JP3489157B2 - Distributed shared memory system and computer - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention
Network system composed of multiple computer systems
In a computer system, physically distributed memory
The above data can be used for multiple applications running on different computers.
Related to distributed shared memory technology for sharing between
In particular, network resources are effectively used, and as a result,
Distributed shared memory that improves access performance to resident memory
About the method. 2. Description of the Related Art In a computer, an application is executed inside the computer.
A virtual entity that executes an application is called a process. Step
Process for a single CPU for a certain period of time
By switching to each process and assigning to each process, CPU
More processes than the number can be run in parallel
You. This is called parallel processing. On the other hand, using multiple CPUs
Process to operate multiple processes simultaneously in parallel.
We call it logical. Conventionally, in a computer having a single CPU configuration, the average
When performing row processing, data exchange between processes
As a rule shared memory could be used. This includes
For example, "UNIX system call programming,
Marc J. By Rochkind, translated by Toshihiro Fukusaki, ass
Key Publishing, 1987, pp. 299-315.
There is a shared memory for stem V. In parallel processing, processes are physically executed simultaneously.
Account that consists of multiple processes
No improvement in application execution performance could be expected. There
In a network computer system, multiple computers
Connect over a network and use multiple CPUs simultaneously
By processing applications in parallel,
Application execution performance. [0005] Sharing in a network computer system
In memory, each memory is
Guarantee consistency between data stored in computer memory
And as if sharing a single piece of data,
Must be shown to the application. This technology is
It is referred to as having memory. A virtual storage mechanism is realized by a paging mechanism.
The computer expands its paging mechanism and distributes
Moly can be realized. That is, usually
At the time of access, the data
If it does not exist, a page fault occurs and the auxiliary storage device
Load the page on which data exists from the memory.
In the case of a page fault, auxiliary
Instead of retrieving the page from the storage device, from another computer
Load on your computer's memory via network
It is realized by doing. [0007] Examples of the prior art of the distributed shared memory include:
The following Ivy system (Yale University, USA) and DAS
H-System (Stanford University, USA)
it can. In the Ivy system, each page has an owner
(Owner) exists and the computer that is the owner
Manages the name of the computer that holds the copy of the page. Also own
Only those who are allowed to write to the page
I have. When a read fault occurs on a computer
Will send a copy of the page from the owner. Also write
When a read fault occurs, a write fault is
The new computer becomes the new owner and the new computer becomes
Duplication of existing pages is invalidated. In a DASH system, 16 bytes of data are
The same control as the Ivy system is performed in units of data.
ing. For details of these prior arts,
The operating system that comes after UNIX
No, Maekawa Mamoru, Mario Tokoro, Kentaro Shimizu, Kyoritsu Shuppan, 19
91 ", pp. 124-141. [0011] In a distributed shared memory, a
Networked for consistency of memory content
The memory contents are exchanged between the computers. At this time,
The problem with the conventional distributed shared memory management method is data
The transfer size and the data transfer timing. [0012] The following problems are related to the Ivy system.
There is. First, regarding the data transfer size, the memory contents
The size of the data transfer between computers for exchange is limited.
Page size (for example, 4 kilobytes)
A small data (for example, 8 bytes) compared to the computer
If shared between computers, the amount of data transferred between computers
Waste. Therefore, substantial network resources
There is a problem that the utilization rate of the source decreases. The computer A and the computer B use an
Two different frequently accessed data on the same page
If there is, computer A and computer B alternately
When memory is written to a page, the page is
Move between A and computer B. Therefore, each computer is
Every time memory access is performed, calculation via the network
Page transfer between machines occurs, and memory access performance decreases.
There is a problem of going down. Second, regarding the timing of data transfer,
Data transfer is self-calculated when reading or writing to memory
When there is no page with shared data on the plane,
Was running out of order. Therefore, computer A and computer B share the same shared data.
If data is written continuously to the data,
Each time the computer writes to memory, the page moves between computers.
Move. Transfer between computers via network
The cost is compared with the memory access speed in your computer
Memory access performance to shared data
There is a problem that is lowered. On the other hand, the DASH system is an Ivy system.
Is a system aiming at solving the first problem. That is,
Caused by sharing memory contents in
Reduce sharing units to eliminate useless data transfers
Was. As a result, unnecessary data transfer has been eliminated,
Access to a large area of the
Frequent failures. Access fall
If the traffic occurs frequently, the total
The data transfer time between the computers becomes remarkable. DASH cis
System, on the other hand, reduces the transfer time for small data.
Specialized network for the purpose of
is suggesting. However, this dedicated network is a general purpose network.
Higher cost due to different network architecture
Problem. Further, the second problem of the Ivy system is as follows.
Multiple computers continuously access the same shared data
Performance degradation in the case of DASH system
It has occurred and has not been resolved. [0018] Due to the above problems, the conventional distributed shared memo
Domain-based parallelism, a typical parallel processing technique.
When applied to processing, communication via the network is frequent
And good value is obtained at the processing speedup rate by parallelization.
I can't. In addition, data with pointer variables as elements
Target, such as when sharing data structures between multiple computers
Data may be scattered and stored on the memory.
In such cases, communication via the network is frequent.
Occurs. An object of the present invention is to solve the above-mentioned problems and to provide a network.
Effective utilization of network resources and distribution with high access performance
It is to provide a shared memory system. Means for Solving the Problems First, the first problem is solved.
Data and notes to share with the application
Can be set, and the set sharing unit is stored.
Has a shared data management table for storing
Shared data manager for managing data management tables
Steps are provided. Specifically, the application is a two-dimensional arrangement.
Data such as a column is divided into two parts, and the boundary part is divided
If you want to share
Set the minute data as a sharing unit. The sharing unit is
Memory continuous area, memory discontinuous area (constant data size,
Data intervals), or a combination of these
specify. Next, in order to solve the second problem, a plurality of
When to guarantee the consistency of shared data on computers
Application from virtual shared space (1) Latest
Data acquisition instruction, (2) Up-to-date data
Decisions are made based on two pieces of information:
Use the share unit set as the unit to be verified. More specifically, shared data is transmitted to all computers.
(2) In the computer that changes the shared data
All shared data will be
Update method or share data on one computer
In a computer that tries to acquire shared data
(1) Shared data at the timing of the latest data acquisition instruction
Network with shared data from computers that hold
How to get through and further increase the reliability of shared data
To combine these two methods, one or more
The danger may be distributed by holding shared data in a computer. When the application is executed, a sharing unit is set in advance.
Setting prevents unnecessary data transfer and
Work resources can be used effectively. Also physical
Applications need to ensure the consistency of the data
By guaranteeing only when necessary,
Network resources by minimizing the transfer between
Can be used effectively. As a result, access to shared data
Performance can be improved. FIG. 1 shows a network computer according to the present invention.
1 is a block diagram illustrating one embodiment of a system 100. In this figure, the network computer system
The system 100 connects a plurality of computers 101 to a network 102
It is configured by connecting with. In the computer 101, 103
Is the data transmission path between each block constituting the computer
The system bus 104 is a process for executing a program.
And 105 are magnetic disks for storing programs and data.
Auxiliary storage device such as disk
A main memory for storing the programs and data,
107 is for controlling input and output with the network 102
Network interface. According to the invention
The shared data management method is, for example, on the computer 101 (1).
Application and application on computer 101 (2)
Application when sharing data when applications execute in parallel
Is done. FIG. 2 shows an example of the present invention for the region division type parallel processing.
One embodiment in which the distributed shared memory according to the invention is applied
FIG. Here, the region division type parallel processing is shown in FIG.
The data array as shown in (a) is shown in FIGS. 2 (c) and (d).
Split as shown, and each one in parallel on a separate computer
This is a type of parallel processing. Fluid analysis problems
For many scientific and technical calculations, such as structural analysis problems, the domain division type
Parallel processing is applicable. As an example of the region division type parallel processing, a two-dimensional
One solution of the thermal diffusion analysis problem will be described with reference to FIG. Real solution
In the analysis method, the region to be analyzed is as shown in FIG.
Array data 202 divided into a two-dimensional mesh
And stored in the main memory 106 as processor 1
04 performs arithmetic processing. At this time, the array data 202
Each element stores the temperature value of each element. The arithmetic processing is
For all the elements of the two-dimensional array in FIG.
Perform calculations using neighboring elements, and calculate them for a certain period of time for each element.
The processing to be set as a later value is determined by a predetermined end condition.
Until the condition is satisfied. Where each
The calculation using neighboring elements of the element is, for example, as shown in FIG.
Thus, when calculating the new value of array element (m, m),
The four neighboring points, that is, (m, m-1), (m + 1, m),
(M, m + 1), (m-1, m) and (m, m) itself
It is to calculate using the value. An example of this calculation is an array
For the value of element (m, m), the value of each neighboring point and (m, m)
m) and the difference between the two values
After multiplying by the coefficient, the value is added to the value of (m, m).
is there. FIGS. 2C and 2D show two analysis methods.
2 shows a region division when processing is performed in parallel on a computer. Up
As described above, in this analysis, calculation of each array element
Column using x = 0 to x = (n / 2) -2
202 (1) and x = (n / 2) +1 to x = (n−
For the column 202 (2) up to 1), the calculation of each area is performed.
It is only necessary that only the appropriate computer can read and write.
What is necessary is just to store it on the local memory 106. It
From x = (n / 2) -1 to x = (n / 2)
Columns 203 (1) and (2) are read and written from both computers.
Must be able to be stored in the shared memory area
There is. It should be noted that a typical analysis problem as described above is
The scale of the problem, that is, the number of elements in the array in FIG.
The order is from several hundred columns to several thousand rows × thousands of columns. As described below, the distributed sharing method of the present invention
In the memory method, data exchange between computers is performed as described above.
A method that can be efficiently performed even for region division type parallel processing
provide. FIG. 3 shows the execution of the same application.
Sharing when sharing data between multiple processes
FIG. 2 is a block diagram illustrating an embodiment of a data management method. In the figure, the address on the computer 101 (1) is
Application 201 (1) and computer 101 (2)
Application 201 (2) shares data
I have. Each of the components constituting the network computer system 100
Computer 101 does not have physically shared memory
Therefore, physically distributed shared data 203 (1) and 2
Hardware consistency and software consistency with 03 (2)
Must be maintained by means of
is there. By this consistency control, on the computer 101 (1)
Application 201 (1) and the computer 101
(2) The application 201 (2) on the
Can share the same data
You. In this embodiment, the consistency of shared data is maintained
A virtual space is shared data management system that exists on each computer.
Provided by stem 204. Application 20
1 performs Read / Write for shared data
Is consistent with the shared data management system 204
Performs with guaranteed virtual space. In this embodiment,
Specifically, the data stored in the shared data replication area 205
Read and write data. In this embodiment, all the computers 101
The shared data management system 204 includes an application
There is a copy 205 of the shared data 203 used in 201
I do. This copy 205 is used when the application
Generated when you declare the use of This is a
Acquire shared data 203 from application 201
This is to minimize the processing delay when a request comes. Both
In the data management system 204, the
Consistency of the shared data copy 205 is guaranteed.
In other words, the shared data management system uses
If the data copy is changed, the shared data copy 20
5 is transmitted to all the computers 101 that hold 5
You. Further, the shared data management system 204
To save the required memory, all computers
Without having a copy of the data, application 201
When an instruction to reflect the latest shared data is issued, the instruction is
A method of creating a copy on the computer where the copying was performed may be adopted.
At the same time, the knowledge of the computer that holds the copy 205 of the shared data
Notify another computer of the bespoke child. Replication of shared data
Network computer system or one or more
It may be held by a computer. FIG. 4 shows a shared data management system according to the present invention.
FIG. 3 is a block diagram showing one embodiment of the present invention. The shared data management system 204 stores the shared data
Data number identification section 401, shared data number addition section 402,
Packet transmitting / receiving unit 403, shared data area management unit 404
Consists of In addition, the auxiliary storage device 105 has a shared
It has a data management table 405. Shared data management table
Bull 405 is pre-loaded in main memory 106
By doing so, the reference and update processing can be accelerated.
it can. About details of shared data management table 405
Will be described later. The shared data number identification unit 401 is
Shared data of the packet sent from the transmitting / receiving unit 403
Identify the number. Memory area corresponding to shared data number
Is checked using the shared data area management unit 404, and
A process of writing data to a memory area to be performed is performed. The shared data number adding section 402
The shared data number and the corresponding
Memory area, read the memory area, and
Generate packets that transmit the contents of shared data to computers
Do. The packet transmitting / receiving unit 403 is a network
-The packet of the packet sent from the interface 107
Parses the command and sends it to the block that processes the packet.
Divide. Also, the shared data number adding unit 402
Network to transmit the packets to other computers
Perform processing for activating the interface 107. The shared data area management unit 404 stores the shared data
Management of data management table 405 and shared data management system
Processing for controlling the block 204 is performed. Also,
Processing request for the shared data of application 201
New entry in the shared data management table 405
Add or delete entries. In addition, the shared data
Shared data number adding unit 4 so that contents are transmitted to other computers
02 and the packet transmitting / receiving unit 405 are controlled. Conversely,
Shared data so that data contents can be received from other computers
It controls the number identification unit 401 and the packet transmission / reception unit 403.
You. This is performed in steps 401 to 404 in the figure.
Each process is software, hardware, or
Is performed by a combination of Software
Regarding the processing of the part performed by the software,
Be executed. In other words, describe the contents of each software process
The stored program is stored in the auxiliary storage device 105,
Main memory 106 connected to system bus 103
And executed by the processor 104.
You. FIG. 5 shows the arrangement when performing the area division type parallel processing.
Indicates the memory location of shared data when column data is divided
This is one embodiment. In this figure, for simplicity, 6 rows × 6 columns
Is divided into two in the column direction and two computers are used
Take parallel processing as an example. In fact, there are more elements
Array is divided into rows and columns so that more
It is assumed that parallel processing is performed by assigning to computers
However, it is possible to directly apply the contents described using this diagram.
Noh. Each element of the array is an 8-byte double-precision real number.
I do. One column of the divided boundary is shared data number "0"
Shared data 501 having a shared data number "1"
Declared as shared data 502. 511 is 6 rows x 3
2 shows an arrangement of a partial array of columns on a main memory. 5
As shown in FIG. 11, the shared data 501 is stored in the main memory.
It is arranged as a discontinuous area. On the other hand, in the present invention, the shared data is declared.
Of shared data copy created in the main memory
One embodiment is a continuous area as shown at 512.
You. Location of shared data at 511 and at 512
Shared data management table for deployment with shared data replication
405. In areas 511 and 512
To map stored data, key the shared data number
Have gone. This is supported by the application 51
1 when it is declared to use shared data
Automatically created in the data management table 405. to this
Therefore, discontinuous shared data is assigned to one identifier (shared data).
Data number) and group discontinuous data
As a result, no useless memory area is required. FIG. 6 shows a shared data management table according to the present invention.
Table 405 shows an embodiment of the present invention.
FIG. 4 is a diagram showing a table structure suitable for management. The shared data management table 405 stores the shared data
Data number entry 601, the total
Management node flag entry 60 indicating whether it is a calculator
2. Shared data address indicating start address of shared data
Entry 603, start address of duplication of shared data
Data replication address entry 604 indicating
Shared data size entry 60 indicating the size of data
5. Computer list entry holding shared data
609. Also, the shared data used by the application
Flag entry 60 indicating whether the data is a continuous area
6. The shared data is composed of several divided data
Data element number entry 607 indicating whether the
Shows how spaced the data is arranged
Hold the shared data offset entry 608. This
This allows the application to
Not only shared data but also fixed size data at fixed intervals
Can be declared as shared data
It is possible. Further, by computer list entry 609,
The shared data indicated by the shared data number.
All computers can be specified. Change shared data
Calculator list entry to notify other calculators
609 may be used. If you do not use it,
A block that communicates with all computers connected to the work
You can use the road cast function. However, requested here
Broadcast function does not cause packet loss.
If the order of packets received by each computer does not change
It is necessary to have such reliability. A copy of the shared data is held by all computers.
If not, a computer that holds a copy of the shared data
Need an item to store the identifier of. FIG. 7 shows pointer variables in a plurality of calculation periods.
Memory allocation of shared data when sharing data structures including
FIG. FIG. 7A shows the declaration of a data structure in the C language.
Here is an example of the word. In this example, the third element c is a positive number type
Is a pointer variable. Below, this structure is shared
An example according to the present invention will be described with reference to the case where the present invention is applied. The data structure shown in FIG.
7B, the memory is arranged as shown at 511 in FIG.
Is placed. In this arrangement, the element c is stored in the area 701 of the element c.
It is the address value in memory that is
The data area 702 is pointed out via a counter. data structure
When sharing the body between multiple computers, the data you want to share is
With the data stored in the elements a, b, d and the area 702
The address value itself stored in element c must be shared.
No need. This is because each computer has independent main memory 10
6 and is shared within the area 511 on each computer.
This is because the data storage addresses do not always match. On the other hand, in the present invention, the shared data is declared.
Of shared data copy created in the main memory
Is shown at 512. In 512, element in 511
c in the area 703 corresponding to the area 701 where
Is the offset value 7 to the start position of the data storage area.
04 is stored. Also specified by the offset value 704
The data length 706 is stored at the head 705 of the data storage area.
And the data 702 is stored successively. to this
More data structures with pointer variables as elements
When sharing between computers, the data to be shared is
It can be stored in a continuous area on the memory. Arrangement of Shared Data in 511 and 512
Of the arrangement with the shared data replication in the shared data pipe
Management table 405. Region 511 and
The association of the data stored in 512 is
It is performed using the existing data number as a key. This response is left
In the same manner as shown in FIG.
1 when it is declared to use shared data
Automatically created in the data management table 405. to this
Therefore, discontinuous shared data is assigned to one identifier (shared data).
Network number), and can be used for sharing between multiple computers.
The frequency of communication via the network can be reduced. FIG. 8 shows a shared data management table according to the present invention.
Table 405 shows another embodiment of the data structure 405, in particular, of the data structure.
Introduces sharing and management of shared data with hierarchical relationships
FIG. In this embodiment, the shared data shown in FIG.
For the data management table 405, the structure definition
An entry indicating the structure size of the entry and 611 is set.
I am. Shared data corresponding to shared data number 601
Is a data structure, the structure definition entry 610 contains
Stores the definition of the data structure and its shared data replication area
The data length on 512 is stored in 611. In this figure
Indicates that the shared data with the shared data number “11” has the data structure
It indicates that it is a structure. Structure definition of shared data management table 405
Of the structure definition table stored in the definition entry 610
An example is shown in FIG. Structure definition table 801
Is an entry that stores the type information of the elements of the data structure.
802, entry 80 storing the size of the element
3. Store the starting position of the element in the data structure
Entry 804, if the element is a pointer type,
The data size that stores the size of the data that it points to
-Consists of an entry 805. Further, in FIG. 8, the hierarchical element flag entry
612 and hierarchical element shared data number entry 613
By establishing a hierarchy, hierarchical
Defined in order to ensure consistency of shared data
Acquisition processing etc., including data defined as lower layers
Thus, it can be executed collectively. For example, in FIG.
In the lower layer of the shared data with the shared data number "5"
As the data, the shared data with the shared data number “6” is set.
In the lower hierarchy of the shared data number “6”.
Is not defined. In this case, for example, shared data
When the consistency assurance process of number "5" is instructed, the shared data
A similar process is executed for the shared data of the number “6”.
You. Also, instruct consistency processing for shared data number "6"
The same process only for the shared data
Execute. It should be noted that shared data having a hierarchical relationship is shared data.
By placing them continuously on the data duplication area 512.
Therefore, when executing the consistency assurance process,
Network to send data to the network.
Can improve the use efficiency of the network. The data structure including the pointer variables shown in FIG.
Structures can also be managed as a hierarchical structure. That is,
Of the data structure elements a through d into one shared unit
And the element c is
What is necessary is just to define the data part 702 which points. Also multiple
Pointer variable is defined in the data structure
For example, in the order in which they appear in the structure definition,
The existing data number is assigned to the hierarchical element shared data number entry 613.
You can register to. A data structure including pointer variables is hierarchically structured.
FIG. 9B shows another embodiment in which the structure is managed. Structure
The structure definition table 802 starts from 803 equivalent to 801
805 entries and if the element is of pointer type,
An entry that stores the shared data number of the data it points to
A bird 806 is provided. FIG. 10 shows a shared data management system according to the present invention.
FIG. 6 shows an embodiment of a packet structure used by the stem 204. FIG.
FIG. The packet 901 has a packet header 902
And data 903. Packet header 902
Are the destination computer 904, the source computer 905, and the shared
Data number 906 and shared data management command 907
Is done. For example, the shared data having the shared data number 12
When Computer 1 notifies Computer 2 of a change in data,
"2" is assigned to the destination computer 904, and 1 is assigned to the source computer 905.
12 to the shared data number 906, a shared data management command
907, the identifier indicating “shared data change notification”
The contents of the changed shared data are stored in the
Packaged on the computer that holds the data or all computers
To send a credit. FIG. 11 shows a shared data management method according to the present invention.
Provide to application 201 to apply the formula
FIG. 6 is a table showing an embodiment of an access function to shared data;
You. Character string starting with “dsm_” in this embodiment
Is an application to use a shared data management system.
Operating system functions that the application calls
Is the name. Applications access shared data
Initialization, consistency assurance, and termination
Perform First, initialization processing of 1001 and 1002
In dsm_open and dsm_open2,
Sharing units used by the application
Declare to the system. Specifically, the shared data is stored in a certain memory area.
Data numbers. In a shared data management system,
Enter the management information entry required to manage the sharing unit.
Secure. The dsm_open of 1001 is a
Declare the memory area as one shared unit. 1002
dsm_open2 indicates that data having a certain size is one.
Memory areas arranged at regular intervals are considered as one shared unit.
Declare. Next, 1003 and 1004 consistency assurance processing
To dsm_acquire, dsm_release
The application has declared as a share unit.
When performing Read / Write operation on memory area
Performs a series of Read / Write operations in 1003 ds
m_acquire and dsm_release of 1004
e, the shared data management system
Ensuring unity consistency. Finally, when the application executes the shared data (shared)
If you no longer need credits, the end of 1005
End processing dsm_close to shared data management system
Issue. With this issuance, the shared
Release the management information area in the data management system
Can be. The initialization processing, the consistency assurance processing, and the termination processing
Processing performed in the shared data management system
Will be described later in detail. FIG. 12 shows a distributed shared memory according to the present invention.
Application for performing region division type parallel processing using
FIG. 4 is a flowchart illustrating an example of a main process of the application.
You. Here, the application on each computer is
For each element of the array data shown in FIGS.
, Top, bottom, left and right adjacent elements as described with reference to FIG.
And the arithmetic processing using the value of its own element. First, in step 1101, an auxiliary storage device
From the array data stored in the application 201
The sequence data necessary for (1) is read into the main memory 106.
Put in. In step 1102, use in application
Declares shared data to be used. Specifically, the partial sequence 5
01 is assigned to the shared data number “0” and the partial array 5
02 is declared as the shared data number “1”
You. In the figure, each declaration is (open 0),
(Open 1), each of which is 100
This is performed by using the first routine. In step 1103,
In the partial sequence handled by application 201 (1)
With the shared data having the shared data number “0”,
The operation to reflect on the shared space is executed using the routine 1004.
Run. In the figure, it is abbreviated as (rel 0). The next step 1104 to step 110
8 correspond to the main loop of the application
You. Synchronize with other applications in step 1104
take. In step 1105, the virtual shared space is
Perform operations to obtain new data. here,
Source is shared data "1", destination is same shared data
Dsm_acquire processing of 1003 as “1”
Execute. This operation is abbreviated as (acq 1) in the figure.
did. In step 1106, all the elements of the partial array
Calculation processing using neighboring elements and its own element
Execute. In step 1107, the
Step 11 to reflect the result in the virtual shared space
Execute in the same way as 03. In step 1108, the calculation
It is determined whether the process has been completed or not.
Returning to step 1104, the processing up to step 1108 is repeated.
Do it back. If the processing has been completed, the process proceeds to step 1109. At step 1109, the shared data is released.
Work. Here, the shared data having the shared data number “0”
Release of shared data having data and shared data number "1"
The processing is performed using the routine 1005. In the figure,
Abbreviated as (close 0) and (close 1), respectively.
did. About application 201 (2)
Are the same as steps 1101 to 1109.
Execute the process. However, since the partial sequence in charge is different,
Step 1112, Step 1114, Step 111
6, the corresponding application 201
For shared data different from the processing steps in (1)
And perform the operation. FIG. 13 shows the initial state of the shared data shown in FIG.
When performing initialization processing, the shared data management system performs
It is a flowchart which shows one Example of a process. Processing steps enclosed by block 1201
Implements applications that have declared use of shared data.
The processing performed by the executing computer 101 is
The processing steps surrounded by the lock 1202 are shared data numbers.
Is performed by a computer that manages shared data
4 shows the processing contents. First, in step 1203, the application
Is shared with the shared data number by the shared data area management unit 404.
Specify the memory area of the shared data and declare the use of shared data.
Call 1001 dsm_open routine to say
Move. In processing step 1204, the designated shared data
Determines computer to manage shared data based on data number
I do. For example, configuring a network computer system
Assign a unique number to the computer in advance, and
Number of the computer that constitutes the network computer system
The remainder divided by the total number matches a certain predetermined number.
May be determined as the management computer. Steps
In step 1205, the shared data determined by the shared data area management unit
Use of shared data for data management computers
Notice. Computers that manage shared data must
In step 1206, the shared data declared has already been used.
Check if it is used. If not used
In step 1210, the shared data management
A new entry is created in the management table 405. Also,
Initialize the shared data holding computer list. use
If so, in step 1207, the shared
609 sharing a computer that declared to use data
Add to the data holding computer list, and in step 1208
To computers registered in the shared data retention computer list
On the other hand, the list of computer
Notice. In a computer that has declared use of shared data,
If the shared data has been declared, step 120
In step 9, the latest shared data is
Data, and use it as a copy of the shared data.
You. If not, step 1211
Secures a memory area for duplication of shared data. At step 1212, the shared data management
Create a new entry in the shared data address
Initial settings such as the address and shared data copy address are performed. FIG. 14 shows one example of the shared data shown in FIG.
When performing persistence assurance processing, the shared data management system
It is a flowchart which shows one Example of a process. Processing steps enclosed by block 1301
Is 1003 ds to get the latest shared data
Processing performed by the computer that issued m_acquire
Indicates the content. First, in step 1302, the application
To get the latest shared data for shared data in the application
, The shared data number (source) and the acquired share
Shared data number indicating the memory area where data is to be stored (move
Destination), and the shared data area management unit 404 sends dsm_
Issue an acquire. Shared data to be obtained in step 1303
Address of the shared data corresponding to the
Whether the shared data size etc. is in the shared data management table 405
Search using the shared data number as a key. Next, as in step 1303, step 1
At 304, based on the shared data number indicating the destination,
A shared data address corresponding to the shared data number,
Check the data size. In step 1305, the shared
Copy the contents of the memory area 512 of the data copy 205 to the destination
The shared data is copied to the memory area 511. When copying in step 1305,
In the division flag 606 of the shared data management table 405
When "1" is stored, the area 512 shown in FIG.
From the storage format of the area 511 to the storage format of the area 511.
This means that when copying data, the shared data
Copy at address interval of 608 minutes offset value
It is implemented by doing. Similarly, shared data management
The structure definition entry 610 in the table 405
And “1” is stored in the hierarchy element flag 612.
Also, when converting the storage format,
To 511. FIG. 15 shows one example of the shared data shown in FIG.
When performing persistence assurance processing, the shared data management system
It is a flowchart which shows one Example of a process. Processing steps enclosed by block 1401
Manages shared data for applications
1004 dsm_relea to reflect on stem
The contents of the processing performed by the computer that issued the
The processing steps surrounded by the block 1402 are the counterparts of the shared data.
Run on all computers that have shared data
This shows the processing to be performed. First, in step 1403, the application
Reflect shared data contents in shared data management system
The shared data number to specify
Issues 1004 dsm_release to the area manager
You. At step 1404, the designated shared data
Data area corresponding to data number is shared data management table
Search for the file using the shared data number as a key. Furthermore, sharing
The data number addition section reads the contents of the corresponding memory area.
However, all computers that hold the specified shared data
Create a packet to propagate changes to it. In step 1404, a memo of shared data
The contents of the storage area 511 are stored in the memory area of the shared data copy 205.
Copy to area 512. At this time, step 13 in FIG.
Processing is performed in a direction opposite to the copy processing performed in step 05.
That is, from the storage format of the area 511 shown in FIG. 5 and FIG.
While converting the storage format of the area 512,
Do pea. At step 1406, the packet transmitting / receiving unit
Is the shared data holding computer in the shared data management table.
The packet created by the shared data number addition unit based on the
To send a credit. In a computer that holds shared data,
At step 1407, the communication
Packet that notifies the change of shared data
Received by the transmitting and receiving unit. Further examine the packet contents and
Knows that it is a change notification of the existing data and recognizes the packet number.
Pass the packet to another part. At step 1408, packet number identification
The unit identifies a shared data number in the packet. Furthermore
Search the shared data management table based on the shared data number
Search for the corresponding shared memory replication address,
Check the ises and more. In step 1409, the shared data identification unit
Based on the shared data change notification arriving from the network
Write to the corresponding memory area. FIG. 16 shows the end of the shared data shown in FIG.
When the shared data management system performs
It is a flowchart which shows one Example of a process. Processing steps enclosed by block 1501
Is a process performed by the computer that requested the release of the shared data.
Processing steps enclosed in block 1502
Manages shared data determined by the shared data number
It shows the contents of processing performed by the computer. First, in step 1503, the application
The shared data number to the shared data area management unit.
1004 dsm to specify and release shared data
_Release is issued. In processing step 1504
Will share data based on the specified shared data number.
Decide which computer to manage. In step 1505,
Shared data management computer determined by the data area management unit
To release the shared data. In the computer that manages shared data, step
In step 1506, the shared data declared to be released
The shared data management data based on the shared data number.
Request release from the shared data holding computer list of the cable
Deleted calculator. In step 1507, the shared data management
Request from the shared data holding computer list of the logical table
As a result of deleting a computer with
Check if the list is empty. Computer list
If it is empty, in step 1508, the shared data
Data management table 40 in the computer managing the
5 is deleted from the entry of the shared data. Computer list
If is not empty, in step 1509 the shared data
For computers registered in the data holding computer list
Notifies the change of the shared data holding computer list. In a computer that has declared release of shared data,
In step 1510, sharing from within the shared data management table
Delete data entries and assign for shared data replication
Release the allocated memory area. The embodiments of the present invention have been described above.
However, finally, referring again to FIG.
A brief summary of the purpose, organization, and effects of the examples
You. [0108] It is composed of a plurality of computers connected via a network.
The network computer system that is created
Whether a single piece of data is shared
Application to realize distributed shared memory
About the shared data used by the
Data area, the timing of shared data consistency assurance, etc.
Get information about shared data and flow between networks
Network resources by reducing the amount of data transfer and the number of transfers
To improve the performance of accessing shared data.
Let In FIG. 4, which is a preferred embodiment of the present invention,
Is the shared data management system 204 is the shared data number
Identification unit 401, shared data number addition unit 402, packet
Consisting of a transmitting / receiving unit 403 and a shared data area management unit 404
Then, the shared data management table 405 is held. Data shared by application 201
(Shared unit) and mapping to memory, shared data area management unit
When declared for 404, the pair of shared data and memory
The response is stored in the shared data management table 405. Shared data
Order shared data changes to ensure data consistency
Computer and the computer that receives the shared data change
And take the following configuration. In the computer instructing the change of the shared data,
Shared to notify other computers of changed shared data
Check the memory area corresponding to the data number, and
Add a shared data number to the contents to create a notification packet.
And a shared data number adding unit 402 for performing the generation. The computer that receives the change of the shared data
The shared data number of the packet transmitted from the network
And identify the memory area corresponding to the shared data number.
Investigate using the data management section and store it in the corresponding memory area.
Shared data number for writing data to
A separate part 401 is provided. Packet transmission and reception between both computers
To control the network interface
A packet transmitting / receiving unit 403 is provided. According to a preferred embodiment of the present invention, application
Information about the shared data used by the
By providing a means to notify the data management system,
The amount of unnecessary data transfer and the number of transfers flowing through the
And that network resources can be used effectively
Both do not transfer useless data, so network
The transmission time is shortened and the access performance to shared data is improved.
For example, in parallel processing of the area division type,
The speed-up rate can be improved. According to the present invention, according to the present invention, network connection
Network computing composed of multiple computer systems
In the machine system
Sharing that makes it appear as if they share one piece of data
Ensure consistency of shared data when implementing memory
Network transfer volume and number of transfers required for
Therefore, network resources can be used effectively
You. In addition, since there is no needless data transfer,
Transfer time is shortened, resulting in access to shared data.
Process performance can be improved.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram of a network computer system according to an embodiment of the present invention. FIG. 2 is a diagram showing an embodiment of a region division type parallel processing which is an example of an application target of the present invention. FIG. 3 is a diagram showing one embodiment of holding shared data according to the present invention. FIG. 4 is a block diagram showing one embodiment of a shared data management system according to the present invention. FIG. 5 is an embodiment showing an application of the shared data to a region division type parallel processing according to the present invention. FIG. 6 is a diagram showing one embodiment of a shared data management table according to the present invention. FIG. 7 is an embodiment showing application of shared data to a data structure according to the present invention. FIG. 8 is a diagram showing one embodiment of a shared data management table according to the present invention. FIG. 9 is a diagram showing one embodiment of a structure definition table according to the present invention. FIG. 10 is a diagram showing one embodiment of a communication packet structure in the present invention. FIG. 11 is a diagram showing one embodiment of a shared data access function according to the present invention. FIG. 12 is a flowchart illustrating an embodiment of an application main process using shared data according to the present invention. FIG. 13 is a flowchart showing an embodiment of a shared data use start process according to the present invention. FIG. 14 is a flowchart showing an embodiment of a shared data acquisition process according to the present invention. FIG. 15 is a flowchart showing an embodiment of a shared data reflection process according to the present invention. FIG. 16 is a flowchart showing an embodiment of shared data release processing according to the present invention. [Description of Signs] 100: Network computer system, 101: Computer, 102: Network, 103: System bus,
104: processor, 105: auxiliary storage device, 106:
Main memory, 107: network interface, 201: application, 202: local data in application, 203: shared data in application, 204: shared data management system, 20
5: Shared data replication in shared data management system, 40
1: shared data number identification unit, 402: shared data number addition unit, 403: packet transmission / reception unit, 404: shared data area management unit, 405: shared data management table, 50
1: Shared data having a shared data number “0” 502:
Shared data having a shared data number "1", 511: Placement of shared data on the main memory, 512: Placement of shared data copy on the main memory, 601: Shared data number, 602-613: Shared data management table 701-706: Arrangement elements of the data structure on the shared data replication area, 801 and 802: Structure definition table, 803-807: Management items of the structure definition table,
901: communication packet, 902-907: communication packet component, 1001-1006: access function of shared data, 1301-1318: main processing step of area division type application using shared data, 120
1: a processing step in the computer requesting use of shared data; 1202: a processing step in the shared data management computer; 1203-1212: a processing step when requesting use of shared data; 1301: a process of acquiring shared data Processing steps in the computer that issued the
5: processing steps for acquiring shared data, 140
1: processing steps in the computer that issued the shared data reflection processing; 1402: processing steps in another computer sharing data; 1403-1410: processing steps when reflecting shared data; 1501: release of shared data In the computer that requested the request, 1502: processing step in the shared data management computer, 1503-151
0: Processing step when requesting release of shared data.

──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Takehisa Hayashi 1-280 Higashi Koikebo, Kokubunji-shi, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd. (72) Inventor Akira Kito 5030 Totsuka-cho, Totsuka-ku, Yokohama-shi, Kanagawa Hitachi (56) References JP-A-5-282165 (JP, A) S. See Tanenbaum, Parall Programming Using Shared Objects and Broadcasting, IEEE COMPUTER, USA, IEEE, August 1992, 10-19 Marc. J. Rochkind, UNIX System Call Programming, Japan, ASCII, November 1, 1987, p299-p315 Maekawa Mamoru, Distributed Operating System Next to UNIX, Japan, Kyoritsu Shuppan, December 25, 1991, p124 −p141 (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 12/06 G06F 15/167 G06F 15/177 G06F 9/46 G06F 12/00

Claims (1)

(57) [Claims] [Claim 1] A plurality of computers are connected by a network,
In a distributed shared memory system in which data distributed on the memory of each computer is shared by each computer, an arbitrary part of an area shared by each computer is specified as a mapping to a one-dimensional array from a computer application, The shared data management means for designating the data distributed on the memory of the computer as the one-dimensional array as a shared unit and managing the shared data in association with the specified shared unit by at least one of the plurality of computers. A distributed shared memory system, comprising: