JPH10143486A - Data transmission/reception method in parallel computers - Google Patents

Abstract

PROBLEM TO BE SOLVED: To transfer data at a high speed by selecting whether to perform transfer by using an area to which an address is allocated for respective processor units(PUs) beforehand or to perform the transfer by a pipeline processing by using the area dynamically allocated at the time of the transfer depending on the length of the data. SOLUTION: In the inside of remote memory write areas 114 and 115, static buffers 116 and 117 to which the address is allocated beforehand for the respective communicating party PUs and dynamic buffers 118 and 119 for which the address is dynamically changed are present. In the case that the data length of transmission data is shorter than a predetermined value, the transmission data are written in the static buffer area of a transmission destination. Also, in the case that the data length of the transmission data is longer than the predetermined value, the address of the dynamic buffer area of the transmission destination is received by using the static buffer area, the transmission data are written in the dynamic buffer area of the received address at the transmission destination and the data are transferred by the pipeline processing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for transmitting and receiving data between PUs in a parallel computer in which a plurality of element computers (processing units, hereinafter referred to as PUs) are connected by a communication network, and more particularly to a high-speed message passing and data security. Related to a data transmission / reception method to be secured.

[0002]

2. Description of the Related Art A parallel computer improves processing speed by connecting a plurality of PUs by a communication network and operating them simultaneously. In particular, the present invention is directed to a distributed memory type parallel computer in which each PU can access only a memory space associated therewith. In a distributed memory parallel computer, data on the memory of another PU cannot be directly accessed. Each time data is required, it is necessary to perform transmission and reception and move the data to its own PU.

In a distributed memory type parallel computer, all data exchange between PUs must be described in a program. Here, data passed between PUs is called a message. Exchanging this message is called message passing. In the parallel computer program, other PUs
If the data required by
The PU sends these data to other PUs in advance, and
If your PU needs data on the memory of the
It is necessary for the PU to explicitly write an instruction to receive these data from another PU in advance in the program of each PU.

In many parallel computer systems, a group of functions (or subroutines) called a message passing library is prepared in advance in order to support transmission and reception of messages between such PUs, and communication is performed by C or FORTRAN. It can be described as a function call from a program. Some message passing libraries have been implemented on different parallel computer hardware to provide a de facto standard communication environment. Oak RidgeNational in the United States
Examples are the PVM developed at the Laboratory and the MPI, which is being standardized in recent years. A parallel program written by calling these communication libraries has a high possibility (portability) that it can be operated only by recompilation on a different parallel computer.

In order to transfer a message between PUs in the communication library, a sending PU calls a message sending function, a receiving PU calls a corresponding message receiving function, and a message is exchanged between these. A transmission / reception method is currently generally used. If the receive function is called before the send function, the receive function blocks (stops) until data arrives; if the send function is called before, it blocks until the start of the receive function or the message Buffered within. This is send / rec
It is called eive method.

As a method for realizing data transfer between PUs at high speed, there is a distributed memory type parallel computer having a remote memory writing mechanism. In remote memory writing, each
The PU can directly transfer data to a specific memory area in the partner PU without intervention of the partner PU. The specific memory area in which remote memory writing can be performed is called a remote memory writing area. Since the real address is continuous in the remote memory write area and is not swapped, each PU can transfer data at any time.

FIG. 2 shows a conventional method for implementing message passing on a parallel computer having a remote memory writing mechanism. The actual data transfer between each PU is performed between the remote memory write areas. First, when the transmission function is called, data is copied from the buffer (201) in the user program on the transmission side to the buffer (202) in the remote memory write area. Next, from there, a buffer (203) in the remote memory write area of the receiving side
Is transferred using remote memory writing. Finally, the reception function is called and the buffer in the user program is written from the remote memory write area (203).
The message is copied to (204), and the delivery of the message is completed.

[0008]

As described above, when data is transferred without the intervention of the partner PU in remote memory writing, the receiving side needs to confirm the use of the remote memory writing area on the receiving side. There is a risk of overwriting a large memory area with write data and destroying data on the receiving side. Therefore, the data transfer cannot be performed continuously without checking the remote memory writing area on the receiving side. Also, data transfer from the user program buffer to the remote memory writing area on the sending side, data transfer from the remote memory area on the sending side to the remote memory area on the receiving side, and data transfer from the remote memory writing area on the receiving side to the And a total of three data transfers were required. Also, the order between data separately sent to the remote memory write area cannot be guaranteed.

It is an object of the present invention to perform remote memory transfer at high speed with a limited amount of transfer memory and to guarantee the order of data sent independently by a partner PU.

[0010]

According to the present invention, there is provided a parallel computer comprising a plurality of computers and a communication path interconnecting them, wherein each computer has a static area which is fixed for each computer as a communication partner. Securing a dynamic reception buffer area and a dynamic reception buffer area dynamically allocated when a communication occurs; and, when the data length of the transmission data is shorter than a predetermined value, the static buffer of the transmission destination. Writing the transmission data in the area, and if the data length of the transmission data is longer than a predetermined value, receiving the address of the dynamic buffer area of the transmission destination using the static buffer area, This is achieved by providing a step of writing the transmission data to the dynamic buffer area of the received address at the destination.

When transferring data longer than a predetermined value, data can be transferred at a high speed with a limited amount of memory by transferring the data by pipeline processing.

Further, the order of the transmitted data can be guaranteed by sequentially connecting the used transfer memories (buffers).

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The details of the present invention will be described below with reference to the drawings.

First, a specific example of the mounting method of the present invention will be described with reference to the drawings. FIG. 1 shows an overall configuration diagram of the present invention. 1
Reference numerals 01 and 102 denote PUs (processor units).
Reference numerals 3 and 104 denote CPUs and 105 and 106 are memories. Reference numeral 107 denotes a communication path (a network that can interconnect the PUs) connecting them. Although the number of PUs is arbitrary, a parallel computer including two PUs is shown for the sake of explanation. Reference numerals 108 and 109 are OSs (operating systems). When executing the user program, the user program is first loaded into the memory from an auxiliary storage device or the like (not shown in FIG. 1) (110, 111). It is assumed that the user program is linked in advance to the message passing library (112, 113) of the present invention. The message passing library is provided with remote memory write areas (114, 115) to which remote memory can be written from other PUs. Further, in the remote memory writing area, there are a static buffer (116, 117) to which an address is assigned in advance for each communication partner PU and a dynamic buffer (118, 119) whose address dynamically changes.

[0015] Of the above components, the message passing library is a component that characterizes the present invention.
The details will be described below.

(A) Configuration of Buffer A buffer configuration for message passing in the present invention will be described with reference to FIGS. As described above, in the present invention, the static buffer (301) and the dynamic buffer (301)
Two types of buffers (401) are used. The static buffer 301 corresponds to the static buffers 116 and 117 in FIG. 1, and the dynamic buffer 401 is the dynamic buffer 118 and the dynamic buffer 118 in FIG.
119.

First, the static buffer stores the communication partner PU (# 0,
A plurality of blocks (FIG. 3)
In FIG. 6, six blocks are shown for PU # 1). Each block of the static buffer is roughly divided into a header (302) and a message body (303).
(305), first address (306), last address (30
7) is included. tag is an identifier for selecting the corresponding send / receive pair, length is the length of the message to be communicated, first address is the first address when a dynamic buffer is allocated, and last address is the last address when a dynamic buffer is allocated. Is an area for storing. The identification information of the destination PU and the source PU in the communication path (network) is separately managed, and the message is transferred in the network. Each static buffer is
Buffers of the same shape for transmission (301) and reception (309) are prepared for each PU, and information such as the use status of the buffer is shared with the communication partner PU. The static buffer is set in advance for each communication partner PU, so that each PU can know the address at the time of initialization. The transmitting side can always transfer data to the receiving buffer on the receiving side as long as the transmitting buffer has a free space.

The dynamic buffer is composed of a plurality of blocks (401) that are not distinguished for each communication partner PU. Each block is divided into a header (402) and a message body (403), and the header is used by the receiving side to confirm whether data has arrived. The configuration of the area of the header portion is the same as the configuration of the static buffer, and the destination P
It is assumed that identification information (address) of the U and the transmission source PU in the network is separately managed. In order to further select a buffer amount according to the message length, each block is managed as a bundle of several blocks (this bundle is indicated by a thick line frame in the plurality of blocks 401 in FIG. 4). Each bundle is registered in the management header (404), and the receiving PU acquires a buffer bundle having an optimal size according to the message length (in the example of FIG. 4, the bundle of one block and the bundle of four blocks are respectively When a message is smaller than the size of one block, a bundle of one block is obtained, and when the message is larger, a bundle of four blocks is obtained.) The receiving PU writes the first address and the last address of the acquired buffer bundle in the first addres
It is stored in s, last address and transmitted to the transmitting PU. Also, the sending side has prepared two dynamic buffers
(405) By using these alternately, pipeline processing becomes possible. Details of the pipeline processing will be described later.

(B) Communication Protocol In general, when the message length is short, it is required that the message be transferred at a higher speed (low latency). On the other hand, when the message length is long, a larger amount of messages are transferred per unit time. (High throughput) is required. In order to satisfy these two incompatible requirements, the present invention provides two communication protocols, a short protocol used for transmitting a message with a short message length and a long protocol used for transmitting a message with a long message length. However, by switching between them, data transfer with less delay is realized.

FIG. 5 shows a timing chart of the short protocol. The short protocol uses a static buffer for message transfer. When the transmission function is called by the user program (501), the transmission side PU transmits the header of the static buffer and the message body to the reception side (503). On the other hand, when the receiving function is called (50
2), the receiving PU receives the message in the static buffer, and transmits a reception completion notification to the transmitting side (504). Message communication can be completed by one round trip data communication.

However, since the remote memory write area is limited, the length and quantity of the static buffer are limited. Therefore, if all messages are transmitted in a static buffer and a large number of messages are to be communicated, it is necessary to wait for the static buffer to become empty, and conversely the communication speed decreases. Therefore, when the message length is long, message transfer is performed using a dynamic buffer that is not distinguished for each PU and therefore can be prepared in large quantities. This is a long protocol. Using a short protocol to send a message with a message length smaller than the static buffer capacity, and sending a message with a message length larger than the static buffer capacity, using the length of the message body part of the static buffer as a boundary Is controlled to use the long protocol.

FIG. 6 shows a timing chart of the long protocol. When the transmission function is called by the user program (601), the transmitting PU first detects the length of the data length to be transmitted (the data amount of the data to be transmitted), and this data amount is larger than the capacity of the static buffer. When,
It is determined that the long protocol is used. If the amount of data to be transmitted is smaller than the capacity of the static buffer, the short protocol described above is used. In the case of the long protocol, the header of the static buffer is transmitted to the receiving side (603). On the other hand, when the receiving function is called (602), the receiving PU receives the information of the message in the static buffer, and sends back the address information of the dynamic buffer according to the information using the static buffer (604). The PU transmits a message based on the buffer information received (605), and finally, the receiving PU transmits a reception completion notification to the transmitting side (606). Although two round trips of data transfer are required and the latency is higher than that of the short protocol, a dynamic buffer can transfer a larger amount of data than a static buffer, so that the throughput can be increased.

Hereinafter, the operation of each protocol will be described in detail. First, the operation of the short protocol will be described with reference to FIG. The transmitting PU stores the message length and the identifier in the header of the static buffer (701). Furthermore, the message is copied from the user area to the message body (70
2). Next, the remote side writes the static buffer to the remote side from the transmitting side (703). On the other hand, the receiving PU
Receives the message in a static buffer (705), copies it from there to the user area, and completes the reception (70
6). Finally, the receiving side sends out a reception completion notification using a static buffer (707), and upon receiving the notification, the transmitting side also ends the processing (704).

Next, the operation of the long protocol will be described with reference to FIG. First, the transmitting PU stores the message length and the identifier in the header of the static buffer as in the short protocol (801). In the case of the long protocol, since the message cannot be sent with the static buffer, only the header is written in the remote memory to the receiving side (802). When the receiving PU receives the information of the message in the static buffer (807),
A dynamic buffer of an appropriate length is secured according to the length, and its start address and end address are obtained (808). Those addresses into the static buffer's first address, last
It is set in the address and sent back to the sending side (809). The transmitting side receives the buffer information (803), repeats the loop until all the messages are transmitted (804), copies the dynamic buffer block as a unit from the user area to the buffer, and transmits it (805). Further, the receiving side repeats the loop until all the messages are received (810),
The message is copied from the received block to the user area (811). Finally, the receiving side sends out the reception completion notification using the static buffer (812), and upon receiving the notification, the transmitting side also ends the processing (806).

(C) Pipeline transfer in long protocol In message passing using remote memory writing, copying from the user area to the remote memory writing area, remote message transfer from the sending side to the receiving side, Copying to the area and transferring the message are required three times. Therefore, at least about three times as long as remote memory writing is required. Therefore, in the present invention, performance is improved by preparing two dynamic buffers on the transmission side and performing pipeline processing.

The operation during pipeline processing will be described below with reference to FIG. In FIG. 9, the buffer (204) in the remote memory writing area of the transmitting PU is different from the two dynamic buffers (405) of the transmitting side in FIG. 4 by the buffer (203) in the remote memory writing area of the receiving PU. Is
The dynamic buffer (401) on the receiving side is represented in a simplified manner. Although the dynamic buffer on the receiving side is prepared in multiple planes, only four planes necessary for transferring ABCD data are shown here. In step 1, the message A is copied from the user area to the remote memory write area on the transmitting side (901). Next, in step 2, message A is transmitted from the transmitting side to the receiving side by remote message transfer (9).
02) At the same time, the message B is copied to the remote memory write area (903). In step 3, the receiving side copies the message A from the remote memory writing area to the user area (904), transfers the message B from the transmitting side to the receiving side by remote memory (905), and transmits the message C from the user area to the transmitting side. Copy to the remote memory write area (906). In step 4, the message B is copied on the receiving side (907), the transfer of the message C from the sending side to the remote side is written to the remote memory (908), and the message D is copied on the sending side (909).

FIG. 10 shows the operation of each of the transmitting PU and the receiving PU in the pipeline processing. On the transmitting side, data A
Is copied from the user area to the remote memory writing area (1001), and then the data A is transmitted to the receiving side and the data B is copied at the same time as the memory B (1002). Finally, data D is transmitted (1005). On the other hand, on the receiving side, first, data A is received from the transmitting side (1006), and then, upon receiving data B, memory A of data A is copied at the same time (1007).
009) Finally, memory copy of data D is performed (1
010).

(D) Interface of Message Passing Library The message passing library is a group of functions for performing message passing in the form of a function call from a user program. Hereinafter, the interface of the function of the message passing library and the operation thereof according to the present invention will be described. The function name, the argument name, and the like are arbitrary, and do not necessarily have to be the same as the specifications described here.

(1) Init () In this function, the message passing library performs necessary initialization operations. When using the message passing library, all PUs must call this function first. When this function is called from the user program, a static buffer is
(FIG. 1: 116, 117) and the dynamic buffer (FIG. 1: 118,
119) is created, and each buffer is initialized. The address of the static buffer is fixed for each PU, and all PUs can communicate the destination address at the time of transmission of the static buffer at the time of this initialization. The following functions can be used only after calling the initialization function.

(2) Send (buf, dest, tag, length) Here, buf is the head address of the user memory storing the message to be transmitted, and dest is the destination PU number (identifying the destination PU in the network. Information), tag represents the identifier of the message, and length represents the length of the message. When the user calls this function, the library sends a message using the short protocol (701 to 704 in FIG. 7) or the long protocol (801 to 806 in FIG. 8) according to the length.

(3) Recv (buf, src, tag) buf of this function is the start address of the user memory for storing the received message, src is the source PU number (information for identifying the PU in the network), tag Represents a message identifier. When the user calls this function, the source PU
The length of the message is received in the header part of the static buffer with the matching number, and the short protocol is adjusted accordingly.
(FIG. 7: 705 to 707) or the message is received by the long protocol (FIG. 8: 807 to 812).

(E) Guarantee of order in non-blocking operation For transmission and reception in the message passing library,
There are blocking functions and non-blocking functions. The blocking function is a function that blocks (stops) the operation of the program from the time the transmission / reception function is called until transmission / reception is completed.The non-blocking function returns after the function is called and before transmission / reception is completed. It is a function that can perform other operations in the meantime. Until the previous section, a blocking function was assumed. In this section, an additional mechanism for realizing non-blocking will be described. In a non-blocking function, a plurality of transmission / reception functions may be issued before transmission / reception is completed. At this time, in the transmission / reception functions having the same tag, the transmission function and the reception function may not correspond in the order of issuance. The method of guaranteeing the order of the transmission / reception functions of the present invention will be described below with reference to FIG.

In the message passing library of the present invention, at the time of transmission, first, message information (tag, length) is set in a header of a static buffer, and the message is notified to the receiving side.
(In short protocol, the body of the message is also sent at the same time.) At this time, the next (11
01), and the next specifies the block of the static buffer to be transmitted next. Each block of the static buffer in use is chained by next, and transmitted in the order of the chain. The static buffer received by the receiver is from the sender,
Like the sender, they are connected by next. Therefore, the search can be performed in the order of the chain, and the issue order of the transmission functions can be determined.

When the transmission side issues a transmission function when the static buffer is not empty, or when the reception side issues the reception function, the header of the static buffer has not been sent yet. It cannot be expressed in the order of the blocks in the static buffer. Therefore, in the present invention, in order to maintain the order of functions that cannot be processed immediately when the function is issued,
A request object for managing the order of issuing unprocessed functions is introduced (1102, 1103). The request object has a tag that holds the information of the message
(1104, 1105), dest (src) (1106, 110
7), length (1108, 1109) and next to retain the order
It has a total of four elements (1110, 1111). Unfinished functions are chained together by next, just like the blocks in the static buffer, and processed by the sender and the receiver in that order. After the blocks of the static buffer are processed in order, the processing proceeds in the order of the request object. With the above method, the order in the non-blocking function is guaranteed.

(F) Additional Interface for Non-blocking Operation If the order is guaranteed, a non-blocking function can be realized by adding the following interface.

(1) Isend (buf, dest, tag, length) Each argument has the same specification as Send. When the user calls this function, if there is free space in the static buffer, connect the buffer to be used to the chain and transfer it. If there is no free space in the static buffer, create a request object and add this function to the chain. connect. The request object chains are processed in the order of issue. After transferring the static buffer to the receiving side, the short protocol (FIG. 7: 70
1 to 704) or the long protocol (FIG. 8: 801 to 80)
Data is transferred asynchronously in 6).

(2) Irecv (buf, src, tag) Each argument has the same specification as Recv. When a user calls a function, a request object is created and connected to the chain. Processing is performed in order from the function at the top of the chain, and a search is performed to determine whether a transmission function having a corresponding tag has been issued from the top of the chain of the static buffer. After the corresponding transmission has been retrieved, the short protocol (704: FIG. 7)
707) or the long protocol (Fig. 8: 807-812)
The data is transferred asynchronously by.

(3) Wait () This function is a function for waiting for completion of the non-blocking function. Since the non-blocking function returns immediately after the function call, the user does not know when the function completes. Therefore, this function is used to indicate the completion of the non-blocking function. When this function is issued, the PU is blocked until the non-blocking function that has not been confirmed to complete is completed. This function is completed when all functions are completed. After this function is completed,
Send / Isend and Recv / Irecv are issued and communication is resumed.

[0039]

According to the remote memory transfer control method of the present invention, in data transfer using remote memory writing, data is transferred using an area to which an address is previously assigned to each PU depending on the length of data. It is possible to select whether or not to perform transfer by pipeline processing using a region that is sometimes dynamically allocated, thereby enabling high-speed data transfer. As shown in FIG. 12, when the pipeline operation is introduced, three message transfers occur simultaneously except for the first and last two times. Therefore, about three times the performance can be obtained as compared with the conventional message passing using remote memory writing.

Further, according to the remote memory transfer control method of the present invention, it is possible to guarantee that data is transferred in the order in which the transmission functions are issued, and that the transferred data is received in the order in which the reception functions are issued. As a result, the order of data in the transmission / reception function that performs the non-blocking operation can be guaranteed.

[Brief description of the drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

FIG. 2 shows a conventional data transfer control method using remote memory writing.

FIG. 3 is an explanatory diagram of a static buffer.

FIG. 4 is an explanatory diagram of a dynamic buffer.

FIG. 5 is a timing chart of a short protocol.

FIG. 6 is a timing chart of a long protocol.

FIG. 7 is a flowchart of a short protocol.

FIG. 8 is a flowchart of a long protocol.

FIG. 9 is an explanatory diagram of a pipeline operation.

FIG. 10 is a flowchart of a pipeline operation.

FIG. 11 is an explanatory diagram for guaranteeing order.

FIG. 12 is an operation diagram of a pipeline operation.

[Explanation of symbols]

101,102 ... element calculator, 103,104 ... CPU,
105,106 ... memory, 107 ... communication path, 108,1
09 ... Operating system, 110,111 ...
User programs, 112, 113 ... message communication library, 114, 115 ... data transfer memory area,
116, 117, 301, 309 ... Static buffer, 11
8, 119, 401, 405, 910, 911 ... dynamic buffer, 201, 204 ... user buffer, 202, 20
3. Data transfer buffer, 302 ... Static buffer header, 303, 308 ... Static buffer message body, 304 ... Message identifier storage area, 3.
05: an area for storing the length of the message, 306 ... an area for storing the head address of the secured dynamic buffer, 3
07: an area for storing the final address of the secured dynamic buffer; 402: a header for confirming whether a message has arrived in the dynamic buffer; 403: a message body of the dynamic buffer; 404... Dynamic buffer management header, 1101... Area storing static buffer order, 1102, 1103.
1104, 1105... Request object area for storing message identifiers, 1106, 1107... Request object areas for storing message transmission / reception partners, 1108, 1109. Area of 1110,11
11 ... Area for storing the order of request objects.

Claims (4)

[Claims] 1. A method of transferring data between computers, comprising a plurality of computers and a communication path interconnecting the computers, wherein each computer has an area fixed for each computer as a communication partner. Securing a dynamic reception buffer area that is dynamically allocated when a communication occurs, and a step of securing a dynamic reception buffer area when the data length of the transmission data is shorter than a predetermined value. Writing the transmission data to the static buffer area; and when the data length of the transmission data is longer than a predetermined value, receiving the address of the dynamic buffer area of the transmission destination using the static buffer area. And a step of writing the transmission data in the dynamic buffer area of the received address at the transmission destination in the parallel computer. . 2. A plurality of element computers are connected by a communication path, and data in a data transfer memory area on an arbitrary element computer is written to a data transfer memory area on an arbitrary element computer. A data transmission / reception method between any two elemental computers in a parallel computer having a remote memory writing mechanism, wherein each elemental computer has a data transfer memory area in response to a communication initialization request from a user program. Securing a static reception buffer area in which an address is fixed in advance for each computer to be communicated with and a dynamic reception buffer area from which a buffer is dynamically allocated each time a communication occurs; In response to a transmission request from the program, the transmission-side element computer stores the transmission data in the user memory in the data transfer memo. Copy to the area, configure a header including information about the data length, if the data length of the transmission data is shorter than a predetermined value, the header and the data prepared on the receiving side element computer If the data length of the transmission data is longer than a predetermined value, the header is written in the static buffer area prepared on the element computer on the receiving side. Writing the data using the remote memory writing mechanism, and the receiving-side element computer, upon the arrival of the header,
The data length is obtained by referring to the header, and when the data length of the transmission data is longer than a predetermined value,
Securing a buffer requiring the data length in the dynamic buffer area and notifying the address information of the buffer to the element computer on the transmission side; and the element computer on the transmission side determines whether the data length of the transmission data is If the length is longer than the predetermined value, the arrival of the address information triggers the data referring to the address information to the secured dynamic buffer of the receiving element computer using the remote memory writing mechanism. A writing step; and, upon a reception request from the user program, the receiving element computer copying the data written in the static buffer area or the dynamic buffer area to a reception buffer on the user memory. A data transmission / reception method comprising: 3. When a plurality of transmission requests are issued from a user program, a step of storing pointer information to a header to be secured for the next transmission request in the header information. Data transmission / reception method described. 4. A parallel computer in which a plurality of element computers are connected by a communication path and which has a remote memory writing mechanism for writing data on an arbitrary element computer to a memory of another arbitrary element computer. A method for transmitting and receiving data between element computers, wherein each element computer includes, on the data transfer memory area, a dynamic reception buffer area from which a buffer is dynamically allocated whenever communication occurs, and two dynamic reception buffer areas. Allocate a transmission buffer. Configuring the dynamic reception buffer area to be composed of a plurality of blocks of a predetermined fixed length, and the transmitting element computer transmits the transmission data in response to a transmission request from a user program. Dividing the packet into a plurality of packets of a predetermined fixed length, forming a header including information on the length of data, and notifying the header to an element computer on the receiving side; Arriving at a timing, referring to the header information, securing a required number of reception blocks of the blocks in the dynamic reception buffer area, and notifying the transmission side element computer of the address information of the buffers. And transmitting the n-th packet of the plurality of packets to the data transfer device when the address information of the buffer arrives. And writing the (n-1) -th packet already copied to the other of the two transmission buffers on the data transfer memory area to the remote memory. Writing to the (n-1) th block of the dynamic reception buffer on the element computer on the receiving side by using the mechanism, and sequentially applying the above two steps to all of the plurality of packets; Triggering the receiving request to cause the receiving element computer to sequentially copy the data written in the plurality of blocks of the receiving buffer to the receiving buffer on the user memory.