JPH09237255A - Dynamic load distribution method for parallel computer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the quantity of load to be required for dynamic load distribution and to improve the efficiency of processing by executing dynamic load distribution so as to request task allocation to another processor after judging that a certain processor can not process a request task newly generated in the processor itself. SOLUTION: When a task generation request is generated in a certain processor, the processor executes practicability/impracticability judgement for the task based upon load information and task information (1). At the time of judging impracticability, the processor acts as a sender, the sender selects a processor having a processor number next to that of the sender itself as a receiver candidate (3) and transmits the task information to the receiver candidate. When the candidate has a margin, i.e., quali-fiction (6), for processing the task at the time of receiving a qualification judgement result from the candidate, the candidate is determined as a receiver (7) and task allocation is requested to the receiver (8).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for performing dynamic load distribution in a real-time parallel computer so that all tasks generated by each processor can be processed within a predetermined frame time. It is a thing.

[0002]

17 shows a conventional parallel computer system in which a plurality of processors 102 to 1mn are connected by a network 101. In order for all the processors in the system to operate efficiently, tasks that occur sequentially are sequentially assigned to these processors for distributed processing. Each processor generates load information from the processing time of the assigned task.

Conventionally, in such a parallel computer system, when a new task generation request is made in a state where each processor has a heavy load, the processor is searched for the lightest load by checking the load information of the other processors. Request the processor to allocate the task, or a processor with a heavy load randomly selects another processor and examines the load information of the selected processor to determine whether the load is lighter than itself. When the other processor has a lighter load, the processor is requested to be assigned the task, and the loads of all the processors are averaged to improve the efficiency of the entire system.

[0004]

In the conventional parallel computer system as described above, when a processor with a heavy load requests a processor with the lightest load to request task allocation, the load information of the other processor is sequentially read. Since it has to be investigated, it may take a lot of time in some cases, and in the method of randomly selecting the other party, the load of the selected processor is heavier than one's own load, so it takes a lot of time to repeat selection. In some cases, even if the load of the other processor is lighter than itself, there is no difference, and the processing time required for assigning the task is longer and meaningless load distribution may be performed. There is a problem that the load for performing dynamic load distribution increases, and the processing efficiency of the original application task decreases.

The present invention has been made to solve the above problems, and in a real-time multiprocessor system in which each processor has to process a task assigned within a predetermined frame time, When a new generation request task occurs in a certain processor, it is first judged whether or not the request task can be processed in the processor, and only when it is determined that the task cannot be processed, the processor becomes a sender and another processor. Dynamic load balancing is performed in order to request task allocation to. As a result, it is not necessary to always try to distribute the load, and thus it is possible to prevent the load of the processor from unnecessarily increasing. In addition, each processor requested by the sender to assign a task determines whether or not the task can be executed, so that the sender does not need to determine whether or not the processor other than itself can execute the task. As a result, it is necessary to determine the receiver of the sender. The amount of load can be reduced. From the above, the purpose is to reduce the amount of load required for dynamic load distribution in the entire system and improve processing efficiency.

[0006]

According to a first aspect of the present invention, there is provided a dynamic load balancing method for a real-time parallel computer, comprising: a means for determining whether a processor is executable when a task generation request is made; When it becomes impossible to execute the processing of the task, the processor becomes a sender, and means for selecting a receiver candidate as a candidate of a processor that substitutes the processing of the task from other processors, Means for transmitting information regarding the task by using the received task information, a means for determining whether the receiver candidate is qualified as a receiver based on the received task information, a means for returning the determination result to the sender, and a means for receiving the information from the receiver candidate. If a sender receives a pass notification, and the sender has enough time to complete the frame time, he or she can select a new receiver candidate. And means to fix, is achieved by providing a means for requesting the processing proxy of the task sender if the acceptance notification has arrived to determine the receiver candidate to the receiver.

The dynamic load balancing method for a real-time parallel computer according to the second aspect of the present invention is a means for each processor to broadcast task information to all other processors and a means for each processor to freely access the shared memory. When a task generation request is issued to a certain processor, the processor determines whether or not the processor can execute the task. If the result of the determination indicates that the task cannot be executed, all the broadcast means are used. By means of transmitting task information to the processor, means for each processor to perform qualification determination as a receiver based on the received task information, means for writing the determination result in the shared memory, and the sender reading the shared memory. Equipped with a means to find one qualified receiver, determine it as a receiver, and request processing on behalf of the task It is achieved by.

A dynamic load balancing method in a real-time parallel computer according to a third aspect of the present invention is a method for transmitting a reject signal to the sender when the receiver cannot accept the task generation request after determining the receiver, and a frame time end. This is achieved by providing the sender with means for iteratively selecting a new receiver candidate and determining a receiver when there is time to spare.

The dynamic load balancing method for a real-time parallel computer according to the fourth aspect of the invention is such that if the receiver cannot accept the task generation request after the receiver is determined,
Means for sending a reject signal to the receiver to the sender;
This is achieved by providing a means for the sender to newly determine a receiver from the remaining receiver qualified persons when the sender has time to complete the frame time.

A dynamic load balancing method for a real-time parallel computer according to the fifth aspect of the present invention is a method for a sender to prepare a list of qualified receivers when a receiver is determined, and a task generation for a processor determined by the receiver. If there is a way to send a list of qualified people with the required information and the receiver cannot accept the task creation request after the receiver has been determined, the receiver will be registered in the qualified list as long as there is time before the end of the frame time. It is achieved by providing means for newly determining a receiver from among existing processors.

Further, in the dynamic load balancing method for a real-time parallel computer according to the sixth and seventh inventions, the receiver candidate is requested to generate when the judgment result is "pass" at the time of judging the receiver candidate qualification. This is achieved by providing means for preliminarily reserving the resources required to generate a task.

According to an eighth aspect of the present invention, there is provided a dynamic load balancing method for a real-time parallel computer, which is a receiver candidate other than the receiver candidate actually determined as the receiver among the receiver candidates whose judgment result is acceptable at the time of qualifying the receiver candidate. This is achieved by providing means for the sender to instruct the cancellation of the reservation of the candidate resource.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION

Embodiment 1. 1 and 2 are processing flow charts showing Embodiment 1 of the present invention. FIG. 1 is a process flow chart when a task generation request is issued to a processor, and FIG. 2 is a process flow chart of a processor which is selected as a receiver candidate by a sender and to which information regarding the task is transmitted. In FIG. 1, when a task generation request occurs in a certain processor, the processor obtains the task scheduling time schedule until the end of the frame time, load information such as memory usage status, and the required memory amount of the task. Executability determination 1 of the task is performed based on the task information regarding the required processing time and the like. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. When it is determined that the processor cannot be executed, the processor becomes a sender, and the sender selects 3 as the receiver candidate for the processor having the processor number next to itself, and sends 4 task information to the receiver candidate. In FIG. 2, the receiver candidate that has received the task information determines whether 51 is qualified to process the task based on the task information and the load information obtained as a result of the task scheduling of the receiver candidate. Reply to the sender. Since another task has already been assigned from the current time to the end time of the frame time, if there is no room to process the task, it is determined to be unqualified, and otherwise, it is determined to be qualified. In FIG. 1, when the qualification determination result from the receiver candidate is 5 and the qualification is 6, the receiver candidate is determined as the receiver 7 and the task allocation request 8 is made. In the case of no qualification 6, if there is time to perform task processing and allocation request by the end of the frame time, if the processor has the next processor number, it is re-executed from the receiver candidate selection 3 If there is no time, the task generation request is rejected at 9.

Embodiment 2 FIG. 3 and 4 are process flow charts showing the second embodiment of the present invention. FIG. 3 is a processing flow chart when a task generation request is issued to a certain processor, and FIG. 4 is a processing flow chart of a processor which is selected as a receiver candidate by a sender and to which information regarding the task is transmitted. In FIG. 3, when a task generation request is made in a certain processor, the processor obtains the task scheduling time schedule until the end of the frame time, load information such as memory usage status, and the required memory amount of the task. Executability determination 1 of the task is performed based on the task information regarding the required processing time and the like. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. If it is determined that the processor cannot be executed, the processor becomes a sender and broadcasts task information 10 to other processors. In FIG. 4, each processor that receives the task information determines whether or not there is a qualification for processing the task based on the task information and the load information obtained as a result of its task scheduling, and shares the determination result. Write 52 to memory. Since another task has already been assigned from the current time to the end time of the frame time, if there is no room to process the task, it is determined to be unqualified, and otherwise, it is determined to be qualified. In FIG. 3, the sender reads 11 the shared memory and retrieves 12 one qualified person who wrote that it is qualified. If you have a qualified person 6
Determines 7 the qualified person as a receiver and makes a task allocation request 8 to the receiver. If there is no qualified person 6, when there is a time margin to process a task or request an allocation by the end of the frame time, 9 is read again from the shared memory and re-executed. Request to generate the task is rejected.

Embodiment 3 5 and 6 are processing flow charts showing Embodiment 3 of the present invention. FIG. 5 is a process flow chart when a task generation request is issued to the processor, and FIG. 6 is a process flow chart of the processor which is selected as a receiver candidate by the sender and to which information regarding the task is transmitted.
In FIG. 5, when a task generation request is generated in a certain processor, the processor obtains the task scheduling time schedule until the end of the frame time obtained as a result of the task scheduling, and load information such as memory usage status,
Execution permission / inhibition determination 1 of the task is performed based on the task information regarding the required memory amount and the required processing time of the task. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. When it is determined that the processor cannot be executed, the processor becomes a sender, and the sender selects 3 as the receiver candidate for the processor having the processor number next to itself, and sends 4 task information to the receiver candidate. In FIG. 6A, the receiver candidate that has received the task information determines whether or not the task processing agent qualifies based on the task information and the load information obtained as a result of its task scheduling. The judgment result is sent back to the sender. Since another task has already been assigned from the current time to the end time of the frame time, if there is no room to process the task, it is determined to be unqualified, and otherwise, it is determined to be qualified. FIG.
When the result of receiving the qualification determination result 5 from the receiver candidate is 6 as a result of the qualification, the receiver candidate is determined as the receiver 7 and the task allocation request 8 is made. FIG.
In (b), the receiver that has received the task allocation request determines whether or not the task can be executed 53 based on its own load information and the information of the task, and if possible, generates 54 the task. An acceptance signal is sent 55 to the sender to report that the task allocation request has been accepted.
Execution is impossible because the receiver receives a task allocation request due to a delay in the network, etc., and the receiver accepts another task allocation request from another processor during that time. If it is determined that the reject signal is sent to the sender 56
I do. In FIG. 5, the sender waits for a response from the receiver 13 after making a task allocation request 13, ends load balancing when an accept signal is returned, and ends the frame time when a reject signal is returned. If there is enough time to process the task or request the allocation, the process is re-executed from the receiver candidate selection 3, and if there is no time, the task generation request is rejected.
As a result of the qualification judgment, if there is no qualified person 6,
If there is enough time, the receiver candidate selection 3 is re-executed 9 and if there is no time, the task generation request is rejected.

Embodiment 4 7 and 8 are process flow charts showing the fourth embodiment of the present invention. FIG. 7 is a processing flow chart when a task generation request is issued to a certain processor, and FIG. 8 is a processing flow chart of the processor which is selected as a receiver candidate by the sender and to which information regarding the task is transmitted. In FIG. 7, when a task generation request is generated in a certain processor, the processor determines load information such as a time schedule of task allocation and memory usage status until the end of the frame time obtained as a result of task scheduling, and the required memory amount of the task. Executability determination 1 of the task is performed based on the task information regarding the required processing time and the like. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. If it is determined that the processor cannot be executed, the processor becomes a sender and broadcasts task information 10 to other processors. In FIG. 8A, each processor that has received the task information determines whether or not the processor is qualified to process the task 51 based on the task information and the load information obtained as a result of its task scheduling. Write 52 results to shared memory. If another task is already assigned from the current time to the end time of the frame time and there is no room to process the task, it is determined that there is no qualification,
If not, it is determined to be qualified. In FIG.
The sender reads 11 the shared memory and retrieves 12 one qualified person who wrote that he / she is qualified. When there is a qualified person, one receiver 7 is decided from among those, and a task allocation request 8 is made to the receiver. In FIG. 8B, the receiver having received the task allocation request determines 53 whether the task can be executed, generates a task 54 if possible, and sends 55 an accept signal to the sender. , Report that the task allocation request has been accepted. The receiver cannot receive the task allocation request due to a delay in the network, etc., so that the receiver receives the task allocation request from another processor in the meantime and the task cannot be executed. When it is determined, the reject signal is transmitted 56 to the sender. In FIG. 7, the sender waits for a response from the receiver 13 after making the task allocation request 8, ends load balancing when an accept signal is returned, and ends the frame time when a reject signal is returned. If there is enough time to process the task or request allocation, the shared memory reading 11 is re-executed, and if there is no time, the task generation request is rejected.
If there is no qualified person as a result of qualification judgment, 6 is also the same.
If there is enough time, read the shared memory 11
Is re-executed, and if there is not enough time, the task generation request is rejected.

Embodiment 5 9 and 10 are process flow charts showing the fifth embodiment of the present invention. FIG. 9 is a process flow diagram when a task generation request is issued to a certain processor,
FIG. 10 is a processing flow chart of the processor which is selected as a receiver candidate by the sender and to which the information regarding the task is transmitted. In FIG. 9, when a task generation request is made in a certain processor, the processor obtains the task scheduling time schedule until the end of the frame time, the load information such as the memory usage status, and the required memory amount of the task. Executability determination 1 of the task is performed based on the task information regarding the required processing time and the like. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. When it is determined that the processor cannot be executed, the processor becomes a sender and broadcasts task information to other processors.
I do. In FIG. 10A, each processor that has received the task information determines whether or not the processor is qualified to process the task 51 based on the task information and the load information obtained as a result of its task scheduling. Write 52 results to shared memory. In FIG. 9, the sender reads 11 the shared memory and retrieves 12 all qualified personnel who have written qualified. As a result of the search, a list 15 of all receiver qualified persons determined to be qualified is created. When there is a qualified person 6, the processor with the smallest processor number among them is determined as a receiver 7 and a task allocation request 8 is made to the receiver. At this time, a list obtained by removing oneself from the created receiver qualified person list is also sent. In FIG. 10B, the receiver that has received the task allocation request determines 53 whether the task can be executed, and if possible, generates a task 54,
An acceptance signal is sent 55 to the sender to report that the task allocation request has been accepted. It is not possible to execute because the receiver receives the task allocation request due to delay in the network, etc., and the receiver accepts the task allocation request from another processor in the meantime. If it is determined, when there is enough time to process the task or request the allocation until the end of the frame time 57, the qualified person is searched 58 from the list of the receiver qualified persons who have already received, and the qualified person In some cases 59, the one with the smallest processor number is newly determined as a receiver 60 and a task allocation request 61 is made. If you do not have time 57 or if you do not have a qualified person 59
Sends 56 a reject signal to the sender for which the task generation request has occurred. In FIG. 9, the sender waits for a response from the receiver after performing the task allocation request 8, and ends load balancing when an accept signal is returned, and ends the frame time when a reject signal is returned. If there is enough time to process the task or request the allocation by 14 then re-execute from the shared memory reading 11
If there is not enough time, the task generation request is rejected if 14. Similarly, in the case where there is no qualified person as a result of the qualification judgment, if there is enough time, then 9 is read again from the shared memory 11 and if there is no time, 9
Request to generate the task is rejected.

Embodiment 6 FIG. 11 and 12 are process flow charts showing the sixth embodiment of the present invention. FIG. 11 is a processing flow chart when a task generation request is issued to a certain processor, and FIG. 12 is a processing flow chart of a processor which is selected as a receiver candidate by a sender and to which information regarding the task is transmitted. In FIG. 11, when a task generation request occurs in a certain processor, the processor obtains the task scheduling time schedule until the end of the frame time, load information such as the memory usage status, and the required memory amount of the task. Whether or not the task can be executed based on task information related to the required processing time, etc. 1
I do. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. When it is determined that the processor cannot be executed, the processor becomes a sender, and the sender selects 3 as the receiver candidate for the processor having the processor number next to itself, and sends 4 task information to the receiver candidate. In FIG. 12A, the receiver candidate that has received the task information determines whether or not the task candidate is qualified to process the task 51 based on the task information and the load information obtained as a result of the task scheduling of the receiver candidate. The judgment result is sent back to the sender. Since another task has already been assigned from the current time to the end time of the frame time, if there is no room to process the task, it is determined to be unqualified, and otherwise, it is determined to be qualified. At this time, if it is determined that you are qualified 59
In the receiver candidate, resources such as a memory and a processing time within a frame time necessary for generating / executing a task are secured 62, and the task is surely requested when a task allocation request is actually made. Allow it to run. In FIG. 11, when the reply 5 from the receiver candidate is received and it is qualified 6, the receiver candidate is determined as the receiver 7 and the task allocation request 8 is made. In FIG. 12B, the receiver that has received the task allocation request makes a determination 53 as to whether the task can be executed, generates a task 54 if possible, and sends 55 an accept signal to the sender. , Report that the task allocation request has been accepted.
The receiver cannot receive the task allocation request due to a delay in the network, etc., so that the receiver receives the task allocation request from another processor in the meantime and the task cannot be executed. When it is determined, the reject signal is transmitted 56 to the sender. In FIG. 11, the sender waits for a response from the receiver 13 after making the task allocation request 8, ends load balancing when an accept signal is returned, and ends the frame time when a reject signal is returned. If there is enough time to process the task or request allocation, the processor with the next processor number is re-executed from the receiver candidate selection 3, and if there is no time, the task Reject the generation request. Similarly, in the case where there is no qualification as a result of the qualification determination, in the case where there is enough time, the process is re-executed from the selection 3 of the receiver candidates, and when there is no time, a request to generate the task is issued. Reject.

Embodiment 7. 13 and 14 are process flow charts showing the seventh embodiment of the present invention. FIG. 13 is a process flow diagram when a task generation request is issued to a certain processor, and FIG. 14 is a process flow diagram of the processor which is selected as a receiver candidate by the sender and to which information regarding the task is transmitted. In FIG. 13, when a task generation request is made in a certain processor, the processor obtains the load information such as the time schedule of task allocation and the memory usage status until the end of the frame time obtained as a result of the task scheduling, and the required memory amount of the task. Whether or not the task can be executed based on task information related to the required processing time, etc. 1
I do. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. If it is determined that the processor cannot be executed, the processor becomes a sender and broadcasts task information 10 to other processors. In FIG. 14A, each processor that has received the task information determines whether or not there is a qualification for processing the task based on the task information and the load information obtained as a result of its own task scheduling. Write 52 results to shared memory. At this time, when it is determined that the user is qualified, 59, the task generation / reception in the receiver candidate is performed.
Resources 62 such as a memory and a processing time within a frame time necessary for execution are secured 62 so that the task can be surely executed when a task allocation request is actually made. In FIG. 13, the sender reads the shared memory 1
1. Search one receiver candidate who wrote that it is qualified 1
I will do 2. When there is a qualified person, the receiver qualified person is determined as the receiver 7 and the task allocation request 8 is made to the receiver. In FIG. 14B, the receiver that has received the task allocation request makes a determination 53 as to whether the task can be executed, generates a task 54 if possible, and sends 55 an accept signal to the sender. , Report that the task allocation request has been accepted. The receiver cannot receive the task allocation request due to a delay in the network, etc., so that the receiver receives the task allocation request from another processor in the meantime and the task cannot be executed. When it is determined, the reject signal is transmitted 56 to the sender. In FIG. 13, the sender waits for a response from the receiver 13 after making the task allocation request 8, ends load balancing when an accept signal is returned, and ends the frame time when a reject signal is returned. If there is enough time to process tasks or request allocation, read shared memory 1
The process is re-executed from 1, and when there is no time, the request to generate the task is rejected. Similarly, in the case where there is no qualified person as a result of the qualified determination, if there is enough time, the processing is re-executed from the reading 11 of the shared memory if there is enough time, and if there is not enough time, the task generation request is issued. Reject.

Embodiment 8. 15 and 16 are flowcharts showing Embodiment 8 of the present invention. FIG. 15 is a process flow diagram when a task generation request is issued to a certain processor,
FIG. 16 is a processing flow chart of the processor which is selected as a receiver candidate by the sender and to which the information regarding the task is transmitted. In FIG. 15, when a task generation request is generated in a certain processor, the processor obtains load information such as the time schedule of task allocation and the memory usage status until the end of the frame time obtained as a result of task scheduling, and the required memory amount of the task. Executability determination 1 of the task is performed based on the task information regarding the required processing time and the like. The processing time required for the task is longer than the time from the current time to the end time of the frame time, or another task is already assigned from the current time to the end time of the frame time, so there is room to process the task. If not, it is determined to be unexecutable, and if not, it is determined to be executable. If it is determined to be executable, task generation 2 is performed in the processor and execution is started. When it is determined that the processor cannot be executed, the processor becomes a sender and broadcasts task information to other processors.
I do. In FIG. 16A, each processor that has received the task information determines whether it is qualified to process the task 51 based on the task information and the load information obtained as a result of its task scheduling. Write 52 results to shared memory. At this time, when it is judged that the receiver is qualified, resources such as a memory and a processing time within the frame time necessary for generating and executing the task are secured in the receiver candidate 62, and the task is actually allocated. Ensure that the task can be performed when requested. FIG.
At 5, the sender reads 11 the shared memory and retrieves 12 all receiver candidates that have been written as qualified.
If there is a qualified person, 6 decides the receiver with the lowest processor number among the receiver qualified persons 7
Then, the resource reservation cancellation instruction 16 is issued to the remaining receiver qualified persons. In FIG. 16B, each receiver qualified person releases 63 the resource related to the task of the own processor upon receiving the reservation cancellation instruction. In FIG. 15, the sender makes a task allocation request 8 to the receiver. In FIG. 16C, the receiver that has received the task allocation request finally determines 53 whether the task can be executed, generates a task 54 if possible, and sends an accept signal to the sender. Send 55 and report that the task allocation request has been accepted. The receiver cannot receive the task allocation request due to a delay in the network, etc., so that the receiver receives the task allocation request from another processor in the meantime and the task cannot be executed. When it is determined, the reject signal is transmitted 56 to the sender. In FIG.
After sending the task allocation request 8, the sender waits for a response from the receiver 13, ends load balancing when an accept signal is returned, and when a reject signal is returned,
If there is enough time to process a task or make an allocation request by the end of the frame time, the process is re-executed from the reading 11 of the shared memory, and if there is no time, the task generation request is rejected. Similarly, if there is no qualified person as a result of the qualification judgment, if there is enough time, the process is re-executed from reading the shared memory 11 if there is enough time, and if there is not enough time, the task is created. Reject the request.

[0021]

As described above, according to the present invention, in a real-time multiprocessor system in which each processor has to process an assigned task within a predetermined frame time, a new generation request is made by a certain processor in the system. When a task occurs, it is first determined whether the requested task can be processed in that processor, and when it is determined that it cannot be processed, dynamic load balancing is requested to another processor for task allocation. By doing so, it is possible to prevent an unnecessary increase in the load on the processor and improve the processing efficiency.

Further, when it is determined that the new generation request task cannot be processed by the own processor and the other processor is requested to perform the processing substitution, each processor performs its own qualification judgment as the receiver when determining the receiver. Therefore, it is possible to prevent a sender, which attempts to perform dynamic load distribution to reduce the load, from increasing the load for the dynamic load distribution processing.

[Brief description of drawings]

FIG. 1 is a flowchart showing a process of a sender according to a first embodiment of a dynamic load distribution method according to the present invention.

FIG. 2 is a flowchart showing processing of a receiver candidate according to the first embodiment of the dynamic load distribution method according to the present invention.

FIG. 3 is a flowchart showing processing of a sender according to the second embodiment of the dynamic load balancing method of the present invention.

FIG. 4 is a flowchart showing processing of a receiver candidate according to the second embodiment of the dynamic load distribution method according to the present invention.

FIG. 5 is a flowchart showing processing of a sender according to the third embodiment of the dynamic load balancing method of the present invention.

FIG. 6 is a flowchart showing processing of a receiver candidate according to the third embodiment of the dynamic load distribution method of the present invention.

FIG. 7 is a flowchart showing processing of a sender according to the fourth embodiment of the dynamic load balancing method of the present invention.

FIG. 8 is a flowchart showing processing of a receiver candidate according to the fourth embodiment of the dynamic load distribution method of the present invention.

FIG. 9 is a flowchart showing processing of a sender according to the fifth embodiment of the dynamic load distribution method of the present invention.

FIG. 10 is a flowchart showing processing of receiver candidates according to the fifth embodiment of the dynamic load distribution method of the present invention.

FIG. 11 is a flowchart showing processing of a sender according to the sixth embodiment of the dynamic load distribution method of the present invention.

FIG. 12 is a flowchart showing processing of a receiver candidate according to the sixth embodiment of the dynamic load distribution method of the present invention.

FIG. 13 is a flowchart showing the processing of the sender according to the seventh embodiment of the dynamic load balancing method of the present invention.

FIG. 14 is a flowchart showing processing of receiver candidates according to the seventh embodiment of the dynamic load distribution method of the present invention.

FIG. 15 is a flowchart showing the processing of the sender according to the eighth embodiment of the dynamic load balancing method of the present invention.

FIG. 16 is a flowchart showing processing of receiver candidates according to the eighth embodiment of the dynamic load balancing method of the present invention.

FIG. 17 is a schematic diagram showing a network of a conventional device and a parallel computer device according to the present invention.

[Explanation of symbols]

101 network, 102 processor, 103
Processor, 1 mn processor.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hiromi Yamazaki 2-3-3 Marunouchi, Chiyoda-ku, Tokyo Sanryo Electric Co., Ltd. (72) Hideaki Yoshioka 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Ryoden Co., Ltd. (72) Inventor Hiroyuki Takano 2-3-3 Marunouchi, Chiyoda-ku, Tokyo Sanryo Denki Co., Ltd. (72) Inventor Toru Hanazawa 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Sanryo Denki Co., Ltd. (72) Inventor ▲ Taka ▼ Masato Hashi 2-3-3 Marunouchi, Chiyoda-ku, Tokyo Sanryo Electric Co., Ltd.

Claims (8)

[Claims] 1. A plurality of processors having equivalent functions are provided, and a task in each processor must be processed within a predetermined frame time, and a task to be processed at the frame time at the start of the frame time. In a real-time parallel computer that makes a determination and performs task scheduling based on the determination result, when a task generation request occurs as a result of the task scheduling in a processor, the processor obtains a frame obtained as a result of the task scheduling. Whether or not the task can be executed is determined based on the load information such as the time schedule of task allocation and the memory usage status until the end of the time, and the task information related to the required memory amount and the required processing time of the task. The task from the time until the frame time end time If the processing execution of the task is impossible because the required processing time is longer, or another task is already assigned from the current time to the frame time end time, the processor It becomes a sender of the task, selects a processor having a processor number next to the processor of the other processors as a receiver candidate, transmits the task information to the receiver candidate, and the receiver candidate itself receives the task information and itself. On the basis of the load information obtained as a result of the task scheduling of the above, it judges whether or not the task can be executed by itself, that is, whether or not it is qualified as a receiver, and returns the judgment result to the sender. The receiver candidate has already been assigned another task between the current time and the frame time end time, etc. If a notification of success notification is sent back, as long as there is enough time to process the task and request allocation by the end of the frame time, re-execute from the selection of a new candidate receiver, and if a notification of success is returned, the receiver A dynamic load balancing method in a parallel computer, characterized in that a candidate is determined as a receiver and a processing substitution of the task is requested. 2. As a result of the task scheduling in a processor, the sender has a function of broadcasting information about a task to all other processors and a shared memory mechanism that each processor can freely access.
When a task generation request occurs, the processor obtains the task scheduling time schedule until the end of the frame time obtained as a result of the task scheduling, load information such as memory usage, and the required memory amount and required processing time of the task. Whether or not the task can be executed is determined based on the task information regarding the result, and as a result, the required processing time of the task is longer than the time from the current time to the frame time end time, or it has already been exceeded from the current time to the frame time end time. If the task cannot be executed due to another task being allocated, the processor becomes the sender of the task and sends the task information to all processors using the broadcasting function. However, each processor receives the task information and the load information obtained as a result of its own task scheduling. Based on that, it is determined whether or not the task can be executed by itself, that is, whether or not it is qualified as a receiver, the determination result is written to the shared memory, and the sender reads the shared memory and writes that it is qualified. If there is no receiver qualified person who did the job, if there is no receiver qualified person, re-execute from reading shared memory as long as there is time to process tasks and request allocation by the end of frame time, 2. The dynamic computer in a parallel computer according to claim 1, wherein when there are receiver qualified persons, the receiver qualified person having the smallest processor number among them is determined as a receiver and requested to process the task. Load balancing method. 3. When the task allocation request is transmitted to the receiver after the receiver is determined, the receiver delays receiving the task allocation request due to a delay in the network or the like. When the receiver cannot accept the task creation request due to, for example, accepting the assignment request of another task of the receiver, the receiver sends a reject signal to the sender, and the sender still has the frame time. The operation rate of a task can be improved by repeating selection of a new receiver candidate and determination of a receiver when there is enough time to process the task or make an allocation request by the end. Dynamic load balancing method in the parallel computer described. 4. When the task allocation request is transmitted to the receiver after the receiver is determined, the receiver delays receiving the task allocation request due to a delay in the network or the like. When the receiver cannot accept the task creation request due to, for example, accepting the assignment request of another task of the receiver, the receiver sends a reject signal to the sender, and the sender still has the frame time. The operation rate of a task can be improved by repeating the selection of a new receiver candidate and the determination of a new receiver when there is enough time to process the task or request an allocation by the end. Dynamic load balancing method in the parallel computer described. 5. The sender prepares a list of qualified receivers at the time of determining the receiver, and the qualified worker list is transmitted to the processor determined by the receiver together with the information necessary for task generation. When the task allocation request was sent to the receiver after the receiver was determined, the receiver was delayed in receiving the task allocation request due to a delay in the network, etc. When the receiver cannot accept the task creation request due to the reason that it has accepted the allocation request, etc., and when there is still time to process the task or request the allocation by the end of the frame time at that time. The receiver is newly selected from the processors whose receiver is registered in the qualified person list,
The dynamic load balancing method in a parallel computer according to claim 2, wherein a task processing agency is requested. 6. The receiver candidate is a resource such as a memory and a processing time within a frame time required to generate a task for which a generation request is made when the judgment result is “pass” at the time of judging the qualification of the receiver candidate. By securing in advance, the task can be surely generated / executed when the generation request is actually received.
Dynamic load balancing method in the parallel computer described. 7. The receiver candidate is a resource such as a memory and a processing time within a frame time required to generate a task requested to be generated when the result of the judgment at the time of qualifying the receiver candidate is “pass”. The task is secured in advance so that the task can be surely generated and executed when the generation request is actually received.
Dynamic load balancing method in the parallel computer described. 8. The sender instructs the cancellation of the reservation of the resource of the receiver candidate other than the receiver candidate actually determined as the receiver among the receiver candidates whose judgment result is passed at the time of the receiver candidate qualification judgment, 8. The dynamic load balancing method for a parallel computer according to claim 7, wherein unnecessary consumption of resources of processors other than the receiver is avoided, and processing efficiency of the entire system can be improved.