.
3 and comprises a transfer control device 16 for controlling data transfer between the shared memory 14 of the node 11 and the shared memory of another node.

Further, the transfer control device 16 issues a transfer instruction for transferring data in the shared memory 14 in the node 11 as a transfer source to a shared memory in another node as a transfer destination.
Arithmetic processing units 61 to 61 via local network 15
6N, a command decoding unit 17 that receives and decodes data from the 6N, a transfer control unit 18 that controls data transfer between nodes according to the command decoded by the command decoding unit 17, and a data transfer process between nodes according to the instruction of the transfer control unit 18. A shared memory access unit 19 for accessing the shared memory 14 of the node 11 via the local network 15;
8, the shared memory access unit 19 or a multicast control unit 20 described below and the inter-node network 1
When multicasting is performed from the data transfer control unit 21 that performs data transfer between the node 3 and the node 11, the operation parameters (valid bit, operation instruction command, final transfer flag,
A multicast control unit 20 that holds a transfer destination node number, a transfer source node number, a shared memory logical address in the transfer destination node, address generation information) and transfer data, and controls a multicast;
One-to-one transfer between nodes that correspond one-to-one (hereinafter 1: 1)
And a temporary save unit 22 that temporarily saves 1: 1 transfer information when the node 11 is the transfer destination node when the multicast transfer destination node between the nodes is the same node. doing.

The shared memory access unit 19 includes address conversion means 23 for converting a logical address to a physical address when accessing the shared memory 14 in the node 11 by transferring data between nodes. The data transfer control unit 21 converts a logical node number into a physical node number when transferring data from the node 11 to another node, and a data transfer from the other node to the shared memory 14 in the own node. At this time, information necessary for the end report (hereinafter referred to as carry-over information) is held, the end of access to the shared memory 14 is monitored, and the final data transferred to the node 11 indicates the end of access to the shared memory 14. When identified, the shared memory access end monitoring means 25 for reporting the end of data transfer to the transfer source node from the handover information, the data transfer means 26 for sending data to the inter-node crossbar network 13, and the node number conversion means 24 Shared memory access end monitoring means 25
And data transfer control means 27 for controlling the data transfer means 26.

The multicast control unit 20 comprises a data holding means 28 comprising a data buffer for holding data to be multicast sent from the shared memory access unit 19, and a parameter buffer for holding operation parameters sent from the shared memory access unit 19. , A parameter holding unit 29 composed of a circuit for generating an address, a multicast end monitoring unit 30 for monitoring the end of a multicast command issued from the node 11, a data holding unit 28, a parameter holding unit 29, and a multicast end monitoring unit 30. Control means 31 for controlling
And

The inter-node crossbar network 13 includes a crossbar control unit 32 for controlling data transfer to a destination node by referring to a destination node number in a parameter from a source node when transferring data between nodes. .

Next, the operation of the first embodiment will be described with reference to FIG.
A description will be given based on FIG. FIG. 3 shows the first and second
FIG. 4 is a flowchart (part 2) of the multicast operation according to the first and second embodiments. The description uses the names and reference numerals in FIG.

First, the arithmetic processing unit 6 for issuing a transfer instruction
The data transfer from the shared memory 14 of the node 11 (hereinafter, the transfer source node) including the 1 to 6N to the shared memory of the other plurality of nodes (hereinafter, the transfer destination node) includes
Shall instruct.

First, the transfer command issued by the arithmetic processing unit 61 in step (hereinafter referred to as S) 1 includes a field to be a multicast command command between nodes (denoted as MC in the figure) (provided that it is MC). The instruction format is not shown). Further, the memory in the system is defined in one logical address space, and the logical address is constituted by the logical node number and the logical address of the shared memory in the node. The data to be multicast is transferred to a different logical address of the shared memory in each node. The transfer control device 16 transmits the transfer command issued by the arithmetic processing device 61 via the local network 15
Received by the instruction decoding unit 17. The command decoding unit 17 analyzes the transfer command in S2 and, if the received transfer command is a multicast command between nodes, sends the control information to the transfer control unit 18 together with an instruction to perform multicast. The transfer control unit 18 receives this control information and
Instructs the shared memory access unit 19 to read the operation parameters from the shared memory 14 via the local network 15 and transfer the operation parameters to the multicast control unit 20. Shared memory access unit 19 receiving the instruction
When accessing the shared memory 14 of the own node, the address conversion means 23 converts the address from the logical address to the physical address, reads out the operation parameters from the shared memory 14 in S4 and sends it to the multicast control unit 20,
The multicast control means 31 identifies the operation parameter sent from the shared memory access unit 19, and
Hold the operation parameters in the parameter holding means 29,
In S6, it is determined whether or not the parameter setting has been completed. If the setting is completed, next, in S7, the data to be transferred from the shared memory 14 via the local network 15 is read out to the shared memory access unit 19,
It issues an instruction to transfer to the multicast control unit 20. Upon receiving the instruction, the shared memory access unit 19 reads the data to be transferred from the shared memory 14 in S8 and sends the data to the multicast control unit 20, and the multicast control unit 31
, The data sent from the shared memory access unit 19 is identified, the data is held in the data holding unit 28 in S9, and whether or not the data setting is completed is identified in S10. If the setting is completed, next, in S11, the operation parameters and data to be transferred are read out according to the instruction of the transfer control unit 18 and sent to the data transfer control unit 27 of the data transfer control unit 21.

Here, the operation of the multicast control unit 20 will be described in more detail. In the multicast control unit 20, the operation parameter of the number of nodes to be multicast is stored in the parameter storage unit 29, and under the control of the multicast control unit 31, Parameter holding means 2
At 9, the following address generation processing of the shared memory in the transfer destination node is performed. For the first node, the parameter corresponding to the first node is read from the parameter buffer in the parameter holding means 29. At the same time, the first transfer data is read from the data buffer of the data holding unit 28. The read parameters and data are combined and sent to the data transfer control unit 21 as the first transfer data for the first node. So far,
This is the operation of S11 described above.

FIG. 5 is a block diagram showing the configuration of the data holding means and the parameter holding means in the multicast control unit. The parameter holding means 29 is provided in the address generation circuit 3
3 and a parameter buffer 34 having a plurality of words.

Following S11, in the parameter holding means 29, the address generation circuit 33 stores the logical address for storing the second transfer data to the shared memory in the first node in S12 in the operation parameter of the first node. Is generated from the address and the address generation information. In S13, it is determined whether or not the next address is the final transfer to the transfer destination node.
4 (hereinafter referred to as address resetting), and again
Return to 11. An operation parameter corresponding to the second node is read out from the second node in the same manner as the first node, and is combined with the first transfer data to form a data transfer control unit 21.
And reset the address. Similarly, the first transfer data is transmitted for all the nodes to be multicast. For the second and subsequent transfer data, the second and subsequent transfer data are read from the data holding unit 28 and sent to the data transfer control unit 21 in the same manner as the first transfer data. At the time of resetting the address, if the transfer data of the next address is identified as the last transfer data of the multicast in S13,
At S15, the final transfer flag is set at the same time as the address resetting.

The data transfer control unit 27 of the data transfer control unit 21 that has received the operation parameters and data sent from the multicast control unit 20 in S11 first proceeds to S16.
Causes the node number conversion means 24 to convert the transfer destination node number of the operation parameter from the logical node number to the physical node number, and causes the data transfer means 26 to transmit the operation parameter and data to the inter-node crossbar network 13 in S17. The data format transmitted at this time has a form as shown in an explanatory diagram illustrating the data format of data transmitted from the node in FIG. 6 to the crossbar network between nodes.

In the crossbar network 13 between nodes,
Upon receiving the operation parameters and data from the transfer control device 16, the crossbar control unit 32 sends the operation parameters and data to the node number specified in the operation parameters in S18. The data format sent at this time is
This is a form as shown in an explanatory diagram for explaining a data format of data transmitted from the inter-node crossbar network to the nodes in FIG.

Next, the operation of the transfer destination node (here, the node 12) will be described with reference to FIG.

The transfer destination node 12 receives the operation parameters and data transferred in S19 by a transfer control device (not shown). In the transfer control device, in step S20, the instruction decoding means decodes the transferred operation parameter, and when the instruction is MC transfer and instructs writing to the shared memory of the own node, the transfer control unit transmits to the shared memory in S21. Instruct ADD in the parameter to write data, and execute S2
In step 2, information (transfer source node number, final transfer flag) (hereinafter referred to as carry-over information) for notifying the transfer source node of the transfer completion report from the operation parameter is sent to the data transfer control unit, and is sent to the shared memory access unit. Send operating parameters and data. The data transfer control unit stores the portable information in the shared memory access end monitoring unit. The shared memory access unit converts an address for the shared memory in the operation parameter from a logical address to a physical address by the address conversion unit, and writes the physical address to the shared memory via the local network. When the data writing is completed, the shared memory notifies the transfer control device via the local network of a shared memory access completion report indicating that the data writing to the shared memory has been completed. In the transfer control device, the above-mentioned end report is decoded by the command decoding unit under the control of the transfer control unit, sent to the data transfer control unit, and the end report is sent to the shared memory access end monitoring means in S23. In the shared memory access end monitoring means, S
At 24, whether or not the transfer data corresponding to the end report is the last transfer data of the data transfer to the node 12 is identified from the last transfer flag of the revolving information. If the transfer data corresponding to the end report is not the final transfer, the shared memory access monitoring unit invalidates the carry-around information in S25 and continuously monitors the end report of the corresponding transfer data in S24. Return. If the transfer data corresponding to the end report is the last transfer,
Similarly, at step 26, the node information conversion unit 24 invalidates the switching information and converts the node indicated by the transfer source node number of the switching information from the logical node number to the physical node number.
Then, via the inter-node crossbar network 13, an end-node data transfer end report indicating that the data transfer to the own node has ended is notified. In the transfer source node, the transfer control device 16 receives an end report from the transfer destination node in S27. The transfer control device 16 determines whether or not the one received in S28 is an end report. If the end report is received, the instruction decoding unit 17 decodes the end report from the transfer destination node under the control of the transfer control unit 18. Then, the packet is sent to the multicast control unit 20 in S29. Subsequently, the multicast control unit 20 sends an end report in S30 to the multicast end monitoring means 3.
0, and identifies the transfer destination node whose completion has been reported, invalidates the corresponding operation parameter, and determines whether or not the multicast has ended in S31. If the multicast has not ended, the process continues in S32. The process returns to S31 to monitor the end report of the transfer data. Lastly, if it is determined in step S31 that all nodes have reported termination, the termination reports are received from all nodes that perform multicasting, all corresponding operating parameters are invalidated, and in step S33, the completion of multicasting is indicated. The arithmetic processing unit 61 that has issued the multicast instruction via the local network 15
Notify to The arithmetic processing unit 61 identifies the end of the multicast command from the end report in S34, and ends a series of operations.

Next, in the first embodiment, in the crossbar network 13 between nodes, multicast and 1:
A case where the transfer destination node of one transfer is the same node will be described.

In this case, contention of the destination node occurs between the multicast and the 1: 1 transfer. The contention arbitration is performed by the crossbar control unit 31, and the multicast is preferentially processed. In the contention arbitration between multicasts, any multicast is processed according to the priority.

In the case of a conflict between the first transfer data of the multicast and the first transfer data, the processing of the 1: 1 transfer is temporarily interrupted, and the 1: 1 transfer is resumed after the end of the multicast. Similarly, when a conflict with the 1: 1 transfer occurs during the processing of the multicast, the processing of the 1: 1 transfer is temporarily suspended, and the 1: 1 transfer is resumed after the end of the multicast. When a multicast occurs during the 1: 1 transfer and a conflict occurs, the crossbar control unit 31 notifies the transfer source and the transfer destination node of the 1: 1 transfer of an interruption instruction for instructing a temporary interruption of the transfer. When the transfer source node receives the interruption instruction,
When the data transfer of the 1: 1 transfer is temporarily interrupted, and the transfer destination node 12 (here, it is assumed that the multicast and the 1: 1 transfer compete with each other at the transfer destination node 12), when the transfer instruction is received, the transfer is performed. Under the control of the transfer control unit in the control device,
The portable information about the 1: 1 transfer of the shared memory access end monitoring means, which is decoded by the instruction decoding unit and is provided in the data transfer control unit, is temporarily saved in the temporary saving unit. When the data transfer by the multicast command is completed, the portable information of the temporary save unit is stored again in the shared memory access completion monitoring means. After the end of the multicast, the crossbar control unit 31 notifies the transfer source node of a restart instruction for restarting the suspended 1: 1 transfer. The transfer source node that has received the restart instruction restarts the 1: 1 transfer.

As described above, when transferring the same data to a plurality of nodes, the arithmetic processing unit 61 that issues a transfer instruction in the node 11 issues a multicast instruction to execute the instruction issuing operation and the shared memory 14 from the shared memory 14. The operation of reading the transfer data can be performed once each. Data transfer from the transfer source node to the transfer destination node via the inter-node crossbar network 13 is performed as follows.
This is performed serially for the number of transfer destination nodes.

FIG. 8 is a time chart of the first and second embodiments of the present invention. The time required from the issuance of the multicast instruction of the arithmetic processing unit 61 to the instruction identification of the transfer control unit 16 is Ti, and the time required to read the parameters from the shared memory 14 and transfer them to the parameter holding unit 29 (35) via the local network 15 is Tp. , The data is read from the shared memory 14 and the local network 1
5 is Td0, the transfer time to the data holding means 28 via
The time for transferring the parameters and data from the transfer source node 11 to the inter-node crossbar network 13 is Td1.
Is the transfer time from the transfer control device of the transfer destination node to Td.
2. At the transfer destination node, the time required to write data to the shared memory is Td3, and the time required to report the completion to the transfer source node is Te (here, 5 times for the two nodes 1a and 1b). It is a time chart of the multicast in which the transfer is performed twice).

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

FIG. 2 is a block diagram showing first and second embodiments of the present invention.

The difference between the first and second embodiments lies only in the configuration of the parameter holding means of the multicast control unit. Therefore, in the reference numerals and names in FIG.
Only the sign of the parameter holding means is to be read as 35. Although not used for description, FIGS. 3 and 4 are also applicable.

FIG. 13 is a block diagram of the parameter holding means used in the second embodiment. The configuration of the parameter holding means 35 used here is such that a parameter register 36 having a size of one parameter word and an address generation circuit 37 are paired and the number of nodes minus one is connected in parallel.

First, the arithmetic processing unit 6 that issues a transfer instruction
Node 11 including 1 to 6N (hereinafter referred to as transfer source node)
A case will be described in which the arithmetic processing unit 21 instructs data transfer from the shared memory 14 to the shared memory of another plurality of nodes (hereinafter, referred to as transfer destination nodes). The transfer command issued by the arithmetic processing unit 61 includes a field for forming a multicast command between nodes (the command format is not shown). The memory in the system is defined in one logical address space, and the logical address is constituted by a logical node number and a logical address of the shared memory in the node. The multicast data is transferred to the same address in the logical address of the shared memory in each node.

In the second embodiment, the same operation as in the first embodiment is performed, and the parameters are stored in the parameter holding unit 35 in the multicast control unit 20. In the second embodiment, the parameters are stored in the parameter register 36 having a size of one word, and the address reset is performed as follows. The first node combines the parameter of the first node in the parameter register 36 in the parameter holding unit 35 with the data read out of the first transfer data from the data buffer of the data holding unit 28, and The data is sent to the data transfer control unit 21 as the first transfer data to the node. In the parameter holding unit 35, in the address generation circuit 37, the second node number is generated from the transfer destination node number in the parameter, and is stored as a parameter for the second node in the parameter register. Write). In the second node, similarly to the first node, the parameter of the second node from the parameter register 36 and the first transfer data are combined and sent to the data transfer control unit 21 to reset the node number. Similarly, the first transfer data is sent to all the multicast target nodes. After transferring the first transfer data for all the nodes to be multicast, the address generation circuit 37 generates the first node number and the storage address of the second transfer data from the address generation information, and stores them in the parameter register again ( Hereinafter, it is referred to as an address reset). For the second and subsequent transfer data, the second and subsequent transfer data are read from the data holding unit and sent to the data transfer control unit 21 in the same manner as the first transfer data, and the node number reset and the address reset are repeated. At the time of resetting the address, if the next transfer data is identified as the last transfer data of the multicast, a final transfer flag is set simultaneously with the reset of the address. Thereafter, the same operation is performed.

In the second embodiment, an end report is made as follows. When the transfer source node receives the inter-node data transfer end report from the transfer destination node as in the first embodiment, the transfer control device 16 controls the transfer control unit 18 to transmit the end report from the transfer destination node. The command is decoded by the command decoding unit 17 and sent to the multicast control unit 20. The multicast control unit 20 sends an end report to the multicast end monitoring means 30, and identifies the transfer destination node whose end is reported. Upon receiving the end reports from all the nodes that perform the multicast, the parameter is invalidated, and a multicast end report indicating that the multicast has ended is sent to the arithmetic processing unit 61 that has issued the multicast instruction via the local network. . The arithmetic processing unit 61 identifies the end of the multicast command from the end report.

The inter-node crossbar network 13
Is the same as in the first embodiment.

As described above, when transferring the same data to a plurality of nodes, the arithmetic processing unit 61 that issues a transfer instruction in the node 11 issues a multicast instruction to execute the instruction issuing operation and the shared memory from the shared memory. It is possible to perform the operation of reading the parameter and the operation of reading the transfer data once each. Further, data transfer from the transfer source node to the transfer destination node via the inter-node crossbar network 13 is performed serially for the number of transfer destination nodes.

Next, a third embodiment of the present invention will be described with reference to the drawings. Note that the same names and reference numerals are used for those that can be used in common with the first and second embodiments.

FIG. 10 is a block diagram showing the third and fourth embodiments of the present invention.

The multi-node information processing system 4 of the present invention
0 is composed of nodes 41 and 42 of the same type, a plurality of nodes (not shown), and an inter-node crossbar network 43 connecting these nodes. Each node (described as a node 41 as a representative) has N operations. Processing equipment 61-6N
And the shared memory 14 shared by the arithmetic processing units 61 to 6N
A local network 15 having an interface between the arithmetic processing units 61 to 6N and the shared memory 14, a local network 15 and a crossbar network 4 between nodes.
3 and a transfer control device 44 for controlling data transfer between the shared memory 14 of the node 41 and the shared memory of the other node.

Further, the transfer control device 44 issues, via the local network 15, a transfer command for transferring the data stored in the shared memory 14 in the transfer source node 41 to the shared memory of another transfer destination node. Instruction decoding unit 17 that receives and decodes from the arithmetic processing units 61 to 6N
A transfer control unit 18 for controlling data transfer between nodes in accordance with the command decoded by the command decoding unit 17;
8, a shared memory access unit 19 that accesses the shared memory 14 of the node 11 via the local network 15 at the time of data transfer between the nodes, and a shared memory access unit 19 A data transfer unit 45 for transferring data between the networks 43;
When the transfer destination node of the 1: 1 transfer between the nodes and the multicast transfer destination between the nodes is the same node, the information of the 1: 1 transfer is temporarily saved when the node 41 is the transfer destination node And a temporary evacuation unit 22 to be operated.

The shared memory access unit 19 includes address conversion means 23 for converting a logical address to a physical address when accessing the shared memory 14 in the node 11 by transferring data between nodes. Data transfer unit 45
When the data is transferred from the other node to the shared memory 14 in the own node, it holds the rotation information necessary for the end report, monitors the end of the access to the shared memory 14, and stores the final data of the transfer to the node 41. Is shared memory 14
A shared memory access end monitoring unit 25 that reports the end of data transfer to the transfer source node from the traveling information when the end of access to the network is identified, and a data transfer unit 2 that sends data to the crossbar network 43 between nodes.
6 and a data transfer control means 27 for controlling the shared memory access end monitoring means 25 and the data transfer means 26.

The inter-node crossbar networks 43 are all of the same type and correspond one-to-one to each node, and have an interface with each corresponding node.
6, 47 and an inter-node transfer controller (not shown), and a crossbar connection between these inter-node transfer controllers,
An inter-node data transfer unit 48 for transferring data between the inter-node transfer control devices, and inter-node transfer control devices 46 and 4
7 and a network control unit 49 for controlling the inter-node transfer control device and the inter-node transfer unit 48 (not shown).

Here, the inter-node transfer control device 46 includes a decoding unit 50 for decoding an operation instruction of data transfer between nodes sent from the node 41 to another node, and a parameter included in a parameter when performing inter-node data transfer. A node number conversion unit 51 for converting a certain node number from a logical node number to a physical node number; A multicast control unit 53 that holds the parameters and data sent from the node 41 and controls the multicast, and a data transfer control unit 54 that controls each unit constituting the inter-node transfer control device 46. Make up.

The multicast control unit 53 further comprises a data holding means 55 composed of a data buffer for holding data to be multicast, and -1 for the number of nodes in the system for holding parameters for multicasting.
A parameter holding means 56 comprising a number of parameter registers and a circuit for generating an address corresponding to each parameter register; and a multicast end monitoring means 57 for monitoring the end of the multicast instruction issued by the node 41. It comprises a multicast control means 58 for controlling a 55, a parameter holding means 56 and a multicast end monitoring means 57.

Next, the operation of the third embodiment will be described with reference to FIG.
1 and FIG. FIG. 11 is a flowchart (part 1) of the multicast operation of the third and fourth embodiments, and FIG. 12 is a flowchart (part 2) of the multicast operation of the third and fourth embodiments. FIG. 9 is a block diagram of a parameter holding unit used in the third embodiment. The configuration of the parameter holding means 56 used here is such that the parameter register 59 having the size of one parameter word and the address generation circuit 60 are each constituted by one set. The description uses the names and reference numerals in FIG.

First, the arithmetic processing unit 6 that issues a transfer instruction
Node 41 including 1, 6N (hereinafter referred to as transfer source node)
It is assumed that the arithmetic processing unit 21 instructs data transfer from the shared memory 14 to the shared memory of another plurality of transfer destination nodes.

First, the transfer instruction issued by the arithmetic processing unit 61 in S41 includes a field to be a multicast instruction command between nodes (the instruction format is not shown). The memory in the system is defined in one logical address space, and the logical address is constituted by a logical node number and a logical address of the shared memory in the node. The data to be multicast is transferred to a different logical address of the shared memory in each node. The transfer control device 44 includes the arithmetic processing device 6
1 is received by the command decoding unit 17 via the local network 15. The instruction decoding unit 17
In step S42, the transfer command is analyzed, and if the received transfer command is a multicast command between nodes, an instruction is given to perform multicasting, and control information is passed to the transfer control unit 18. The transfer control unit 18 receives the control information, reads the operation parameters and data from the shared memory 14 via the local network 15 to the shared memory access unit 19 in S43, and
Instruct to transfer to. Upon receiving the instruction, the shared memory access unit 19 reads the operation parameters and data and transfers the read operation parameters and data to the data transfer unit 45. The data transfer unit 45 having received the operation parameters and the data transmits the operation parameters and data to the inter-node crossbar network 4 by the data transfer unit 26 under the control of the data transfer control unit 27 in S44.
Send to 3. The data format of the transfer data at this time is the same as in the first and second embodiments, and is shown in FIG.

In the crossbar network 13 between nodes,
The inter-node transfer control device 46 receives the operation parameters and the data from the node 41, decodes the operation parameters received in S45 by using the decoding unit 50 under the control of the data transfer control unit 54, and determines that the operation parameters are multicast targets. Is identified, and if it is identified as a target, S
At 46, the operation parameters and data are sent to the multicast control unit 53. In the multicast control unit 53, under the control of the multicast control unit 58, the operation parameters are held in the parameter holding unit 56 in S47, and it is determined whether or not the data other than the operation parameters is the data to be multicast in S48. If the data is to be multicast, the data is stored in the data holding unit 55 in S50.
To hold. The multicast control unit 53 identifies whether or not the parameter of the number of nodes to be multicast is held in the parameter holding unit 56 in S51. If all the parameters are held, the multicast control unit 53 is instructed to multicast in S52 by the instruction of the data transfer control unit 54. The parameters of all nodes and the first transfer data are read and sent to the transfer unit 52, and the transfer destination node number in the parameters is converted to the node number conversion unit 51.
Send to

Subsequently, under the control of the multicast control means 58, in S53, the logical address for storing the second transfer data to the shared memory in the first node is changed to the address in the operation parameter of the first node and the address generation information. In S54, it is determined whether or not the next address is the final transfer to the transfer destination node. If not, it is sent to the parameter holding means 56 in S55, and is stored again in the parameter buffer 59 in S47 (hereinafter, address resetting). And returns to S52 again. At the time of resetting the address, if the transfer data of the next address is identified as the last transfer data of the multicast in S54, the last transfer flag is set simultaneously with the resetting of the address in S56.

The node number conversion unit 51 converts the node numbers in the parameters from the logical node numbers to the physical node numbers for all the nodes to be multicast at the same time in S57, and sends them to the transfer unit 52. The transfer unit 52 determines, from the parameters sent from the multicast control unit 53 in S58, the information indicating that the data corresponding to the parameters is multicast data and the transfer destination node number sent from the node number conversion unit 51, The information is sent to the network control unit 49 as information, and the parameters and data are sent to the inter-node data transfer unit 48. The network control unit 49 controls the inter-node data transfer unit 48 to transfer the parameters and data corresponding to the network control information to the designated node in S59 based on the network control information sent from the inter-node transfer control device 46. Send information.

FIG. 14 shows a data format between the inter-node transfer control device 46 and the network control unit 49. The inter-node data transfer unit 48 includes a network control unit 49
In step S60, parameters and data corresponding to the network control information are sent to the inter-node transfer control device 47 corresponding to the transfer destination node (here, the node 42 will be described). At this time, parameters and data are sent to all nodes to be multicast at the same time. The first transfer data read from the data holding unit 55 is copied in the inter-node data transfer unit 48 to all the number of nodes to be multicast, and sent to the transfer destination inter-node transfer control device together with the parameters.

FIG. 7 shows a data format between the inter-node transfer control device 46 and the inter-node data transfer section 48. Upon receiving the parameters and data in S61, the inter-node transfer control device 47 (here, the transfer destination node is assumed to be the node 42) decodes the parameters received in S62 by the decoding unit under the control of the data transfer control unit. If the result is the transfer to the node 42, the parameter and the data are sent to the transfer unit, and the transfer unit sends the parameter and the data to the node 42 in S63.

FIG. 7 shows a crossbar network 43 between nodes.
4 shows a data format between the node and the node. In the node 42, upon receiving the parameter and the data, in S6
In step 4, the instruction decoding means decodes the transferred operation parameter, and when MC transfer is instructed and write to the shared memory of the own node is instructed, the transfer control unit sends the shared memory ADD in S65 to S65. Instructs data write, and in step S66, information (transfer source node number, final transfer flag) (hereinafter referred to as carry-over information) for notifying the transfer source node of the transfer completion report from the operation parameters to the data transfer control unit. To send the operation parameters and data to the shared memory access unit. In the data transfer control unit,
The portable information is stored in the shared memory access end monitoring means. The shared memory access unit is configured to convert an address for the shared memory in the operation parameter by the address conversion unit
The logical address is converted to a physical address and written to the shared memory via the local network. Shared memory is
When the data writing is completed, a shared memory access completion report indicating that the data writing to the shared memory has been completed is notified to the transfer control device via the local network. In the transfer control device, under the control of the transfer control unit, the above-described end report is decoded by the command decoding unit, and sent to the data transfer control unit. In S67, the end report is sent to the shared memory access end monitoring means. In step S68, the shared memory access end monitoring means identifies whether or not the transfer data corresponding to the end report is the last transfer data of the data transfer to the node 42 from the final transfer flag of the revolving information. If the transfer data corresponding to the end report is not the final transfer, the shared memory access monitoring unit proceeds to S6.
In S9, the process returns to S68 in order to invalidate the carrying information and continue to monitor the end report of the corresponding transfer data. If the transfer data corresponding to the above-mentioned end report is the final transfer, the portable information is similarly invalidated in S70,
The inter-node transfer control device 46 corresponding to the node indicated by the transfer source node number of the handover information is notified of the inter-node data transfer end report indicating that the data transfer to the own node is completed. At the transfer source node, the transfer control device 46
In step S71, an end report is received from the transfer destination node. The transfer control device 46 determines whether or not the one received in S72 is an end report.
Under the control of 4, the end report from the transfer destination node is decoded by the decoding unit 50, and is sent to the multicast control unit 53 in S73. Subsequently, the multicast control unit 53 sends an end report to the multicast end monitoring means 57 in S74, identifies the transfer destination node for which the end report has been made, invalidates the corresponding operation parameter, and determines in S75 whether the multicast has ended. If it is determined that the multicast has not ended, the process returns to S75 to continue monitoring the end report of the corresponding transfer data in S76. Lastly, if it is determined in S75 that all nodes have completed the end report, the end reports are received from all the nodes performing the multicast and all the corresponding operation parameters are invalidated, and it is indicated in S77 that the multicast has ended. A multicast end report is sent to the arithmetic processing unit 61 that has issued the multicast instruction via the local network 15. The arithmetic processing unit 61 identifies the end of the multicast command from the end report in S78, and ends a series of operations.

As in the first embodiment, the transfer data is written in the shared memory, and the shared memory reports the completion of the access to the shared memory and the monitoring means 94 for the completion of the shared memory access.
Notify 1. Shared memory access end monitoring means 941
Identifies the end report and sends an inter-node data transfer end report to the inter-node crossbar network 13, as in the first embodiment.

Next, in the third embodiment, when the transfer destination node of the multicast and 1: 1 transfer is the same node in the inter-node crossbar network 43 as in the first embodiment, the transfer destination Node contention occurs. The contention arbitration is performed by the network control unit 49, and the multicast is preferentially processed. In the case of contention between the first transfer data of each of the multicast and the 1: 1 transfer, in the case of contention with the 1: 1 transfer during the multicast processing, and in the case of the contention of the multicasts, the same as in the first embodiment. Is performed. Network control unit 49 when multicasting occurs during 1: 1 transfer and conflicts occur
Notifies the inter-node transfer control means corresponding to the transfer source node and the transfer destination node of the 1: 1 transfer, of an interruption instruction instructing a temporary interruption of the transfer. Upon receiving the interruption instruction, the inter-node transfer control means corresponding to the transfer source node temporarily suspends the 1: 1 transfer data transfer. The inter-node transfer control device 47 corresponding to the transfer destination node (here, it is assumed that the transfer destination node of the multicast and the 1: 1 transfer competes with 42).
Then, when the interruption instruction is received, the interruption instruction is decoded by the decoding unit under the control of the data transfer control unit, and is sent to the node 42 via the transfer unit. In the node 42, under the control of the transfer control unit in the transfer control unit, the instruction is decoded by the instruction decoding unit, and the rotation information of the shared memory access end monitoring unit in the data transfer unit is temporarily stored in the temporary save unit. Evacuate to When the data transfer by the multicast command is completed, the portable information of the temporary save unit is stored again in the shared memory access completion monitoring means. After the end of the multicast, the network control unit 49 notifies the inter-node transfer control means corresponding to the transfer source node of a restart instruction for restarting the interrupted 1: 1 transfer. The transfer source node that has received the restart instruction restarts the 1: 1 transfer.

As described above, when transferring the same data to a plurality of nodes, the arithmetic processing unit 61 that issues a transfer instruction in the node 41 issues a multicast instruction to execute the instruction issuing operation and transfer from the shared memory. Read data and transfer operation between crossbar networks between nodes.
Each can be done once. Further, data transfer from the inter-node crossbar network to the transfer destination node is performed simultaneously for all transfer destination nodes.

FIG. 15 is a time chart for the third and fourth embodiments of the present invention. Arithmetic processing unit 6
Ti is the time required from issuance of the multicast command 1 to the command identification of the transfer control device 44, Tp0 is the time required to read the parameters from the shared memory 14 and transfer them to the transfer control device via the local network 15, and the transfer control device 4
The time required from 4 to the parameter holding means 56 is Tp
1. The time for reading data from the shared memory 14 and transferring it to the transfer control device 44 via the local network 15 is Td0, the time for transferring data from the transfer control device 44 to the data holding means 55 is Td1, and the parameters and data are The time required to transfer data from the crossbar network 43 to the transfer control device of the transfer destination node is Td2, the time required for the transfer destination node to write data to the shared memory is Td3, and the time required to report completion to the transfer source node is Td2.
e (here, 2 as in the first embodiment).
It is a time chart of the multicast in which transfer is performed five times to two nodes 1a and 1b).

Next, a fourth embodiment of the present invention will be described.

FIG. 10 is a block diagram showing the third and fourth embodiments of the present invention.

The third and fourth embodiments are different only in the configuration of the parameter holding means of the multicast control unit. Therefore, in the reference numerals and names in FIG. 10, only the reference numeral of the parameter holding means is replaced with 35. The configuration of the parameter holding means is shown in FIG. Although not used for description, FIGS. 11 and 12 are also applicable.

First, the arithmetic processing unit 6 that issues a transfer instruction
The operation of the case where the arithmetic processing unit 21 instructs data transfer from the shared memory 14 of the node 41 including 1 to 6N (hereinafter referred to as a transfer source node) to the shared memory of another plurality of nodes (hereinafter referred to as transfer destination nodes) will be described. Arithmetic processing unit 6
The transfer command issued by No. 1 includes a field to be a multicast command between nodes (command format is not shown). The memory in the system is defined in one logical address space, and the logical address is constituted by a logical node number and a logical address of the shared memory in the node. The multicast data is transferred to the same address in the logical address of the shared memory in each node.

In the fourth embodiment, the same operation as in the third embodiment is performed, and the parameters are stored in the parameter holding means 56 in the multicast control unit 53. In the fourth embodiment, the parameter is stored in only one word,
The address reset is performed as follows. After transferring the parameter corresponding to the first transfer data, the parameter generation circuit generates a storage address of the second transfer data from the address generation information and stores the storage address in the parameter register again. For parameters corresponding to the second and subsequent transfer data, the address resetting is repeated in the same manner as the first transfer data. At the time of resetting the address, if the next transfer data is identified as the last transfer data of the multicast, a final transfer flag is set simultaneously with the reset of the address. Thereafter, the same operation as in the third embodiment is performed. FIG.
6 between the node 41 and the inter-node crossbar network 43, FIG. 7 between the inter-node crossbar network 43 and the node 42, FIG. 14 between the inter-node transfer control device 46 and the inter-node data transfer unit 48, and FIG. Transfer control device 4
6 and each data format of the network control unit 49 are shown. In the fourth embodiment, an end report is made as follows. As in the third embodiment, the multicast termination monitoring means 57, which has received the inter-node transfer termination report, invalidates the parameters of the parameter holding means 56 upon identifying that the termination report has been received from all nodes to be multicast. , And sends a multicast end report indicating that the multicast has ended to the node 41 via the transfer unit 52.
Subsequent steps are performed in the same manner as in the third embodiment. The contention arbitration in the inter-node crossbar network 43 is the same as in the third embodiment.

As described above, when the same data is transferred to a plurality of nodes, the arithmetic processing unit 61 that issues the transfer instruction in the node 41 issues the multicast instruction to thereby execute the instruction issuing operation and the node from the shared memory. It is possible to perform the operation of reading the parameters and the data between the crossbar networks once each. Further, data transfer to the transfer destination node via the inter-node crossbar network is performed simultaneously for all transfer destination nodes.

[0074]

As described above, according to the present invention, a data transfer command between nodes, which is a transfer command transmitted from an arithmetic processing unit in an arbitrary node to a transfer control device in the same node, includes Including a multicast command for transferring data to other nodes and including a multicast control unit in the transfer control device for receiving the multicast command and controlling the multicast to the plurality of nodes, the same memory is shared by the plurality of nodes. When transferring data, there is an effect that the load on the arithmetic processing unit can be reduced. Further, when the same data is transferred to the shared memory of a plurality of nodes, there is an effect that multicasting can be enabled for an arbitrary node.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating the principle of a multi-node information processing system according to the present invention.

FIG. 2 is a block diagram showing first and second embodiments.

FIG. 3 is a flowchart (part 1) of a multicast operation according to the first and second embodiments.

FIG. 4 is a flowchart (part 2) of a multicast operation according to the first and second embodiments.

FIG. 5 is a block diagram showing a configuration of a data holding unit and a parameter holding unit in the multicast control unit.

FIG. 6 is an explanatory diagram illustrating a data format of data transmitted from a node to a crossbar network between nodes.

FIG. 7 is an explanatory diagram for explaining a data format of data transmitted from an inter-node crossbar network to a node and a data format between an inter-node transfer control device and an inter-node data transfer unit in the third and fourth embodiments; is there.

FIG. 8 is a time chart of the first and second embodiments.

FIG. 9 is a block diagram of a parameter holding unit used in the second embodiment.

FIG. 10 is a block diagram showing third and fourth embodiments.

FIG. 11 is a flowchart (part 1) of a multicast operation according to the third and fourth embodiments.

FIG. 12 is a flowchart (part 2) of the multicast operation of the third and fourth embodiments.

FIG. 13 is a block diagram of a parameter holding unit used in the third embodiment.

FIG. 14 is an explanatory diagram illustrating a data format between an inter-node transfer control device and a network control unit.

FIG. 15 is a time chart for the third and fourth embodiments.

FIG. 16 is an explanatory diagram illustrating a data format between a node and a crossbar network between nodes according to the fourth embodiment;

FIG. 17 is a time chart showing the operation of the conventional technique.

[Explanation of symbols]

 1, 2, 11, 12, 41, 42 Node 3, 4, 16, 44 Transfer control device 5, 13, 43 Crossbar network between nodes 6, 7, 8, 20, 53 Multicast control unit 10, 40 Multi-node information processing System 14 Shared memory 15 Local network 17 Command decoding unit 18 Transfer control unit 19 Shared memory access unit 21, 54 Data transfer control unit 22 Temporary save unit 23 Address conversion unit 24 Node number conversion unit 25 Shared memory access end monitoring unit 26 Data transfer Means 27 Data transfer control means 28, 55 Data holding means 29, 35 Parameter holding means 30 Multicast end monitoring means 31 Multicast control means 32 Crossbar control units 33, 37, 60 Address generation circuit 34 Parameter buffers 36, 59 Parameters Register 45 Data transfer unit 46, 47 Inter-node transfer control device 48 Inter-node data transfer unit 49 Network control unit 50 Decryption unit 51 Node number conversion unit 52 Transfer unit 56 Parameter holding unit 57 Multicast end monitoring unit 58 Multicast control unit 61, 6N Arithmetic processing unit

Claims (8)

[Claims] A plurality of arithmetic processing units; a shared memory shared and used by the arithmetic processing units; a transfer control unit for controlling a data transfer operation in accordance with a data transfer request of the arithmetic processing units; In a multi-node information processing system, a plurality of nodes each including a shared memory and a local network interconnecting the transfer control device are interconnected by an inter-node crossbar network to transfer data between arbitrary nodes. Multicast for transferring data from its own node to a plurality of other nodes within an inter-node data transfer instruction which is a transfer instruction sent from an arithmetic processing unit in the arbitrary node to a transfer control unit in the same node And a multicast that receives the multicast command and controls multicast for a plurality of nodes. A multi-node information processing system comprising a multicast control unit. 2. The multi-node information processing system according to claim 1, wherein said multicast control unit is provided in said transfer control device. 3. The multi-node information processing system according to claim 1, wherein said multicast control unit is provided in said inter-node crossbar network. 4. An instruction decoding unit for decoding the transfer instruction sent from the arithmetic processing unit in the transfer control device, and an address conversion for converting a logical address to a physical address when accessing the shared memory. A shared memory access unit including means for accessing the shared memory; a data holding unit for holding data for single-cast and multicast transfer control sent from the shared memory access unit; and a data sent from the shared memory access unit. Parameter holding means for holding the operation parameters for the number of nodes required for transfer and generating an address, multicast end monitoring means for monitoring the end of a multicast command issued by the node to which the node belongs, and the data holding means and the parameter Holding means and the multicast end monitoring means Control means for controlling multicasting for a plurality of nodes including multicast control means for controlling the number of nodes, and node number converting means for converting a logical node number to a physical node number when performing single-cast and multicast inter-node data transfer Memory access end monitoring means for monitoring the end of access to the shared memory in the own node from the other node, transfer means for sending data to the crossbar network between the nodes, the node number conversion means, and the shared memory access A data transfer control unit that includes a data transfer control unit that controls the end monitoring unit and the transfer unit and controls data transfer between the own node and the crossbar network between the nodes; and temporarily stores information on data transfer between nodes from another node. Temporary evacuation A transfer control unit for controlling the instruction decoding unit, the shared memory access unit, the multicast unit, the data transfer control unit, and the temporary save unit, and a transfer destination in the inter-node crossbar network. 3. The multi-node information processing system according to claim 2, further comprising a crossbar control unit that controls data transfer to the node. 5. The multi-node information processing system according to claim 4, wherein said parameter holding means has a one-word parameter register for holding a parameter and a circuit for generating an address. 6. An instruction decoding unit for decoding the transfer instruction sent from the arithmetic processing unit in the transfer control device, and an address conversion for converting a logical address to a physical address when accessing the shared memory. A shared memory access unit for accessing the shared memory, the shared memory access end monitoring means for monitoring the end of access to the shared memory in the own node from another node, and data for the crossbar network between the nodes. A data transfer control unit for controlling data transfer between the own node and the crossbar network between the nodes, the data transfer control unit including: A temporary save unit that temporarily saves information on data transfer between nodes from A transfer control unit for controlling the instruction decoding unit, the shared memory access unit, the data transfer control unit, and the saving unit, and each of the nodes corresponding to each node in the crossbar network between the nodes. A plurality of inter-node transfer controllers having an interface with the other nodes, an inter-node data transfer unit for performing cross-bar connection of all inter-node transfer controllers and performing data transfer between the inter-node transfer controllers, and the inter-node transfer The multi-node information processing system according to claim 3, further comprising a control unit and a network control unit that controls the data transfer unit between nodes. 7. A decoding unit for decoding an instruction for data transfer between nodes sent from the node in the inter-node transfer control device, and a logical node for performing single-cast and multicast inter-node data transfer. A node number conversion unit that receives a number and converts it into a physical node number; a transfer unit that transfers transfer data in the inter-node transfer control device to a corresponding node or the inter-node data transfer unit; Data holding means for holding data, parameter holding means for holding parameters corresponding to the number of partner nodes required for transfer data transmitted from the corresponding node and generating addresses, and termination of multicast issued by the corresponding node. Multicast end monitoring means for monitoring, the data holding means, and the parameter holding means A multicast control unit that includes a multicast control unit that controls the multicast end monitoring unit and controls a multicast to a plurality of nodes; and controls the decryption unit, the node number conversion unit, the transfer unit, and the multicast control unit. The multi-node information processing system according to claim 6, further comprising a transfer control unit. 8. The multi-node information processing system according to claim 7, wherein said parameter holding means has a one-word parameter register for holding a parameter and a circuit for generating an address.