JPS62126457A - Distributed data processing system - Google Patents

Abstract

PURPOSE:To attain the consistent holding of data by requesting the execution of processing for distributed data to a processor element holding the dispersion data. CONSTITUTION:Distributed data and processing procedure for processing the distributed data are collected as an object 16 and the object is individually controlled by respective processor element 10, 10A-10E. When a processing request for distributed data included in another object is generated in a certain processor element, the processing is requested to another object based on the inter-object communication and the inter-processor element communication and the processing of the distributed data can be attained by using only a message relating the processing request and the processing report as the object for the communication without moving the data itself among respective processor elements. Consequently, a communication neck based upon the inter-processor element communication can be released and the consistent holding of data can be attained.

Description

[Detailed Description of the Invention] [1 Already Required] The distributed data processing method of the present invention collects distributed data and processing procedures for processing the distributed data as an object in a multiprocessor system composed of a plurality of processor elements. In this case, the object is managed separately in each processor element, and a processing request for distributed data owned by another object occurs in one processor element.

Through inter-object communication and inter-processor element communication, processing requests are made to other objects and the data itself does not need to be moved between each processor element.
Distributed data processing can be realized by communicating only messages related to processing requests and processing reports, making it possible to eliminate communication bottlenecks and maintain data consistency through communication between processor elements.

[Industrial application field]

Distributed data processing method 2 In particular, in a multiprocessor system composed of a plurality of processor elements, when processing data distributed among the plurality of processor elements, a processing request is sent to the processor element that owns the data. The present invention relates to a distributed data processing method that aims to reduce the communication load by doing so.

[Conventional technology]

In a multiprocessor system made up of a large number of processor elements without secondary storage, suppose that each processor element performs similar processing on a substantial amount of data.

In this case, it is practically impossible for each processor element to have a complete amount of data, since the memory of the processor element is limited.

Therefore, a large amount of data is divided and distributed among a large number of processor elements, and when each processor element needs this distributed data, it is automatically transmitted using the communication path between the processor elements. A method is being considered in which the data is loaded into the memory of a processor element and a series of processes are executed.

[Problem that the invention seeks to solve]

According to the conventionally considered method described above, there is a problem in that when data loading occurs frequently, the throughput decreases due to the communication bottleneck of data loading, and the processing efficiency of the entire system deteriorates.

In particular, in the case of processing that involves not only data reference but also data modification, communication is required to return the modified data again, resulting in a large communication bottleneck. In addition, in the case of processing that involves changing data, other processor elements similarly access the distributed data after loading the data, updating the data, and returning the data. Then, there is a problem in that old data is processed and it is difficult to maintain data consistency.

An object of the present invention is to solve the above-mentioned problems, and to provide a distributed data processing method that eliminates communication bottlenecks in data loading and makes it possible to maintain data consistency.

[Means for solving problems]

FIG. 1 is a block diagram for explaining the basic configuration of the present invention.

In FIG. 1, 10. IOA~IOE are each CP
Processor element (PE) with U.

Reference numeral 11 (same for IIA; the same shall apply hereinafter) is an inter-processor element communication control unit that controls the transmission and reception of messages between processor elements; 12 is a message transmission processing unit that processes message transmission requests from objects; and 13 is an object of its own processor element. 14 is a message reception processing section that processes message reception for objects of its own processor element; 15 is an object identifier for each object of the entire system;

16 is an object; 17 is a data section that holds distributed data; 18 is a processing procedure section that includes a group of processing instructions for the data section 17;

20 represents a communication path used for communication between processor elements. In FIG. 1, the communication path 20 is a ring connection, but it may also be a bus connection, a lattice connection, etc. 2. The present invention does not depend on the communication form.

The multiprocessor system according to the present invention includes processor elements 10. IOA, 1OB... and 1
It consists of a communication path 20. Note that one processor element or another processor.

It may also be configured to act as a host processing device and manage the entire system.

Each processor element 10 etc. has an object 16 distinguished by an object identifier. The object 1-16 consists of a data section 17 and a processing procedure section 18, and the data section 17 has distributed data obtained by dividing related data. The processing procedure section 18 has a group of instructions for interpreting messages and processing one distributed data according to a processing request message. The contents of the data section 17 are as follows for each object 1.
Although the contents of the processing procedure section 18 may be different for each object 16, the contents of the processing procedure section 18 may be common to each object 16. The contents of the processing procedure section 18 are
If each object 16 is the same.

Management becomes easier.

When the object 16 processes distributed data held by another object 16, the message transmission processing unit 1
2, the message transmission processing unit 12 that sends messages regarding processing contents uses the object management table 15.
, and if the destination object 16 is within the own processor element, the message sent processing unit 13 is activated.

If it exists in another processor element, it requests the inter-processor element communication control unit 11 to transfer the message.

The message sent processing unit 13 connects the specified message to the queue of the destination object in the object management table 15.

The inter-processor element communication control unit 11 has two communication channels 20
The inter-processor element communication control unit 11 that received the message sends the message to the inter-processor element communication control unit in another processor element via the message reception processing unit 14, and the message reception processing unit 14, Message set processing section 13
Start. The message set processing unit 13 queues messages in the object management table 15. Object 1 requested for processing by queuing
6 obtains a message from the object management table 15 and interprets the message using the processing procedure section 18;
Execute processing.

After the execution is completed, the processing results are collected as a message if necessary and sent back to the requesting object 16, but this message communication is performed in the same way as the message communication at the time of the request.

[Effect]

According to the present invention, when a certain object 16 needs to process other distributed data, only a message indicating what kind of processing is requested to the object that has the distributed data is sent. Ru.

Therefore, generally, one piece of distributed data itself does not flow on one communication path 20 except in the case of an initial load, and only a message indicating the contents of a processing request flows, so that the amount of data communication between processor elements is significantly reduced.

Further, since processing for one piece of distributed data is performed by the processing procedure unit 18 of the object 16 that has the distributed data, the processing is serialized and the consistency of the data is not impaired.

〔Example〕

FIG. 2 is an example of an object management table used in the present invention, and FIG. 3 is an explanatory diagram of object processing.

FIG. 4 is an explanatory diagram of the message transmission process, FIG. 5 is an explanatory diagram of the message reception process, and FIG. 6 is an explanatory diagram of the message set process.

The object management table 15 is configured, for example, as shown in FIG.
), object PE address (OBJ PE A
DDR), object state (OBJ 5TATU,
S), has a message queue (MSG QUE) field. Note that other attribute fields may be provided as necessary.

The object identifier is a number assigned to uniquely identify each object in the system. The address of the processor element where the object exists is set in the object PE address field. The object status field contains object status information such as whether the object is waiting for processing or is being processed. Messages requesting processing for the object are queued in the message queue field.

Each object receives and interprets messages sent to it, and stores in advance in the processing procedure section 18 what processing to do with the messages. A processing request for distributed data is realized by sending a message M esg indicating the processing content in the 9th order format to the object Obj[i] that holds the distributed data.

5end (Obj[t] + Mesg) Each object executes processing for the sent message Mesg, as shown in FIG. The following number■
~■ correspond to the numbers shown in FIG.

(2) Determine whether a message sent from another object is set in the message queue of the own object in the object management table 15. If the message is set, control is transferred to process ■.

- If the message is not set, set the object status in the object management table 15 to WAIT status.

■ Suspends execution and waits for a processing request from another object. When a message is received, the WAIT state is released and the process proceeds to process ■.

■ Messages are retrieved from the message queue in principle in a first-in, first-out (FIFO) manner.

■ Interpret the sent message through processing by the processing procedure unit 18 of its own object.

Execute the requested process. Also, if necessary, during this processing, a message is sent to other objects to request processing or report processing.

If the execution is finished, return control to process ■.

Repeat the process for the next message.

When an object sends a message to another object, the message transmission processing unit 12 shown in FIG. 1 is activated, and the message transmission processing unit 12 processes as shown in FIG. 4. The following numbers Φ to ■ correspond to the numbers shown in FIG.

■ Search the object management table 15 using the object identifier as a key, and find the address of the processor element that actually contains the object (OBJ P
Extract EADDR).

■ Extract the own processor element address (PE ADDR) from a predetermined storage area.

■ Compare the above OBJ PE ADDR and its own processor element address (PE ADDR), and if they are the same, transfer control to process ■.

If it is different, it is necessary to send a message to an external processor element, so control is transferred to process (2).

■ Call the message set processing unit 13.

Performs message set processing and waits for the first message transmission request.

(2) Calling the inter-processor element communication control unit 11 and performing inter-PE communication via the 1 communication path 20.

Send a message. Then, it waits for a second message transmission request.

The message reception processing unit 14 performs the processing (1) and (2) shown in FIG. 5 in a primary manner.

■ The inter-processor element communication control unit 11 allows P
Receive messages using E-to-E communications.

■ Call the message set processing unit 13.

Message set processing is performed and the received message is queued in the object management table 15. Then, it waits for the first message to be received.

The message set processing unit 13 is activated by the message transmission processing unit 12 or the message reception processing unit 14,
Processing is performed as shown in FIG. The following numbers ■ to ■ correspond to the numbers shown in FIG.

■ Extract the destination object identifier from the message.

(2) Using the object +-it identifier as a key, search the object management table 15 and extract the message queue of the corresponding entry.

■ Set a message at the end of the message queue.

■ The object state of the current object is WA.
Determine whether or not it is IT. If it is not WAIT, the message set processing is immediately terminated.

■ If the object is in WAIT state, its W
Release the AIT state and resume execution of the object.

〔Effect of the invention〕

As explained above, according to the present invention, processing for one distributed data is executed by the object of the processor element that has the distributed data.

Communication bottlenecks caused by data loading can be eliminated, and data consistency can be maintained. Furthermore, since the data section and the processing procedure section are integrated as objects and distributed to each processor element, for example, when the number of processor elements increases or when objects are added/deletion, the object management table 15 is simply updated. This can be handled simply by updating, allowing for a flexible system configuration.

[Brief explanation of drawings]

FIG. 1 is a diagram for explaining the basic configuration of the present invention. Fig. 2 is an example of an object management table used in the present invention, Fig. 3 is an illustration of object processing, Fig. 4 is an illustration of message transmission processing, Fig. 5 is an illustration of message reception processing, and Fig. 6 is a message set. A processing explanatory diagram is shown. In the figure, 10 is a processor element, 11 is an inter-processor element communication control section, 12 is a message transmission processing section, and 13 is a message set processing section. 14 is a message reception processing section, 15 is an object management table, 16 is an object, 17 is a data section, 18
2 represents a processing procedure section, and 20 represents a communication path.

Claims (1)

[Claims] A plurality of processor elements (10, 10A, 10B,
In a distributed data processing method in a multiprocessor system configured with...), each of the above processor elements holds an object (16) that summarizes data distributed to each processor element and a processing procedure for processing the distributed data. The object management table (15) holds information about which processor element each of the objects (16) exists in; 15) and means (12, 13, 1) for transmitting and receiving messages regarding processing requests or processing reports.
4), and means (11) for communicating messages between the processor elements in response to a request from the inter-object message transmitting/receiving means, requesting the processor element holding the distributed data to process the distributed data. , a distributed data processing method characterized in that it is executed.