JPH07152708A - Distributed shared memory system in loosely-coupled computer system - Google Patents

Abstract

PURPOSE:To attain synchronous control and consistency guaranteed control by the small number of network packets and to improve access performance to a distributed shared memory in a distributed shared memory system in a loosely coupled computer system. CONSTITUTION:A main memory 106 is provided with a consistency guaranteed control specifying instruction string memory 201, a shared data memory 203, a shared data consistency guaranteed control instruction string memory 204, a shared data management table 206, a priority queue 205 with a shared data queuing counter, and so on. When an instruction string in the memory 201 is executed or a packet is received from another computer, an instruction string in the memory 204 is executed. When the instruction string in the memory 204 refers to and changes table 206 and the queue 205, the consistency of the memory 203 and the synchronism of parallel programs can be controlled.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a loosely coupled computer system composed of a plurality of computers coupled by a network, in which data on physically distributed memory is being executed on another computer. The present invention relates to a distributed shared memory method for sharing by a plurality of programs, and more particularly, to a distributed shared memory method that effectively uses network resources and consequently improves the access performance to the distributed shared memory.

[0002]

2. Description of the Related Art In a shared memory in a loosely coupled computer system, the memory is physically distributed, so that the consistency between the data stored in the memory of each computer is guaranteed, and one data You need to make the program look like it's shared. This technology is called distributed shared memory.

Examples of prior art distributed shared memory include:
There is a method of specifying in the program for each unit of shared data to be accessed and transferring the data in the shared unit of the shared data. For example, a paper by AS Tanenbaum et al. "Parallel
Programming Using Shared Objects and Broadcastin
g ", IEEE, Computer, October 1992, pp. 10-19. In the Tanenbaum et al. Method, a write operation to the shared data memory provides a special instruction to the shared data memory for each sharing unit. This special write operation broadcasts the contents to be written and updates the shared data memory existing in each computer. This broadcast guarantees that all computers have the same arrival order of broadcasts issued by multiple computers by ordering the broadcasts by one computer. By using this special broadcast protocol, it is possible to guarantee consistency on each computer when writing to shared data, and simply refer to the shared data memory of each computer when reading. Good.

However, if the write operation is performed a plurality of times by the own computer before the read operation or the write operation is performed on the shared data by another computer, there is a problem that useless transfer occurs.

There is also a method of designating a section for guaranteeing consistency when accessing in units of sharing. For example, the Munin system is a system that realizes this method. M
The unin system is described in JB Carter et al. "Implementation a
nd Performance of Munin ", the 13th ACM Symposium
ium on Operating Systems Principles), October 1991,
It is disclosed on pages 152-164. In Munin, when performing a write or read operation to the shared data memory, an instruction that requests a consistency guarantee before the operation and an instruction that cancels the consistency guarantee after the operation are described in the program. The instruction that requires the consistency guarantee is called the acquire instruction, and the instruction that releases the consistency guarantee is called the release instruction. acquire
When there is a write operation between the instruction and the release instruction, the change is not reflected in the shared data existing on each computer, and the change is not reflected until the release instruction is issued. In the Munin system, paging of the virtual memory mechanism is combined, the computer that issued the acquire instruction becomes the owner of the shared data, and after the consistency guarantee is released by the release instruction, the changes corresponding to the shared data are paged. Hold. The transfer of the retained page is delayed until the next read operation is performed on the shared data memory of the other computer that has the acquire instruction, and unnecessary data transfer can be reduced. But Munin
In this method, the acquire instruction and the release instruction may overlap on multiple computers, so a synchronization means is separately provided to control the write operation of shared data between multiple computers.

[0006] On the other hand, barrier synchronization is a widely used synchronization means in a proto-coupled computer that is connected by a network and each computer has an independent address space. In order to realize barrier synchronization, there are a method of providing a dedicated signal line separately from the network and a method of realizing by network packet. The former method of providing a dedicated signal line requires special hardware and is expensive. On the other hand, the latter needs to transmit a network packet for each computer. For barrier synchronization,
"Parallel Processing Machine, Shinji Tomita, Toshinori Sueyoshi, Ohmsha, 1
989, pages 133-134.

[0007]

The problems in the prior art are as follows. In order to solve this problem, Tanenbaum et al. Transfer data in units of shared data.However, since data is always transferred between computers for write operations, write operations are continuously executed for the same data. No. 1 that wasteful transfer occurs when doing it
There are challenges.

The Munin system solves the first problem as described below. With respect to the update of the shared data memory, a section specifying means for ensuring consistency is provided in the program, and only the change at the end of the section is transferred. By doing so, even if the write operation for the shared data is performed a plurality of times in the consistency guarantee section, the data transfer to another computer occurs only once after the end of the consistency guarantee section. However, since the synchronization means is provided separately from the consistency assurance means, in the loosely coupled computer system that realizes synchronization by exchanging packets, the second problem is that the packet amount increases and the network is burdened. is there.

An object of the present invention is to effectively use network resources in a distributed shared memory system for sharing data on the memory of a physically distributed loosely coupled computer with a plurality of programs running on another computer. In order to use it, the consistency guarantee control and the synchronization control are realized with a small number of network packets, and as a result, the access performance to the distributed shared memory is improved.

[0010]

In the present invention, firstly, the first
In order to solve the above problem, a consistency guarantee means for designating the consistency guarantee of data shared by a plurality of computers in an arbitrary section is provided on the memory in each computer. By providing this means, even if there are a plurality of write operations in an arbitrary section for which consistency is to be guaranteed, it is possible to ensure consistency with only one transfer.

Next, in order to solve the second problem, when a request for shared data including a request for consistency guarantee is received from another computer within the consistency guarantee section, the contents of the same shared data are received. Have a means to store the demand. By providing this means, in the distributed shared memory on the loosely coupled computer system, it is possible to guarantee the consistency of shared data including synchronous processing without increasing the amount of packets.

Further, the number of waits for the consistency guarantee request is controlled corresponding to the means for accumulating a request for the shared data including the consistency guarantee request from the other computer in the consistency guarantee section. And a means for storing the priority of requests for shared data. Furthermore,
The consistency guarantee means is a shared data specification means for specifying shared data with an identifier and a length common to each computer, and a consistency guarantee specification means for specifying the shared data consistency guarantee in an arbitrary section in the program. And, the means for canceling the consistency guarantee is provided to cancel the specification of the consistency guarantee. By using these means, it is possible to perform finer synchronization processing at the beginning of the specified arbitrary section until the consistency of the shared data is guaranteed.

Further, the above-mentioned consistency guarantee designating means is separated into two, and means for reserving the consistency guarantee control and means for waiting for the reservation to be accepted are provided. By doing so, it is possible to execute another calculation multiple times without interrupting the program while designating the consistency guarantee control.

[0014]

According to the present invention, it is possible to reduce network packets required for synchronization for guaranteeing consistency of shared data and synchronization for controlling parallel programs, so network resources are effectively utilized. be able to. Further, since the synchronization for controlling the parallel program can be performed and the consistency of the shared data can be guaranteed at the same time, the burden on the programmer who writes the parallel program can be reduced.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an embodiment of a loosely coupled computer system 100 according to the present invention.

In this figure, a loosely coupled computer 100 is composed of a plurality of computers 101 connected by a network 102.
In the computer 101, 103 is a system bus, which is a data transfer path between blocks constituting the computer, 104 is a processor that executes programs, 103 is an auxiliary storage device such as a magnetic disk that stores programs and data, and 106 Is a main memory for storing the program and data during execution of the program, and 107 is a network interface for controlling input / output with the network 102. The distributed memory system according to the present invention is applied when data is shared, for example, when a program on the computer 101 (1) and a program on the computer 101 (2) execute in parallel.

FIG. 2 is a block diagram showing an embodiment of the main memory 106 according to the present invention.

The main memory 106 is a shared data consistency guarantee control designation instruction string memory 201, a work memory 202, a shared data memory 203, a shared data consistency guarantee control instruction string memory 204, a shared data wait counter-equipped priority queue 205. ,
It comprises a shared data management table 206.

Shared data consistency guarantee control instruction sequence memory 201 and shared data consistency guarantee control instruction sequence memory 204
The data stored in is read as an instruction to the processor 104 via the system bus 103, and the working memory 2
02, shared data memory 203, priority queue 205 with shared data waiting counter, read / write operation for data in shared data management table 206, or control command for network interface 107, and data transfer. As the instructions stored in the shared data consistency guarantee control instruction instruction memory 201, there are an instruction to specify the consistency guarantee for the shared data memory 203 and an instruction to cancel the consistency guarantee. The processor 104 reads these consistency guarantee designation instruction and consistency guarantee release instruction via the system bus 103, and controls commands to the network interface 107 via the system bus 103.
And transfer data.

Shared data consistency guarantee control instruction string memory 20
The command of 4 uses the shared data management table 205 and the priority queue 206 with a shared data waiting counter, and operates so that the data content of the shared data memory 203 matches the content of the shared data memory 203 in another computer 101. When the processor 104 reads the consistency guarantee control instruction stored in the shared data consistency guarantee control instruction sequence memory 201, the processor 104 calls the instruction in the shared data consistency guarantee control instruction sequence memory 204.

Next, the instructions in the shared data consistency guarantee control instruction sequence memory 204 are read by the processor 104, and by referring to the shared data management table 205 and the shared data waiting counter-attached priority queue 206, necessary information is obtained. The data is transferred to another computer 101 via the network interface 107 and the network 102.

When a packet is transferred from another computer 101 to the network interface 107 via the network 103, the processor 104 generates an interrupt and detects that the packet has arrived from another computer. Note that the method of detecting that the network packet has been sent from another computer 101 can also be implemented by the processor 104 periodically monitoring the network interface 107.

When the processor 104 detects that a network packet has been transferred from another computer 101, it executes the instruction in the shared data consistency guarantee control instruction sequence memory 204. The instruction in the shared data consistency guarantee control instruction string memory 204 refers to the shared data management table 205 and the priority queue 206 with the shared data waiting counter, and stores it in the shared data memory 203 via the system bus 103 if necessary. Update the shared data. The shared data management table 205 is used to maintain the consistency of the data stored in the shared data memory 203 of the own computer 101. Priority queue 206 with shared data waiting counter is another computer 1
This is a buffer that holds requests for shared data from 01.

FIG. 3 shows a shared data consistency guarantee control designation instruction sequence memory 20 stored in the main memory 106 according to the present invention.
FIG. 1 is a diagram for explaining one embodiment of a work memory 202 and a shared data memory 203 in more detail.

The instruction string 301 is a one written in the language C so that the instruction string including the consistency guarantee control instruction stored in the shared data consistency guarantee control instruction instruction memory 201 is activated by the computer application. This is an example. The instruction sequence 301 is composed of an instruction 302 for generating shared data, an instruction 303 for specifying consistency guarantee, an instruction 304 for updating shared data, and an instruction 305 for canceling consistency guarantee. When writing a parallel program, the programmer
The generation and update of shared data, the consistency guarantee and its cancellation are specified in consideration of the application priority of each computer.

The instruction 302 is composed of a number "2" indicating arbitrary shared data and a portion for designating the size of the shared data. In this example, the element size is integer type int and size 1
00 array data is generated, and a pointer to the array data area is stored in table. When this instruction is executed, the shared data 320 indicated by the table is generated in the shared data memory 203 of each computer. There are a plurality of shared data in the shared data memory 203.

The instruction 303 is a consistency guarantee specification instruction for specifying the consistency guarantee of the shared data 320 generated by the shared data generation instruction 302. In this command 303, a shared data number for designating shared data for which consistency is to be guaranteed, the number of waiting requests for shared data, and a request priority for shared data are designated.

The command 304 is a command for updating the shared data for which the consistency guarantee is designated in the command 303.

The command 305 is a consistency guarantee cancellation command for canceling the consistency guarantee of the number "2" indicating the shared data 320.

It should be noted that in the instruction 304, the shared data 320 has a plurality of update operations. Command 30
For multiple update operations on the shared data 320 in 4, the shared data 320 on each computer 101 up to the consistency guarantee cancellation instruction 305.
The changes to the can be postponed to reduce the amount of data transferred over the network 108.

The commands 302, 303 and 305 issue control commands to the network interface 107. Network interface that received this control command
The shared data consistency assurance unit in 107 registers and changes the shared data in the shared data management table, and if necessary, the shared data in the other computer via the network interface 107.
Controlling consistency of 20. The work memory 202 stores variables temporarily used in the own computer, for example, the variable “i” 310 used in the instruction sequence 301.

FIG. 4 shows an example of the shared data management table 205 and the priority queue 206 with the shared data waiting counter.

The shared data management table 205 is shared data 3
Shared data number 401 and shared data start address 4 for each 20
02, shared data size 403, owner computer number 404, reserved flag 405, busy flag 406, shared data waiting queue number 408. The shared data waiting queue number 408 is
One of a plurality of priority queues 206 with a shared data waiting counter is shown. Each priority queue 206 with a shared data waiting counter comprises a request waiting counter 407 and a request waiting computer list 410. The request waiting computer list 410 is composed of a request waiting computer 411 and a priority queue having a set of request waiting priority 412 as an entry. The shared data number 401 is a number for identifying a plurality of shared data existing in the shared data memory 203, and is a number common to each computer 101 of the loosely coupled computer system 100. The position and size of the area of the shared data 320 are designated by the shared data start address 402 and the shared data size 403. The owner computer number 404 indicates the computer having the latest shared data indicated by the shared data number 401. When the owner computer number 404 and the own computer number 404 are the same, the owner of the shared data is the own computer. The reserved flag 405 indicates that the consistency guarantee designation command 303 has already been executed. By checking this reserved flag, an error occurs if multiple consistency guarantee requests are issued for the same shared data. Flag in use
Reference numeral 406 is a flag that becomes “1” after the consistency guarantee designation command 303 becomes the owner computer and before the consistency guarantee cancellation command is issued. Shared data waiting queue number
Reference numeral 408 denotes an entry in the priority queue 206 with a shared data waiting counter, which is valid only when the owner computer 404 is its own computer.

Priority queue with shared data wait counter
206 includes a request waiting counter 407 and a request waiting list 410. The request waiting counter 407 is changed by the addition processing with the request waiting number designated by the consistency guarantee designating unit 303, and the request waiting list 410 is controlled using this counter. The request waiting list 410 holds each entry of the request waiting computer 411 and the request waiting priority 412, and for the shared data which is the owner of the own computer, the consistency guarantee designation instruction 320 from another computer A priority queue that stores requests. As for the priority of the request waiting list 410 in this embodiment, the larger the value, the higher the priority. A detailed description of the embodiment using the priority queue 206 with the shared data waiting counter is shown in FIGS.
It will be described later using 2.

FIG. 5 shows an embodiment of the network packet 501 transmitted / received by the network interface 107 via the network 102 according to the present invention.

The network packet 501 is composed of a packet header 502 and data 503. The consistency guarantee packet header 501 includes a destination computer 504, a source computer 505, a shared data number 506, a shared data management command 507, and shared data management information 508. In order to control the consistency of the shared data 320 corresponding to the shared data number 506 from the source computer 505, the shared data management command 507 and shared data management information 508 to the destination computer 504 are transmitted.

FIG. 6 is a diagram for explaining the shared data management command 507 and shared data management information 508 in the packet header 502 of FIG. 5 in more detail.

There are two shared data management commands 507 of this embodiment: a consistency guarantee reservation command “ACQUIRE” 601 and a shared data use permission command “ACQUIRE_ACK” 602.
Corresponding to each shared data management command 507, shared data management information 508 follows.

The consistency guarantee reservation command “ACQUIRE” 601 corresponds to the waiting counter setting number 611 and the request waiting priority 612, and the shared data use permission command “ACQUIRE_ACK” 602 corresponds to the change owner computer 621 and the request waiting computer list 622. To do. When the shared data management command 507 is the consistency guarantee reservation command “ACQUIRE” 601, the data 503 is empty.

Shared data use permission command “ACQUIRE_AC
In the “K” 602, the latest data held by the owner computer is packed in the data 503.

All the above shared data management commands are broadcast to all computers. This broadcast is premised on that all computers have the same arrival order of broadcasts transmitted from each computer. In order to broadcast, for example, the value of the destination computer 504 of the packet header is set to "-1",
In the network interface 107, it may be decided in advance to broadcast when transmitting a packet to the destination computer “−1”.

FIG. 7 is a timing chart showing the operation when a consistency guarantee control command is issued by a plurality of computers 101 in one embodiment of the present invention.

Sections 701, 703, and 705 indicate that the owner computer owns the shared data with the shared data number “2”, and sections 702 and 704 indicate the consistency guarantee designation instruction 304 (1). Indicates the section in which the program that issued is suspended. The instruction in the shared data consistency guarantee control instruction string memory 204 of the computer 101 (i), when the processor consistency guarantee designation instruction 304 (1) is designated, executes the consistency guarantee reservation command “ACQUIRE” 601 (1). Broadcast to all computers. It is assumed that the computer 101 (j) that is the owner at that time has already issued the consistency guarantee cancellation instruction 305.
Computer 101 (j) is a consistency guarantee reservation command "ACQUIRE" 601
When (1) is received, the consistency guarantee permission command "ACQUIRE_A
Broadcast "CK" 602 (1) to all computers. The program of the computer 101 (i) is suspended until the consistency guarantee permission command “ACQUIRE_ACK” 602 (1) is received. Calculator 101
When (i) receives the consistency guarantee permission command “ACQUIRE_ACK” 602 (1), the owner computer changes from computer 101 (j) to computer 101 (j).
Change to (i). When another computer such as the computer 101 (k) receives the consistency guarantee permission command “ACQUIRE_ACK” 602 (1), the owner computer changes from the computer 101 (j) to the computer 101 (i) in the shared data management table. To reflect. The consistency guarantee reservation command "ACQUIRE" 601 (2) generated by the consistency guarantee specification command 304 (2) from the computer 101 (k) before the consistency guarantee cancellation command 305 (2) is executed on the computer 101 (i). ) Reaches the owner computer 101 (i), the consistency guarantee permission command “ACQUIRE_ACK” 602 (2) is not transmitted until the consistency guarantee cancellation instruction 305 (1) is executed. Calculator (k)
Waits for the consistency guarantee permission command “ACQUIRE_ACK” 602 (2), the synchronization for the consistency guarantee can be realized.

FIG. 8 is a PAD diagram showing an instruction sequence stored in the shared data consistency guarantee control instruction sequence memory 204 when the network packet 501 is received from the network interface 107.

First, by the packet header analysis processing in step 801, the network packet 501 is converted into each destination computer 504, source computer 505, shared data number 506, shared data management command 507, shared data management information 508, and data 503. Disassemble. Next, in step 802, the contents of the shared data management command 507 are checked, and the consistency guarantee permission command “ACQUIRE_ACK” or the consistency guarantee reservation command “ACQUI
"RE". If it is the consistency assurance permission command “ACQUIRE_ACK” command, the process proceeds to step 803. In step 803, the changed owner computer 621 in the received packet checks whether it is the own computer, and if so, the busy flag 406 is set to "1" and the owner computer 404 is set as the own computer as shown in step 804. If the changed owner computer 621 in the received packet must be the own computer, the owner computer 40 in the shared management table 205
Change 4 to owner calculator 621.

In step 802, the shared data management command 507
Is the consistency guarantee reservation command “ACQUIRE” 601, the owner computer 404 and the own computer of the shared management table 205 are checked as shown in step 3, and if it is the owner computer, the “ACQUIRE” reception processing of step 807 is executed. To do.

FIG. 9 is a PAD diagram for explaining the "ACQUIRE" reception process 804 in detail. First, in step 901, in addition to the value of the request waiting counter 407 in the priority queue 206 with the shared data waiting counter specified by the shared data number 506 specified in the received packet and the waiting counter setting number 902 in the received packet, The new value is set as the new request waiting counter 407 value.

Next, at step 902, it is checked whether the busy flag 406 in the shared management table 205 is "1" and the request waiting counter 407 is "0". If the busy flag 406 is not “1” or the request waiting counter is not “0”, the request waiting priority queue of the shared data management table 205 is according to the request waiting priority 612 in the received packet as in step 903. Add to 206. If the busy flag 406 is “1” and the request waiting counter 407 is “0”, the shared data waiting queue number 408 is set as in step 904.
From the priority queue with shared data wait counter, and if there is a wait, in step 905, according to the request wait priority 612 in the received packet, the shared data management table 2
In addition to the request waiting priority queue 206 of 05, the head of the shared data waiting queue number is taken out in step 906 to be the next owner computer. If you don't wait in step 904,
The sender computer 505 of the received packet is the next owner computer. By the processing of step 905 and step 906, it is possible to preferentially accept the consistency guarantee reservation command “ACQUIRE” that arrived first.

In step 904, if there is no waiting in the shared data waiting queue number, as shown in step 907, the sender computer of the received packet is the next owner computer. After checking the shared data waiting queue number in step 904, the next owner computer has been decided,
For this owner calculator, go to step 908, "ACQUIRE_AC
"K" transmission processing, CHANGE_OWNER transmission processing in step 909 is executed.

In step 903, it becomes possible to order requests from a plurality of computers by priority. If it is desired to perform processing in the order in which the requests are issued without ordering, all the priorities from the plurality of computers should be set equal.

FIG. 10 shows "ACQUIRE_A" in step 908.
FIG. 8 is a PAD diagram showing details of “CK” transmission processing.

In the "ACQUIRE_ACK" transmission process, first, in step 1001, the owner computer column 404 of the shared data management table 205 is set as the next owner computer.
Next, in step 1002, the owner computer 404 of the shared data management table 205 is the destination computer 504 of the packet header, the own computer is the source computer 505 of the packet header, the shared data number 401 is the shared data number 506 of the packet header, and “AC
QUIRE_ACK ”command 602 to share data management command 50
7. The request waiting computer list 206 of the queue 206 with shared data waiting counter corresponding to the shared data number 506 is copied to the request waiting computer list 622 of the packet header. Next, in step 1003, data of the data size is copied from the start address 402 of the shared data management table 205 to the data part of the packet. Finally, in step 1004, the network packet is transferred to the network interface 107 and is instructed to be sent to the destination computer 504.

The CHANGE_OWNER transmission process at step 909
The task of setting the shared data waiting queue number 206 of step 1002 in the “ACQUIRE_ACK” transmission process to the shared data waiting queue number 622 of the packet header is changed to the next owner computer, the change owner computer 630, and step 1003 is omitted.

FIG. 11 is a timing chart showing the operation when the number of request wait settings in the consistency guarantee designation command 302 is designated in one embodiment of the present invention.

Sections 1101, 1102, 1105, and 1106 indicate that the owner computer owns the shared data with the shared data number “2”, and sections 1103 and 1105 specify the consistency guarantee designation instruction 304. Indicates the section where the issued program is suspended. In the section 1101, it is assumed that the consistency cancellation command has already been issued. In the consistency guarantee designation command 304 (1) in the computer 101 (i), the request waiting number "1" is added to the shared data number "2", and the request is designated to have the priority "0". The computer 101 (i) broadcasts the consistency guarantee reservation command “ACQUIRE” 601 (1) and transmits the request to the computer 101 (j) which is the owner computer.

The request waiting counter 407 of the queue 206 with the shared data waiting counter corresponding to the shared data number "2" of the computer 101 (j) is 0 at this time, and the consistency guarantee designation instruction 304 (1) The designated request wait setting number “1” is added, and the value of the request wait counter 407 becomes “1”. When the request waiting counter 407 is other than "0", the consistency permission command "ACQUIRE_ACK" is not returned. Where computer 101
From (k), a consistency guarantee reservation command “ACQUIRE” 601 (2) is broadcast by the next consistency guarantee designation command 304 (2). At this time, the request waiting setting number is designated as "-1" and the priority is designated as "1". Here, "-1" is added to the value of the request waiting counter 407 of the queue 206 with the shared data waiting counter corresponding to the shared data number "2" of the computer (j), and the request waiting counter 407 becomes "0". . Also,
In the above-mentioned consistency guarantee reservation command 601 (1), the request waiting priority queue 206 stores the request of the priority “0” from the computer (i), and the computer 101 (k) of the priority “1”. The consistency guarantee permission command “ACQUIRE_ACK” 603 (1) is broadcast with the computer 101 (k) having a higher priority as the owner computer in comparison with the request from the computer 101 (k). k). The request from the computer 101 (i) is issued to the owner computer (k) with the consistency guarantee permission command "ACQUI
RE_ACK ”603 (1) is transmitted together as shared data management information. The request of the computer 101 (i) is delayed until the consistency release instruction 305 (1) is executed by the computer 101 (k).

In this way, in the consistency guarantee specification instruction 304, by simultaneously specifying the number of request waiting settings and the priority, not only the synchronization required for the consistency guarantee but also the entire parallel program is controlled. Synchronization can also be realized.

FIG. 12 shows the consistency guarantee designation instruction shown in FIG.
FIG. 9 is a diagram showing a state of a queue 206 with a shared data waiting counter corresponding to a shared data number “2” in the computer 101 (i) when the request waiting set number is designated by 302.

Reference numerals 1201, 1202 and 1203 denote priority queues with a shared data waiting counter at a certain time in the computer 101 (j).
The state of 206 is shown. 1201 is the section 110 in FIG.
11, 1202 indicates the section 1102 and 1203 in FIG. 11 immediately after the section 1202 in FIG. 11, and the consistency guarantee permission command “ACQUIRE_AC
The state immediately before the transmission of "K" 602 (1) is shown. Network packet source computer fields 504 (1), 504
(2) shows "i" and "k", which means that the commands are issued from the computers 101 (i) and 101 (k), respectively. Request waiting counters at 1201, 1202, 1203 407 (1), 407
(2) and 407 (3) are added by the waiting counter setting numbers 611 (1) and 611 (2) of the consistency guarantee reservation commands 601 (1) and 601 (2). As can be seen in 1203, the priority 612 (2) of the request from the computer 101 (k) that arrived later was the computer 101 (i) that arrived earlier.
Request priority list is higher than 612 (1)
The beginning of 410 (3) is replaced. Next, the request waiting counter 40
Since 7 (3) is inspected to be "0", the request from the computer 101 (k) at the head of the request waiting list 410 (3) is executed first.

FIG. 13 is a diagram showing the operation of the embodiment in which the consistency guarantee designation instruction 304 is separated into the consistency guarantee reservation instruction 1301 and the consistency guarantee permission wait instruction 1302.

When the computer 101 (i) is the owner computer which owns the shared data of the object number “2”,
When the consistency guarantee reservation command 1301 is executed in 101 (k), the consistency guarantee reservation command "ACQUIRE" 602 is displayed on the computer.
sent to (i). When the consistency guarantee designation command 304 is executed, the program on the computer (k) is kept waiting until the consistency guarantee permission command 602 is received from the computer (i).
However, the consistency guarantee specification instruction 304 is replaced with the consistency guarantee reservation instruction 1
By dividing into 301 and the consistency guarantee permission wait instruction 1302, it is possible to separately perform a calculation that does not access shared data between the consistency guarantee reservation instruction and the 1301 and consistency guarantee permission wait instruction 1302. These instructions 1301 and 1302 can be realized without changing the shared data management table 205, the configuration of the shared data waiting counter-attached priority queue 206, and the shared data consistency guarantee control instruction memory 204.

[0062]

According to the present invention, a plurality of computers are connected by a network, each computer has an independent address space, and a shared data memory for sharing data among a plurality of computers on a memory in each computer. In a loosely-coupled computer system with, it is possible to reduce the number of network packets required for synchronization to guarantee consistency of shared data and synchronization for controlling parallel programs, so network resources can be effectively utilized. can do. Also,
Since the synchronization for controlling the parallel program can be performed and the consistency of the shared data can be guaranteed, the burden on the programmer who writes the parallel program can be reduced. As a result, in particular, it becomes possible to efficiently deal with the problem of parallel processing in which the calculation time is longer than the communication time and it is difficult to automatically distribute the load.

[Brief description of drawings]

FIG. 1 is a block diagram of a loosely coupled computer system showing an embodiment of the present invention.

FIG. 2 is a block diagram of a main memory and a network interface showing an embodiment of the present invention.

FIG. 3 is a diagram showing an embodiment of a structure of a main memory according to the present invention.

FIG. 4 is a diagram showing an example of a structure of a shared data management table according to the present invention.

FIG. 5 is a diagram showing an embodiment of a communication packet structure according to the present invention.

FIG. 6 is a diagram showing an embodiment of a communication packet structure according to the present invention.

FIG. 7 is a timing chart showing an embodiment of the processing of the shared data consistency guarantee control instruction string memory in the present invention.

FIG. 8 is a PAD diagram showing an example of processing of a shared data consistency guarantee control instruction sequence memory in the present invention.

FIG. 9 is a PAD diagram showing an example of processing of a shared data consistency guarantee control instruction string memory in the present invention.

FIG. 10 is a PAD diagram showing an embodiment of processing of a shared data consistency guarantee control instruction sequence memory in the present invention.

FIG. 11 is a timing chart showing an example of processing of a shared data consistency guarantee control instruction string memory according to the present invention.

FIG. 12 is a diagram showing an embodiment of a shared data management table and a request waiting counter in the present invention.

FIG. 13 is a timing chart showing an embodiment of processing of a shared data consistency guarantee control instruction string memory in the present invention.

[Explanation of symbols]

100: loosely coupled computer system, 101: computer, 102: network, 103: system bus, 104: processor, 10
5: Auxiliary storage device, 106: Main memory, 107: Network adapter, 108: External network, 201: Shared data consistency guarantee control instruction command memory, 202: Working memory, 203: Shared data memory, 204: Shared Data consistency guarantee control instruction memory, 205: Shared data management table, 206: Priority queue with shared data wait counter, 3
01: Shared data consistency guarantee control command command sequence including commands, 302: Shared data generation command, 303: Shared data consistency guarantee designation command, 304: Command sequence for updating shared data, 30
5: Shared data consistency guarantee cancellation command, 310: Data in working memory, 320: Shared data in both shared data areas, 4
01 to 406,408: Shared data management table entry, 40
7: Request waiting counter, 410: Request waiting computer list, 41
1 to 412: Request waiting computer list entry, 501: Network packet, 502: Packet header, 503: Data in network packet, 504 to 508: Packet header entry, 601: Consistency guarantee reservation command, 602: Consistency Guarantee permission command, 611 to 612: Consistency guarantee reservation command entry, 621 to 622: Consistency guarantee permission command entry, 701,703,705: Section that is the owner of the shared data of shared data number "2", 702,704: Consistent 801 to 807, 901 to 908: processing steps when a network packet is received, 901 to 908: processing steps for issuing a consistency guarantee permission command,
1101,1102,1105,1106: Owner section of shared data of shared data number "2", 1103,1104: Section waiting for permission of consistency guarantee, 1201 to 1203: Shared data wait counter and priority with shared data wait counter Queue, 1301: Consistency guarantee reservation command, 1302: Consistency guarantee permission wait command.

Front page continuation (72) Inventor Satoshi Yoshizawa 1-280 Higashi Koikekubo, Kokubunji, Tokyo Inside Hitachi Central Research Laboratory (72) Inventor Hideki Murayama 1-280 Higashi Koikeku, Kokubunji City, Tokyo Inside Hitachi Central Research Center (72) Inventor Kominori Yamada 5030 Totsuka-cho, Totsuka-ku, Yokohama, Kanagawa Prefecture Software Development Division, Hitachi Ltd. (72) Toru Horimoto 5030 Totsuka-cho, Totsuka-ku, Yokohama, Kanagawa Hitachi Software Co., Ltd. Within the development headquarters

Claims (9)

[Claims] 1. A plurality of computers are loosely coupled in a network,
In a distributed shared memory system in a loosely coupled computer system in which each computer has an independent address space and shares data distributed in each computer's memory, consistency of shared data distributed in each computer's memory in an arbitrary section At least one of the plurality of computers is provided with a consistency guarantee means for receiving a request for the shared data including a request for guarantee and a means for accumulating a request for the shared data from another computer. Distributed shared memory method for loosely coupled computer systems. 2. The means for accumulating a request for the shared data from the other computer according to claim 1, and a counter means for counting the number of waits for the request for the consistency guarantee of the shared data, and the shared data. A distributed shared memory system in a loosely coupled computer system characterized by having a means for storing the priority of requests. 3. The consistency guarantee means according to claim 1, wherein the shared data designating means for designating the shared data with an identifier and a length common to each computer, and the consistency guarantee of the shared data are programmed. The consistency guarantee specification means for specifying in the arbitrary section of and the consistency guarantee cancellation means for canceling the specification of the consistency guarantee are provided, and at the beginning of the specified section until the consistency of the shared data is guaranteed. A distributed shared memory method for loosely coupled computer systems characterized by synchronization. 4. The method according to claim 2, wherein when the value of the counter means is other than a specific value, a request for guaranteeing consistency of the shared data and a request for the shared data from the other computer are stored. A distributed shared memory system in a loosely coupled computer system characterized by being stored in a means. 5. The means for accumulating a request for the shared data from the other computer according to claim 2, wherein the order of holding the request for guaranteeing consistency of the shared data is changed according to the priority. Distributed Shared Memory System for Loosely Coupled Computer Systems. 6. The method according to claim 3, wherein the consistency guarantee designation means counts the number of waiting requests for the consistency guarantee of the shared data provided in the means for storing the request for the shared data from the other computer. A distributed shared memory system in a loosely coupled computer system, characterized in that it comprises means for setting a value to be added to the counter means for performing the addition. 7. A distributed shared memory system in a loosely coupled computer system according to claim 3, wherein said consistency guarantee designating means comprises means for designating a priority of a request for consistency guarantee. 8. The loose coupling according to claim 3, wherein the consistency guarantee designating means comprises means for reserving the designation of the consistency guarantee and means for waiting for the reservation to be accepted. Distributed shared memory method for computer systems. 9. A plurality of computers are loosely coupled in a network,
In a distributed shared memory system in a loosely coupled computer system in which each computer has an independent address space and shares data distributed in each computer's memory, the shared data distributed in each computer's memory is a common identifier for each computer. And shared data specifying means to specify by length,
It is provided with a consistency guarantee specification means for specifying the consistency guarantee of the shared data in an arbitrary section in the program and a consistency guarantee cancellation means for canceling the specification of the consistency guarantee, and accepts the request for the shared data. Consistency assurance means, counter means for counting the number of waiting requests for consistency guarantee of the shared data from other computers, and means for accumulating the priority of the request for the shared data are provided. A priority queue with a counter that stores requests for shared data, and at least an identifier, start address, and
And a shared data management table having information on the length of shared data and storing information for associating with the priority queue with a counter, and for a request accepted by the consistency assurance means, Loose coupling characterized by controlling the access of the shared data in the memory in the own computer or the transmission / reception of the shared data in the memory of another computer by using the priority queue with the counter and the shared data management table Distributed shared memory method for computer systems.