JPH0594425A - Task managing method using multiple list - Google Patents

Abstract

PURPOSE:To shorten a time required for a load distribution by constituting a queue of a low-order list in which low-order cells are connected, pointers indicating the head and the tail, multiple list constituted of a high-order list in which high-order cells are connected, and pointers indicating the head and the tail. CONSTITUTION:At the time of taking out the fixed number B of tasks from the queue, task sources 1-1 to 1-S store tasks information in the low-order cell, and connects the low-order cell with the low-order list, when the tasks are prepared. When the number of the tasks is turned to B, the task sources 1-1 to 1-S store the pointers indicating the head and the tail in the high-order cell, connects the high-order cell with the high-order list, and repeats the processing of changing the point indicating the tail so that the connected high- order cell can be pointed. Then, a processors 3-1 to 3-P retrieves the high-order cell positioned at the head of the queue by using the pointer indicating the head at the time of the completion of the entire taken-out tasks, takes out the entire tasks of the connected low-order list, and repeats the processing of changing the pointer indicating the head so that the next high-order cell can be indicated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for managing a large number of tasks in a queue in a system for parallel processing a large number of tasks that can be independently executed. The present invention relates to a method for storing and retrieving task management information that reduces the number of times the list is traversed when retrieving a plurality of tasks from a queue at one time.

[0002]

2. Description of the Related Art FIG. 11 shows S (≧ 1) task sources (9-
Independently executable tasks generated from 1 to 9 to A) are distributed and processed by P (> 1) processors (11-1 to 11-P) using the task distribution means (10). It is a conceptual diagram of a general parallel processing system.

The tasks generated by the task sources (9-1 to 9-S) are managed by putting task management information in a queue in the task distribution means (10).

D. Knuth: The Art of Computer Program
According to ing, Addison-Wesley, Reading. Mass., 1973, there are ring representations using arrays and lists as methods of realizing cues.

The ring representation using an array has a problem that the maximum value of the queue length is fixed and it is impossible to manage the number of tasks exceeding the maximum value.

On the other hand, the use of a list allows the number of tasks that can be managed to be dynamically changed by adding cells.

When a queue is formed by a list, a list is formed by connecting cells storing task management information, a pointer indicating the beginning of the list, and a pointer indicating the end of the list.

When connecting the generated task to the queue, the cell storing the task information is connected to the beginning or end of the list, and the pointer indicating the beginning or end of the list is changed to point to the connected cell. It was done by things.

When the task is taken out from the queue for allocation, the pointer at the head of the list is used to search for the cell at the head of the queue, and the pointer at the head of the list points to the next cell. It was done by changing so.

A conventional load distribution technique in a parallel processing system as shown in FIG. 11 is Yung-Terng Wang, Robert JT M.
orris: Load Sharing in Distributed Systems, IEEE
Transactions on Computers, Vol. C-34, No. 3, pp.
Classified and organized by 204-217, Mar. 1985.

When the task distribution means (10) distributes one task as a unit, the processors (11-1 to 11-1)
1-P) idle time can be shortened, but the number of times of load distribution is increased, so that information required for load distribution, communication overhead, and exclusive control that arbitrates access to the same queue by multiple processors Overhead time increases.

As a method for solving this problem, a fixed number B
A method for collectively distributing (> 1) tasks is shown in the above-mentioned document.

When the total number of tasks is known, the number of tasks to be distributed at one time is increased at first and decreased as the processing progresses to reduce the number of times of load distribution, thus reducing the overhead. A method of doing so is proposed in Japanese Patent Application No. 2-413633 (Japanese Patent Laid-Open Publication No. HEI-6) as a load distribution method for a parallel processing system.

In the above load distribution method, it is necessary to take out a plurality of pieces of task information from the queue at once.

When a plurality of tasks are fetched from the queue at one time, a pointer at the head of the list is used to search for a cell located at the head of the queue, and the cells are sequentially traced until the number reaches the desired number. This is done by changing the pointer indicating the beginning of the list to point to the cell next to the last fetched cell.

[0016]

In the above load distribution method, it is necessary to take out a plurality of pieces of task information from the queue at once.

However, in order to extract a plurality of task management information from the list at once, it is necessary to trace the pointer several times for the tasks to be extracted.

Therefore, there is a problem that the pointer tracing time becomes long when a large amount of task information is taken out, and the time required for load distribution becomes long.

Further, when a plurality of processors access the queue, the time for locking and exclusively using the queue becomes long, the contention between the processors becomes large, and the waiting time for acquiring the lock becomes long. There was a point.

An object of the present invention is to shorten the time required for load distribution by reducing the number of times the pointer is traced when retrieving a large number of task information.

[0021]

In order to achieve the above object, the present invention provides S (≧ 1) task sources, independently executable tasks generated from the task sources, and P (> 1). 1)
In a parallel processing system comprising individual processors and task distribution means for distributing a desired number of tasks to individual processors,

When managing tasks waiting to be executed in the task distribution means by a queue,

A lower list (8) connecting lower cells storing management information of the task, a pointer (7-1) indicating the head of the lower list, and a pointer (7-2) indicating the end of the lower list. ), And a higher list (4) that connects upper cells (7) that store the number of tasks (7-3) connected to the lower list, and a pointer (5 indicating the beginning of the upper list). ) And a pointer (6) indicating the end of the upper list,

When the generated task is connected to the queue, the lower list is created by the lower cells storing the task information, and the pointers indicating the beginning and end of the lower list are connected to the lower list. By creating an upper cell that stores the number of tasks, connecting to the beginning or end of the upper list, and changing the pointer indicating the beginning or end of the upper list to point to the connected upper cell,

When retrieving tasks from the queue for task distribution,

The pointer indicating the head of the upper list is used to search for the upper cell located at the head of the queue,

If the number of tasks to be fetched exceeds the number of tasks linked to one upper cell, all tasks in the lower list linked to the upper cell are fetched and the number of tasks linked to the lower list Is set to 0, and a number obtained by subtracting the number of tasks connected to the upper cell from the number of fetched tasks is set as a new task number to be fetched. Change to point to the cell, repeat the ejection operation,

If the number of tasks to be extracted is less than or equal to the number of tasks connected to one upper cell, lower cells of the number of tasks to be extracted are extracted from the lower list connected to the upper cell and connected to the upper cell. It is characterized in that a number obtained by subtracting the number of tasks to be extracted from the number of existing tasks is set as a new value of the number of tasks connected to the upper cell to perform extraction.

[0029]

According to the task management method of the present invention, by managing the task information in a multiple list, when a large number of tasks are taken out at one time, a list having a length exceeding the number of tasks linked to the lower list is traced. Can be taken out without.

Embodiments will be described below with reference to the drawings.

[0031]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a diagram showing an embodiment of the present invention.
1-S is a task source, 2 is a queue composed of multiple lists of the present invention, and 3-1 to 3-P are processors.

FIGS. 2 and 3 are diagrams showing an upper list, an upper cell and a lower list which form the queue of the present invention.

A lower list 8 connecting lower cells storing task management information, a pointer 7-1 indicating the head of the lower list, and a pointer 7-2 indicating the end of the lower list.
A multiple list consisting of an upper list 4 connecting upper cells 7 storing the number of tasks 7-3 connected to the lower list, a pointer 5 indicating the head of the upper list, and the end of the upper list. A queue is constituted by the pointer 6 shown.

In the following, the processor 3-1 to 3-P uses two types of cases in which a fixed number B (> 1) of tasks are collectively taken out from the queue and when the number of tasks to be taken out at one time is changed. Examples of the invention will be described.

First, a case where a fixed number B (> 1) of tasks are collected from the queue will be described.

When the task is generated, the task sources 1-1 to 1-S store the task information in the lower cells and connect them to the lower list.

When the number of tasks reaches B, pointers indicating the beginning and end of the lower list are stored in the upper cell.

Since the number of tasks linked to the lower list is always B, it is not necessary to store the number of tasks linked to the lower list in the upper cell.

Finally, the upper cell is connected to the end of the upper list, and the pointer indicating the end of the upper list is changed to point to the connected upper cell.

The above processing is repeated.

When the processors 3-1 to 3-P have finished processing all the extracted tasks, the processors 3-1 to 3-P search for the upper cell located at the head of the queue using the pointer indicating the head of the upper list, and search for one lower Since the number of tasks connected to the list is always B, all tasks in the lower list connected to the upper cell are taken out and the pointer indicating the beginning of the upper list is changed to point to the next upper cell.

The above processing is repeated.

As described above, when the number B of tasks to be taken out at a time is known in advance, the lower list is traced by setting the number of tasks connected to one lower list when connecting the tasks to the queue to B. You can retrieve tasks without having to.

Next, a case where the number of tasks taken out from the queue at one time is changed will be described.

When the task is generated, the task sources 1-1 to 1-S store the task information in the lower cells and connect them to the lower list.

Then, the pointers indicating the beginning and the end of the lower list and the number of tasks connected to the lower list are stored in the upper cell.

There is no limitation on the number of tasks linked to the lower list, and for example, it may be a fixed number or the number of tasks generated during a fixed time.

Finally, the upper cell is connected to the end of the upper list, and the pointer indicating the end of the upper list is changed to point to the connected upper cell.

The above processing is repeated.

The processors 3-1 to 3-P take out the tasks from the queue 2 after finishing the processing of all the taken out tasks.

FIG. 4 is a flow chart showing an example of the procedure for extracting the task of the present invention.

(1) Pointer 5 indicating the beginning of the upper list
Is used to search for an upper cell located at the head of the queue (steps 111 and 112).

(2) The number of tasks 7-3 connected to the lower list of the upper cell is compared with the number of tasks to be taken out, and the number of tasks to be taken out is the number of tasks 7-connected to the lower list.
If it exceeds 3, the following procedure (3), the number of tasks to be taken out is 1/2 or more of the number of tasks 7-3 connected to the lower list and the number of tasks 7-3 or less connected to the lower list. For example, if the number of tasks to be taken out is less than 1/2 of the number of tasks 7-3 linked to the lower list, the following procedure (4) is executed (steps 121 to 12).
3).

(3) All the tasks in the lower list 8 connected to the upper cell are taken out (step 131), and the number of tasks 7-3 connected to the lower list in the upper cell is subtracted from the number of tasks to be taken out. Is set as the number of tasks to be newly extracted (step 132), the number of tasks 7-3 connected to the lower list is set to 0 (step 133), and the link is followed to move to the next higher cell (step 134).
The pointer 5 indicating the head of the upper list is changed to point to the cell (step 135), and the procedure (2) is executed.

(4) The number of lower cells obtained by subtracting the number of tasks to be taken from the number of tasks 7-3 connected to the lower list from the head of the lower list 8 connected to the upper cell is traced (step 141). ), From the lower cell to the lower list 8
Until the end (step 142), the number obtained by subtracting the number of tasks to be taken from the number of tasks 7-3 connected to the lower list is set as a new value of the number of tasks 7-3 connected to the lower list (step). 143).

(5) From the head of the lower list 8 connected to the upper cell, the lower cells of the number of tasks to be extracted are extracted (steps 151 and 152), and the number of tasks 7-3 connected to the lower list is extracted. The number obtained by subtracting the number of tasks to be taken out is set as a new value of the number of tasks 7-3 connected to the lower list (step 153).

The above processing is repeated.

In this example, in order to reduce the number of times the lower list is traced, if the number of tasks to be taken out is 1/2 or more of the number of tasks connected to one upper cell, the task is connected to the rear of the lower list. I tried to take out the task that is present, but if it is more problematic that the task that is taken out from the task at the back of the list does not become the FIFO rather than shortening the time to follow the list, follow the above procedure (2). You can do the following:

(2 ') The number of tasks 7-3 connected to the lower list of the upper cell is compared with the number of tasks to be extracted, and the number of tasks to be extracted is the number of tasks 7 connected to the lower list.
If it exceeds -3, the procedure (3) is executed. If the number of tasks to be extracted is 7-3 or less, the procedure (5) is executed.

That is, in FIG. 4, step 123 is directly passed to step 151.

Finally, as a concrete example, an example of the number of times the list is traced when the number of tasks to be taken out from the queue at one time is changed is shown.

The number of tasks whose task source is connected to the lower list is 1
Fixed at 6.

The procedure (2) in which the number of times of tracing the list is small is used instead of the procedure (2 ') in the procedure for the processor to take out the task.

First take 30 tasks, then 2
An example of extracting 0 tasks is shown in FIGS.

(1-1) FIGS. 5 to 6

Since the number of tasks 30 to be taken out exceeds the number of tasks 16 connected to the head upper cell, all 16 tasks are taken out.

The number of connected tasks is set to 0, the link is followed to move to the next upper cell, and the pointer indicating the head of the upper list is changed to point to the cell.

It is necessary to retrieve 14 more tasks.

(1-2) FIGS. 6 to 7

Since the number of tasks to be taken out 14 is 1/2 or more of the number of tasks 16 connected to the upper cell and less than the number of connected tasks, the lower list is traced 16-14 = 2 times, and the remaining 14 tasks are executed. Is extracted, the number of connected tasks is set to 2, and the extraction is completed.

(2-1) FIGS. 7 to 8

Since the number of tasks 20 to be taken out exceeds the number of tasks 2 connected to the head upper cell, all two tasks are taken out.

The number of connected tasks is set to 0, the link is followed to move to the next upper cell, and the pointer indicating the head of the upper list is changed to point to the cell.

It is necessary to retrieve 18 more tasks.

(2-2) FIGS. 8 to 9

Since the number of tasks to be extracted 18 exceeds the number of tasks 16 connected to the upper cell, all 16 tasks are extracted.

The number of tasks connected is set to 0, the link is followed to move to the next upper cell, and the pointer indicating the head of the upper list is changed to point to the cell.

Two more tasks need to be retrieved.

(2-3) FIGS. 9 to 10

Since the number of tasks 2 to be taken out is less than 1/2 of the number of tasks 16 connected to the upper cell, two tasks are taken out from the head of the lower list and the number of connected tasks is 1
As a result of 4, the taking out is completed.

[0081]

As described above, according to the task management method of the present invention, when a plurality of pieces of task information are fetched from the task queue at one time, the number of times the list is traced can be reduced, so the time required for task fetching can be reduced. Can be shortened.

Therefore, the competition when a plurality of processors take out a task from a shared queue can be reduced.

Further, when the tasks generated by a plurality of task sources are managed by a shared queue, the upper cells can be created independently by each task source, and therefore the queue can be accessed only when the upper cells are connected to the upper list. Since it is sufficient, the contention when connecting the task information to the queue can be reduced.

Therefore, the time required for task management can be shortened.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an embodiment of the present invention.

FIG. 2 is a diagram of an upper level list of the present invention.

FIG. 3 is a diagram of an upper cell and a lower list of the present invention.

FIG. 4 is an example of a processing flow of the present invention.

FIG. 5 is a diagram showing a specific example of the present invention.

FIG. 6 is a diagram showing a specific example of the present invention.

FIG. 7 is a diagram showing a specific example of the present invention.

FIG. 8 is a diagram showing a specific example of the present invention.

FIG. 9 is a diagram showing a specific example of the present invention.

FIG. 10 is a diagram showing a specific example of the present invention.

FIG. 11 is a conceptual diagram of a general parallel processing system.

[Explanation of symbols]

1-1 to 1-S task source 2 queues 3-1 to 3-P processor 4 upper list 5 pointer to the beginning of the upper list 6 pointer to the end of the upper list 7 upper cell 7-1 to the beginning of the lower list Pointer 7-2 Pointer indicating the end of the lower list 7-3 Number of tasks connected to the lower list 8 Lower list connecting lower cells storing management information of tasks 9-1 to 9-S Task source 10 Task allocation Means 11-1 to 11-P Processor

Claims (1)

[Claims] 1. S (≧ 1) task sources, independently executable tasks generated from the task sources, and P (≧ 1)
In a parallel processing system including individual processors and a task allocation unit that allocates a desired number of tasks to the individual processors, the task allocation unit manages the tasks when the tasks waiting for execution are managed by a queue. A lower list (8) connecting lower cells storing information, and a pointer (7
-1) and a pointer (7-
2) and the number of tasks connected to the lower list (7-
Queued by a multiplex list composed of an upper list (4) connecting upper cells (7) storing 3), a pointer (5) indicating the head of the upper list, and a pointer (6) indicating the end of the upper list. And connecting the task generated by any of the S task sources to the queue, the lower list is created by the lower cells storing the task information, and the beginning and end of the lower list are An upper cell storing the number of tasks connected to the pointer and the lower list is created, connected to the beginning or end of the upper list, and a pointer indicating the beginning or end of the upper list is pointed to the connected upper cell. When the task is taken out of the queue in order to distribute the task to any of the P processors, the top of the upper list is indicated. The upper cell located at the head of the queue is searched using the interface, and if the number of tasks to be extracted exceeds the number of tasks connected to one upper cell, the tasks in the lower list connected to the upper cell All are taken out, the number of tasks connected to the lower list is set to 0, and the number of tasks taken out is subtracted from the number of tasks connected to the upper cell. Move to the upper cell, change the pointer indicating the beginning of the upper list to point to that cell, repeat the fetch operation, and if the number of tasks to fetch is less than or equal to the number of tasks linked to one upper cell, From the lower list connected to the upper cell, the lower cells of the number of extracted tasks are extracted, and the number obtained by subtracting the number of tasks to be extracted from the number of tasks connected to the upper cell is set to the upper cell. A task management method using a multiple list characterized by setting a new value for the number of connected tasks.