JP2003036179A - Virtual shared memory, virtual shared storing method and virtual shared storing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the overhead of a virtual shared memory and a burden on a user. SOLUTION: A virtual shared memory 1 extracts a reading side thread program and a writing side thread program from a program for performing reading operation and writing operation and performs synchronization control between the reading side thread program and the writing side thread program according to a relation between a reading side thread and a writing side thread such as a dependency relation, based on reading operation and writing operation with respect to the writing side thread and the reading side thread to be executed on the same or mutually-different processors 2A and 2B.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used in various parallel / distributed computers used by connecting to a network, and a memory distributed between computers (hereinafter referred to as "distributed memory") is used as a shared memory on software. Virtualize,
In order to efficiently perform write and read operations to such virtualized shared memory (hereinafter referred to as "virtual shared memory"), virtual shared memory that performs multiple parallel processing using the principle of multithread processing. , A virtual shared storage method and a virtual shared storage program executed by a computer.

[0002]

2. Description of the Related Art Shared memory distributed between computers (hereinafter,
It is called "shared distributed memory". ) Management, if write and read operations of virtual shared memory are performed in parallel, it is necessary to give a guarantee to the read side that accurate data has already been written. Synchronous control between the two is required.

Regarding such synchronous control between a write operation and a read operation, conventionally, a user program explicitly issues a read operation command on the assumption that data has already been written in a memory to be referenced. .

In response to the issued read operation command, it is checked by the computer hardware or the operating system (OS) whether or not the necessary data is written in the memory to be read, and the write operation is still being performed. If not, it is waiting for the read operation.

[0005]

Due to such waiting, a lock operation and a semaphore operation for controlling exclusive access to the shared memory, an operation for maintaining the consistency of contents among the distributed memories, and the like. In, an overhead occurs in memory management. Also, write operations and /
Or, if the timing of the read operation is incorrect, the time and / or number of waits on the read side increases, which increases the overhead.

In the user program describing parallel distributed processing, the synchronization timing of the write operation and the read operation is predicted and described in advance, but if the read target data is not written contrary to the prediction, the read target data is written. Waiting occurs until issuing and waiting for a read operation that cannot be performed.

In the user program, it is difficult to accurately predict the synchronization between the write operation and the read operation in order to suppress the issuance of the read operation that cannot be executed.
Designing a program in consideration of synchronization imposes a considerable load on the user.

An object of the present invention is to provide a virtual shared memory, a virtual shared storage method and a virtual shared storage program executed by a computer, which can reduce the overhead.

Another object of the present invention is to provide a virtual shared memory, a virtual shared storage method, and a virtual shared storage program executed by a computer that reduce the burden on the user.

[0010]

A virtual shared memory according to the present invention is the same as or different from a thread program extracting means for extracting a read side thread program and a write side thread program from a program for performing a read operation and a write operation. The read side thread program and the write side thread according to the relationship between the read side thread and the read side thread based on the read operation and the write operation. And a synchronization control means for performing synchronization control with the thread program.

According to the present invention, the read side thread program and the write side thread program are extracted from the program such as the user program which performs the read operation and the write operation, and according to the relation between the write side thread and the read side thread, Since synchronization control is performed between the reading side thread program and the writing side thread program,
A read operation can be performed with the data already written.

As a result, there is almost no need to wait for a write operation, and the overhead is reduced. Further, since the timing of the write operation and / or the read operation can be appropriately controlled, the waiting time and / or the number of times of waiting can be significantly reduced, and the overhead can be reduced.

Further, since the language processing system can appropriately predict the synchronization of the write operation and the read operation, it is not necessary to design the program in consideration of the synchronization, and the burden on the user is reduced.

Preferably, the read operation and the write operation are treated as a write event and a read event, respectively. In addition, the read thread program and the write thread program are respectively converted by the language processing system from the existing program code and then converted into a format suitable for the read operation and the write operation in the same or different processors. The read thread program and the write thread program.

The virtual shared storage method according to the present invention includes a step of extracting a read side thread program and a write side thread program from a program for performing a read operation and a write operation, and a write executed on a processor which is the same as or different from each other. Synchronization control between the read-side thread program and the write-side thread program according to the relationship between the write-side thread and the read-side thread based on the read operation and the write operation for the side thread and the read-side thread And a step of performing.

According to the present invention, it is possible to reduce the overhead and the burden on the user when performing virtual shared storage.

The virtual shared storage program executed by the computer according to the present invention is the same as or different from the step of extracting the read side thread program and the write side thread program from the program for performing the read operation and the write operation. The read-side thread program and the write-side thread program according to the read operation and the relationship between the write-side thread and the read-side thread based on the write operation. And a step of performing synchronous control between the two.

According to the present invention, it is possible to reduce the overhead and the burden on the user when the virtual shared storage is executed by the computer.

[0019]

Embodiments of a virtual shared memory, a virtual shared storage method, and a virtual shared storage program executed by a computer according to the present invention will be described in detail with reference to the drawings. In the first embodiment described with reference to FIGS. 1 to 4, a mechanism for writing and reading to and from the shared distributed memory between the same or different processors is realized by hardware, and the virtual shared memory is read. The read operation and the write operation are described as a thread program as shown in FIG. In addition, such a thread program is converted from an existing program code by a language processing system.

Further, the program shown in FIG. 1 is automatically converted by the program loader into a thread program for writing and reading to and from the distributed real memory as shown in FIG. 3 at the time of program startup including process startup. It The program shown in FIG. 2 is executed by the hardware device shown in FIG.

The virtual shared memory 1 extracts a read side thread program and a write side thread program from a program for performing a read operation and a write operation and, as will be described later, on the same or different processors 2A and 2B. A relationship between the write thread and the read thread based on the read operation and the write operation based on the write operation, and
For example, synchronization control is performed between the read side thread program and the write side thread program according to the dependency relationship.

The processors 2A and 2B have arithmetic units 3A and 3B, data transfer units 4A and 4B, and distributed memories 5A and 5B, respectively. Prior to execution, the write-side thread program is loaded into the processor 2A and is already loaded into the processor 2B when the program is loaded or the process is started (see FIG. 2). At the same time, conversion tables between the virtual shared memory 1 to be written and read and the distributed memories 2A and 2B are created in the transfer devices 4A and 4B, respectively.

The writing side thread program is executed by the thread execution unit 6A of the arithmetic unit 3A. If a write command (write) is executed during program execution, data (Data
a) is written in the distributed memory 5A.

Here, a case where the read side threads are executed by the same processor, that is, the processor 2A will be described. When the writing side thread program executes the thread execution end instruction, the execution end instruction sends an activation signal to the reading side thread designated as the continuing thread, that is, the thread 2 of the process 2 shown in FIGS. 1 and 2. Issue. In this operation, as shown in FIG. 4, the reading side thread is registered in the synchronization waiting thread management table 7A, and the synchronization counter of the reading side thread is decremented by 1 at the same time as the registration.

The reading thread which received the activation signal,
That is, thread 2 of process 2 shown in FIGS. 1 and 2.
Is managed in the synchronization waiting management table 7A, and when all the activation signals due to other events arrive and the synchronization counter of the synchronization waiting management table 7A becomes zero, and the activation synchronization condition is satisfied, the execution waiting thread queue 8A Can be put in.
At the same time, this thread is deleted from the synchronization wait management table 7A. Then, this thread is taken out of the execution waiting thread queue 8A and executed by the thread execution device 6A. The thread execution device 6A uses the data (D-D) from the address (Local-Phys-Address-A: XA in FIG. 2) of the distributed memory 5A when executing a read instruction (read) during thread execution.
read ata).

Next, a case where the reading thread is executed by another processor, that is, the processor 2B will be described. In the writing side thread program, the data sending instruction (send) is executed after the writing is completed. The address of the distributed memory 5A holding the data to be transferred by the data transmission command (Local-Phys-Address-A: X in FIG. 2).
A), the destination processor (Destination: processor 2B in this case), and the name of the thread on the reading side to be activated (Thread-
ID) is transmitted to the data transfer device 4A. In addition, Thread-ID
Contains the processor name and process name as well.
And in FIG. 2, thread 1 of process 3 corresponds to this.

The data transfer device 4A reads the data (Data) acquired from the distributed memory 5A by the data reading device 9 from the address (Local-Phys-Address-A: XA in FIG. 2) of the distributed memory 5A, and the data transmission device 4A. 10 transfers to the transfer destination processor (Destination: processor 2B in this case).

At this time, the data transfer device 4A uses the address (Local-Phys-Address-A: XA in FIG. 2) of the distributed memory 5A.
To the address of virtual shared memory 1 (Logical-Mem-Addres
s: The address is converted to X in Fig. 2 and the address of the virtual shared memory 1 (Logical-Mem-Addre
ss: X in Figure 2) and the read thread name (Thread
-ID) is transmitted by the data transmission device 10.

The data transfer device 4B uses the transferred data.
(Data), address of virtual shared memory 1 (Logical-Mem-Add
ress: X in Fig. 2) and the name of the read thread to be activated (Thre
When the data receiving device 11 receives the ad-ID, the data writing device 12 causes the address (Log
ical-Mem-Address: The address conversion from X in FIG. 2 to the address of the distributed memory 5B (Local-Phys-Address-B) is performed, and the transferred data is transferred to the address of the distributed memory 5B (Local-Phys-Address). -Write to B). When the writing is completed, a start signal is issued. In this operation, the thread synchronization management device 13 is activated, and the synchronization counter of the thread on the reading side designated by Thread-ID in the synchronization waiting thread management table 7B is decremented by 1.

The reading side thread which received the activation signal,
That is, the thread 1 in the process 3 is managed in the synchronization waiting management table 7B, and all the activation signals due to other events arrive and the synchronization counter of the synchronization waiting management table 7B becomes zero, so that the activation synchronization condition is satisfied. Then, it is put into the waiting thread queue 8B. At the same time, this thread is deleted from the synchronization wait management table 7B. Then, this thread is taken out from the execution waiting thread queue 8A and executed by the thread execution device 6B. The thread execution device 6B uses a read instruction (read) during thread execution.
Address of the distributed memory 5B (Local-Phys-
Address-B: Data is read from XB in FIG.

Next, a second embodiment will be described with reference to FIGS. 1, 2, 5, and 6. In the second embodiment,
This is a software implementation of a mechanism for performing writing and reading, and other event processing and thread execution (internal operation processing) as parallel operations.

This embodiment also presupposes the execution of the read thread program and the write thread program of FIG. 2 loaded prior to execution. In addition, the software is a commercially available hardware device 2 as shown in FIG.
It is assumed that the hardware device 21
Includes an arithmetic unit 22, a memory unit 23, a data transmission / reception unit (interruption unit) 24, and a bus 25 for transferring data between them.

The operation of this embodiment will be described mainly with reference to FIG. As described in the first embodiment, the thread execution management unit 26 runs in the arithmetic unit 22 that executes the writing side thread. At the same time, when the transfer data is received from another processor, it is treated like any other externally generated event.

When a write command (write) is executed during execution of the write side thread program, the data (Data) is transferred to the address (Local-Phys-Address-A: Figure) of the distributed memory 27 specified by the operand of the write command. In 2, write to XA).

When the execution of the thread execution end instruction ends the execution of this thread, the read side thread designated by the operand of the thread execution end instruction is registered as the continuation thread.

Here, the case where the reading thread is executed in the same processor will be described. In the continuation thread registration, the reading side thread is registered in the synchronization waiting thread management table 28, and at the same time, the synchronization counter in the synchronization waiting management table 28 of the reading side thread is decremented by 1.

When the start-up synchronization condition is satisfied by the arrival of all start-up signals due to other events to the thread on the reading side and the synchronization counter becoming zero, the signals are put into the execution waiting queue 29. At the same time, this thread is deleted from the synchronization wait management table. Execution then begins when this thread is dequeued.
When a read command (read) is executed during the execution of this thread, the address (Local-Phys-Address-
A: Access XA) in Figure 2.

Next, the case where the reading thread is executed in another processor will be described. The data transmission processing unit of the thread execution management unit 26 performs address conversion to an address (Local-Mem-Address: X in FIG. 2) of the virtual shared memory, and drives a data transfer device (not shown) to transfer data.

At this time, the data processing unit of the thread execution management unit 26 has the address (Local-Phys-Address-A: XA in FIG. 2) of the distributed memory 27 holding the data to be transferred,
Destination Processor (Destination: Processor 2 in FIG. 2)
B) and the thread name (Thread-ID) on the reading side to be activated are transmitted to the data transmission device 30. The thread name includes the processor name and the process name in which the continuation thread is executed, and corresponds to the thread 1 in the process 3 executed in the processor 2B in FIG.

The data transmission device 30 uses the written data.
(Data) is the address of the distributed memory 27 (Local-Phys-Addres
sA: Read from the transfer destination processor (DeA
stination: In FIG. 2, data is transmitted to the processor 2B). At this time, the data transmission device 30 determines that the data (Dat
a) with the virtual shared memory address (Local-Mem-Addre
ss: X in Fig. 2) and the name of the thread on the reading side to be activated (Thre
transfer ad-ID).

When the destination processor receives the data,
The event is picked up by the data receiving device 31 and held in the interrupt queue 32. At this time, the data receiving device 31 uses the address (Local-Mem-A) of the virtual shared memory.
ddress: X in FIG. 2) is used to obtain the address (Local-Phys-Ad-B: XB in FIG. 2) of the distributed memory of the transfer destination processor, and the transferred data (Data) is assigned to the distributed memory address (Local
-Phys-Ad-B: Write to XB) in Figure 2. When this writing is completed, the event is put in the interrupt queue 32.

In the arithmetic unit 22, the thread execution management unit 2
6 works. The thread execution management unit 26 has a thread execution unit 26a and an interrupt processing execution unit 26b. The thread execution unit 26a extracts the execution waiting thread held in the execution waiting thread queue 33 from the head of the queue and executes it. The execution thread (read thread) is
When executing the read instruction in the thread, the read access is performed to the address (Local-Phys-Ad-B: XB in FIG. 2) of the distributed memory.

The execution management of the threads including the writing side thread and the reading side thread is performed as follows. 1. When the execution of one thread is completed, the control is transferred to the interrupt processing execution unit 26b. The interrupt process execution unit 26b repeats the next operation until the interrupt queue 32 becomes empty. 2. An interrupt is taken out from the head of the interrupt queue, the cause of the interrupt is checked, and the corresponding interrupt processing routine is executed. 3. Of the threads registered in the synchronization waiting thread management table 28, the synchronization counter is decremented by 1 for the thread waiting for the event (thread synchronization processing). 4. As a result, when the value of the synchronization counter becomes zero, the thread is deleted from the synchronization waiting thread management table 28 and put in the execution waiting queue. 5. When the processing of the interrupt processing execution unit 26b is completed, the control is transferred to the thread execution unit 26a.

According to the above embodiment, a program including writing and reading in the virtual shared memory is thread-programmed, and the synchronization between the writing side thread and the reading side thread is controlled by using the synchronization signal. , Overhead and burden on the user are reduced.

In the user program, the write operation and the read operation are described without considering the synchronization of the write operation and the read operation on the virtual shared memory, and the language processing system performs the write operation and the read operation for the memory resource. By analyzing the interdependencies between them, locations where synchronization control is required are automatically extracted, divided into program pieces that include write operations and program pieces that perform read operations, and each is threaded.

At the time of program execution, the virtual shared memory space is mapped to the distributed memory space distributed in each processor, and the write operation and the read operation are executed by the thread including each operation. Is performed on distributed memory on the processor.

The writing side thread writes in the distributed memory existing on the writing side, and transmits a writing end notification to the reading side thread after the writing is completed. When the reading side thread is executed in the same processor as the writing side thread, the reading side thread is started when the writing end notification is received, and reads the distributed memory on the processor.

On the other hand, when the read side thread is executed on another processor, the data transfer device operates according to the write end notification to transfer data from the write side distributed memory to the read side distributed memory. At the same time, the name of the reading thread is transmitted to the reading processor. After reading the data in the distributed memory, the data transfer device on the read side transmits a write end notification to the read side thread and activates the read side thread.

Therefore, according to the present embodiment, the program for performing the write operation and the read operation to the virtual shared memory is divided into the write side thread program and the read side thread program, and the write operation and the read operation are respectively performed. Synchronous control between threads for a writing thread and a reading thread that are handled as events and executed on the same or different processors according to the inter-thread dependency based on the order relationship between the writing event and the reading event. I do.

As a result, the write operation and the read operation can be executed at appropriate timing, and the following effects can be obtained. 1. The memory lock operation and the memory unlock operation at the time of writing and reading are not necessary. 2. Simplify management for maintaining data consistency between distributed memories. 3. The wait time and / or number of waits for write and read operations can be significantly reduced and process state control associated with write and read operations can be simplified. 4. Since the read operation and the write operation to the shared memory can be simply described in the user program without considering the write operation timing and the read operation timing between the distributed memories, the program design becomes easy.

As a result, the distributed shared memory management mechanism can be simplified, and the configuration of the parallel distributed computer and the design of the parallel distributed operating system can be simplified.

[Brief description of drawings]

FIG. 1 is a program block diagram showing the structure of a thread program that describes writing and reading to and from a virtual shared memory in the first and second embodiments of the present invention.

FIG. 2 is a program block diagram showing a structure of a thread program including write and read operations to a distributed memory executed in the first and second embodiments of the present invention.

FIG. 3 is a block diagram of hardware showing a mechanism of data transfer between distributed memories and a thread synchronization control.

FIG. 4 is a diagram showing in detail the data transfer device and the arithmetic device of FIG.

FIG. 5 is a block diagram of a hardware device according to a second embodiment of this invention.

FIG. 6 is a diagram showing an operation process of software for data transmission / reception and thread execution management accompanying writing and reading of data in FIG. 4;

[Explanation of symbols]

1 virtual shared memory 2A, 2B processor 3A, 3B, 22 arithmetic unit 4A, 4B data transfer device 5A, 5B, 27 distributed memory 6A, 6B thread execution device 7A, 7B, 28 Synchronization waiting thread management table 8A, 8B, 29 Waiting thread queue 9 Data reading device 10 Data transmitter 11 Data receiving device 12 Data writing device 13 Thread synchronization management device 21 Hardware device 23 Memory device 24 Data transmitter / receiver 25 bus 26 Thread execution manager 26a Thread execution unit 26b Interrupt processing execution unit 30 data transmitter 31 Data receiver 32 interrupt queue

Continuation of front page (51) Int.Cl. ⁷ Identification code FI theme code (reference) G06F 15/177 681 G06F 15/177 681B

Claims (9)

[Claims] 1. A means for extracting a read side thread program and a write side thread program from a program for performing a read operation and a write operation, and a write side thread and a read side thread executed on the same or different processors. On the other hand, there is provided means for performing synchronous control between the read side thread program and the write side thread program according to the relationship between the read side thread and the read side thread based on the read operation and the write operation. Virtual shared memory characterized by the following. 2. The virtual shared memory according to claim 1, wherein the read operation and the write operation are treated as a write event and a read event, respectively. 3. A format suitable for performing read and write operations in the same or different processors after converting the read thread program and the write thread program from existing program code by a language processing system, respectively. The virtual shared memory according to claim 1 or 2, wherein the read thread program and the write thread program converted into the above are used. 4. A step of extracting a reading side thread program and a writing side thread program from a program for performing a reading operation and a writing operation, and a writing side thread and a reading side thread executed on the same or different processors. On the other hand, according to the relationship between the read side thread and the read side thread based on the read operation and the write operation, the step of performing synchronous control between the read side thread program and the write side thread program. A virtual shared storage method characterized by the above. 5. The virtual shared storage method according to claim 4, wherein the read operation and the write operation are treated as a write event and a read event, respectively. 6. A format suitable for performing read and write operations in the same or different processors after converting the read thread program and the write thread program from existing program code by a language processing system, respectively. 6. The virtual shared storage method according to claim 4, wherein the read thread program and the write thread program converted into the above are used. 7. A step of extracting a reading side thread program and a writing side thread program from a program for performing a reading operation and a writing operation, and a writing side thread and a reading side thread executed on the same or different processors. On the other hand, according to the relationship between the read side thread and the read side thread based on the read operation and the write operation, the step of performing synchronous control between the read side thread program and the write side thread program. A virtual shared storage program executed by a computer characterized by the following. 8. The virtual shared storage program executed by a computer according to claim 7, wherein the read operation and the write operation are treated as a write event and a read event, respectively. 9. A format suitable for performing read and write operations in the same or different processors after converting the read thread program and the write thread program from existing program code by a language processing system, respectively. 9. The virtual shared storage program executed by a computer according to claim 7, wherein the read thread program and the write thread program are converted into the above.