JPH02171952A - Dispatching system for multiprocessor - Google Patents

Abstract

PURPOSE:To reduce the lock competition of a task control queue by eliminating task search processing for selecting a task to be successively dispatched in dispatch processing following POST processing and WAIT processing. CONSTITUTION:For a processor 902 and a processor 903, exclusive control processing such as task control queue 200, resource waiting queue 600 and processor waiting queue 400 and queue relating to dispatch processing are stored in a common memory 901. At the time of POST processing, a POST receiving task is dispatched without executing task search processing and a POST generating task losing the processor using right is registered in the processor waiting queue, and at the time of WAIT processing, the POST generating task registered in the processor waiting queue is dispatched and the task search processing is not executed. Consequently, the locking time of the task control queue at the time of dispatch processing can be shortened and overhead due to the lock competition of the task control queue can be reduced.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a multiprocessor system in which a plurality of processors share resources, and particularly relates to a multitasking dispatch method.

[Prior art] A processing unit (task) executed within a computer system is
Use various shared resources. In a system such as a multiprocessor system in which a plurality of tasks operate simultaneously, contention (lock contention) regarding the right to exclusively use a shared resource occurs among these tasks, so exclusive control is required for this purpose. For example, when multiple tasks are processing sales in an online system, different tasks refer to and update current sales j stored in a common storage area. While a task is referencing and updating this "current sales amount", it is necessary to prohibit another task from accessing this data.

Many of the computer systems currently in practical use perform this exclusive control using the WAIT-PO3T method (Suspend φResume (Suspend lRe5u+++e) method,
Block-Activa
te) method, sleeve wake-up (Sleep-
This is realized by the Wakeup) method. Regarding the WAIT・pos'r method, please refer to Wang (K, uwan) S Briggs (F, A, Br) published by McGraw Hill.
1gg5) Co-authored “Computer Architecture and Parallel Processing (Con+puter A
architecture and ParallelP
rocessing) J.

The WAIT-POST method consists of a WAIT process that puts task 2, which has failed in occupying the shared resource, into a "waiting state" when the shared resource is being occupied by task 1, and task 3 (task 3) that releases the shared resource. 3=Task 1) performs inter-task exclusive control of the shared resource exclusive usage right by POST processing that releases the [waiting state] of Task 2.

Hereafter, a lock is taken to obtain the exclusive right to use the shared resource. Expressed as securing lock rights, etc.

Further, when an attempt is made to take a lock, but the lock cannot be taken because the task is occupied by another task, this is expressed as a lock failure, lock acquisition failure, etc.

When each task fails to lock, it transfers control to the WAIT processing routine. The WAIT processing routine puts the task in which the lock has failed into a "wait state" and calls the dispatcher. The dispatcher assigns another task in the "ready state" to the processor that was executing the task in the "waiting state". When there are multiple tasks in the "ready state", the dispatcher selects which task in the "ready state" to allocate a processor to, depending on the execution priority set in advance for each task and the dispatch processing algorithm. Dependent.

On the other hand, each task checks whether there is a "waiting" task (WAIT task) immediately before releasing (unlocking) the shared resource, and if there is a WAIT task, it passes control to the PO5TO5 task. For POSTO8, the WAIT task is returned to the "ready state". The task returned to the "executable state" is later assigned a processor (1113) by the dispatcher, obtains the right to exclusively use the shared resource, and resumes processing.

Below, the task that starts the POST process is the POST issuing task, and the W that is released from the "waiting state" by the POST process.
The AIT task is called the PO5T receiving task. POS
T-issue task, POST acceptance task, WAIT task, etc. are not unique names of each task. Each task may be a POST issuing task or a POST receiving task.

[Problems to be Solved by the Invention] However, this method has a problem in that the number of dynamic steps required to execute a task increases due to lock contention between tasks. That is, in a multiprocessor, the number of dynamic steps required for execution per task increases compared to when a task is executed by a single processor.

This increase in the number of dynamic steps is due to WAI due to lock contention.
These are an increase in overhead for T processing and POST processing (primary factor), and an accompanying increase in overhead for dispatch processing (secondary factor).

The primary cause is that as the number of processors increases, the number of tasks operating simultaneously increases, resulting in an increase in the lock acquisition failure rate. This increase in the lock acquisition failure rate increases the overhead of retry processing, that is, WAIT-POST processing.

A secondary factor is that task switching increases in proportion to the increase in WAIT-POST processing, and the number of dispatch processing increases. The dispatcher locks the task management queue that centrally manages control information such as the status and execution priority of all tasks in the system until it selects the next task to be dispatched from among multiple ready-to-run tasks. There must be. This lock is essentially unavoidable for updating the task management queue.

In a multiprocessor system, dispatchers operate on multiple processors simultaneously, so an increase in the number of dispatch processes leads to lock contention in the task management queue. This lock contention increases the number of steps required for dispatch processing, further increasing the number of dynamic steps per task.

SUMMARY OF THE INVENTION An object of the present invention is to reduce lock contention in a task management queue that occurs due to an increase in the number of dispatch processes, and to solve the problem of an increase in dispatch overhead.

[Means for solving the problem] In order to solve the above problem, so-called WAIT-PO
In a multitasking multiprocessor system that uses the ST method to perform exclusive inter-task control of exclusive usage rights to shared resources, after the PO5T receiving task is transitioned to an executable state by POST processing, the POST issuing task is executed. The POST issuing task is interrupted, the control information regarding this POST issuing task is registered in the storage means (processor waiting queue), and the processor that was executing the POST issuing task is assigned to the POST receiving task.

Furthermore, the task (WAIT issuing task) that has failed to secure the exclusive right to use the shared resource is placed in a waiting state by WAIT processing, and the POST issuing task registered therein is retrieved from the storage means (processor wait queue). Extract the control information and apply W to this POST issuing task.
Assign the processor that was executing the AIT issue task.

[Operations] The method of the present invention makes it possible to omit the process of selecting the next task to be dispatched from among a plurality of executable tasks (hereinafter referred to as task search process) in the dispatch process that follows the POST process and the WAIT process. can.

This reduces the lock contention of task management queues by shortening the locking time of the same task management queue.

That is, at POST8, the POST receiving task is dispatched without performing task search processing, and the POST issuing task that has lost the right to use the processor is registered in the processor waiting queue. During WAIT processing, POST issued tasks registered in the processor waiting queue are dispatched, and task search processing is not performed. As described above, the locking time of the task management queue during dispatch processing can be shortened, and the overhead caused by lock contention of the task management queue can be reduced.

[Example] An example of the present invention will be described below with reference to FIGS. 1 to 5. This embodiment is an example of an online database system.

Before describing the dispatch method of the present invention, an overview of the online database system of this embodiment will be explained. FIG. 1 shows a schematic configuration of the system. A first processor 902 and a second processor 903 are coupled via a shared memory 901. The two processors can each run a program on the shared memory 90] and refer to data on the shared memory 901 to perform processing in parallel. Task management queue 200. Queues related to exclusive control processing and dispatch processing, such as a resource waiting queue 600 and a processor waiting queue 400, are also stored on the shared memory 901. Each processor can send and receive data to and from the terminal device 906 via the communication control device 904 . Each processor also uses a database 907 on the disk via the disk controller 905.
can be accessed.

Hereinafter, in this specification, a group of data reference/update processing in an online database system will be referred to as a transaction. For example, "In response to a deposit withdrawal request, check the deposit balance, subtract the withdrawal amount from the deposit balance,
A series of processes such as "updating the deposit balance in the database" is called a transaction. The transaction is
It starts with data transmission from the terminal device 906 as a trigger.

Next, the basic task management method will be explained. When a transaction is entered, the supervisor creates one or more tasks corresponding to this transaction. Specifically, when a supervisor generates a task, the supervisor creates a task control block T C that stores control information for each task.
B (Task Control Block) 300
These TCBs are registered and managed in the task management queue 200 until the end of the transaction.

Each task being registered in the task management queue 200.

It can be in any of the following states: "wait state", "ready state", "suspended state", or "execution state".

A task in a "waiting state" cannot be executed by assigning a processor until the event that caused it to wait is completed and the task is in an "executable state." A task in the “ready state” is
As soon as any processor in the system becomes free, the processor can be assigned and transitioned to the "running state." A task in a “suspended state” is a task that issues a POST macro, transfers the right to use the processor to a POST receiving task, and is registered in the processor waiting queue 400. A task in the "execution state" is a task that has obtained the right to use the processor and is being executed.

The dispatcher allocates processors to tasks that are "ready to run." If there are multiple tasks in the "executable state", a processor is assigned to the selected task according to the execution priority assigned to each task.

Next, the data structure of each queue and management block will be explained. These queues and management blocks are stored in shared memory 90.
1 and accessible from all processors in the system.

The task management queue 200 is a queue for managing information necessary for controlling each task, and as shown in FIG. 2, it is composed of one task management queue header 210 and a chain of zero or more task control blocks TCB 300. . Only one task management queue 200 is provided in the entire system. The task management queue header 210 includes a task management queue lock word 201, a task management queue head pointer 202.
It consists of a task management queue end pointer 203.

The task management queue lock word 201 is used for exclusive access control of the task management queue 200.

The task control block TCB300 stores the following information as shown in FIG.

(1) Task status 301 The task states mentioned above (wait state, executable state).

(suspended or running state).

(2) Execution priority 302 Stores the execution priority assigned to the task.

(3) Next TCB pointer 303 Stores the next TCB address when chaining TCBs to form a task management queue. T at the head of the task management queue
In the case of CB, this field stores 0.

(4) Task information save area 304 Used to save the contents of PSW (program status word), registers, etc. so that the task can be restarted when execution of a task is interrupted.

The resource waiting queue 600 is a queue in which tasks that have failed to lock a shared resource are registered, and as shown in FIG.
Control BLock) 700 and 0 or more lock requests Pro 7riL RB (Lock Reque
stBlock) Consists of 800 chains. One resource waiting queue 600 is provided for each shared resource.

The weight control block WCB700 stores the following information as shown in FIG.

(1) Resource wait queue lock word 701WCB
Used for exclusive control of own updates. 1 if occupied by this WCB
is stored, and 0 is stored during release.

(2) Resource lock word 702 Indicates the occupied state of a shared resource that is subject to exclusive control. Stores 1 when the shared resource is occupied, and stores O when the shared resource is released.

(3) Resource waiting queue head pointer 703 Stores the address of the LRB at the head of the resource waiting queue. If the resource waiting queue is empty, O is stored.

(4) Resource waiting queue end pointer 704 Stores the address of the LRB at the end of the resource waiting queue. If the resource waiting queue is empty, 0 is stored.

(5) Occupied task TC pointer 705 Stores the TCB address of the task occupying the shared resource. This field has no meaning while the shared resource is being released.

The lock request block LRB800 stores the following information as shown in FIG.

(1) E CB (Event Control B
lock) 801 Generally, in the WAIT-POST method, between the posT issuing task and the WAIT task,
This is control information that manages the occurrence or non-occurrence of a waiting event. P
When the O5T bit is O, it indicates that no event has occurred, and when it is 1, it indicates that the event has occurred. When the WAIT bit is O, it is not WAIT, 1
When , it indicates WAIT. In this embodiment, it is used to notify the release of shared resources.

(2) Next LRB pointer 802 Stores the address of the first LRB when chaining LRBs to form a resource waiting queue.

(3) WAIT task TCB pointer 803 current LRB
The TCB address of the WAIT task corresponding to the WAIT task is stored.

The processor waiting queue 400 is a queue in which tasks that have lost the right to use the processor due to PO3T issuance are registered, and as shown in FIG. 4, the processor waiting queue header 4
10 and 0 or more processor request blocks P RB (
Processor Request BLock) 5
It consists of a chain of 00. Only one processor waiting queue 400 is provided in the entire system. The processor wait queue header 410 includes a processor wait queue lock word 401, a processor wait queue head pointer 402,
It consists of a processor wait queue end pointer 402.

The processor wait queue lock word 401 is used for exclusive access control of the processor wait queue 400.

The processor request block PRB500 stores the following information as shown in FIG.

(1) Next FRB pointer 501 When configuring a processor wait queue by chaining FRBs,
Stores the address of the next FRB.

(2) Processor-waiting task TCB pointer 502 Stores the TCB address of the processor-waiting task corresponding to the current FRB.

The above task management queue 200. Resource waiting queue 6
00 and the processor waiting queue 400 are shown in FIG. LRB 800 in resource waiting queue 600. Each PRB 500 in the processor wait queue 400 has a pointer to a TCB 300 in the task management queue 200. By following this pointer, the TCB 300 corresponding to the LRB 800 retrieved from the resource waiting queue 600 can be found without searching the task management queue 200. Further, the TCB 300 corresponding to the PRB 500 taken out from the processor waiting queue 400 is also stored in the task management queue 20.
You can find it without searching for 0.

As a result, at POST8, resource waiting queue 6
Task (P) corresponding to LRB800 extracted from 00
If the right to use the processor is given to the OST receiving task (OST receiving task), dispatch processing can be performed without searching the task management queue 200. Similarly, during WAIT processing, PRB5 taken out from the processor wait queue 400
If the processor usage right is given to the task corresponding to 00 (POST issued task), dispatch processing can be performed without searching the task management queue 200.

Finally, this dispatch method will be explained in detail using FIG.

First, lock processing will be explained. When each task 100 locks a shared resource, it issues a LOCK macro 101 and calls a lock processing routine 103. If the necessary shared resource cannot be secured because it is being exclusively used by another task, the task that issued the LOCK macro 101 is registered in the resource waiting queue 600 (108);
- Call the WAIT processing routine in the supervisor,
The task status 301 in the TCB 300 of this task is
Waiting state" (process 109).

Registration of a task in the resource waiting queue 600 in process 108 means that the LRB 800 pointing to the TCB 300 of this task is
, and connect it to the LRB chain of the resource waiting queue 600. At this time, exclusive control for updating the resource waiting queue end pointer 703 is performed using the resource waiting queue lock word 701.

The task in the "waiting state" is changed to the "ready state" by PO5T processing (process 117) issued by another task.
The "wait state" will continue until it is returned to.

Next, it is checked whether a processor-waiting task, that is, a "suspended" task is registered in the processor-waiting queue 400 (process 110). If there is a task in the “suspended state”, the process 11 takes it out from the processor waiting queue 400.
1) Call the high-speed dispatcher (process 112)
, assigns a processor to this task and puts it in the "execution state" (processing 106).

Retrieval of a task from the processor waiting queue 400 in process 111 refers to the PR pointing to the TCB 300 of this task.
B500 is removed from the PRB chain of the processor wait queue 400. At this time, exclusive control for updating the processor wait queue head pointer 403 is performed using the processor wait queue lock word 401.

If there is no task in "suspended state", call dispatch Y (process 113), task search (process 121)
A task to be dispatched is selected from among the tasks in the "executable state" and a processor is assigned to it (processing 124). This task search (process 113) is a process of sequentially tracing the chain of TCHs connected to the task management queue 200 and searching for a TCB whose task status 301 is "executable". While task search processing is being performed, other processors are using task management queue 2.
In order not to update TCB 300 which is chained to 00,
Exclusive control is performed using the task management queue lock word 201.

A task to which a processor is assigned enters the "execution state" and continues processing until the right to use the processor is taken away due to an interrupt or the like.

Next, unlock processing will be explained. When the task 100 that has been using the shared resource exclusively issues the UNLOCK macro 102 and calls the unlock processing routine 104 to release the shared resource. In the unlock processing routine 104, it is checked whether a task in a "waiting state" is registered in the resource waiting queue 600 (processing 114).

If there are no "waiting" tasks, release the resources and return. If there is a task in the "waiting state", the task is taken out from the resource waiting queue 600 (116), and POST is performed to notify the release of the shared resource.
A command for O8 is called (process 117).

Retrieval of a task from the resource waiting queue 600 in process 116 means that the lock request block LRB 800 pointing to the TCB 300 of this task is taken out from the resource waiting queue 600.
It is to remove it from the LRB 800 chain. At this time,
Exclusive control for updating the resource waiting queue head pointer 704 is performed using the resource waiting queue lock word 70.
This is done using 1.

The POST 8-chin moves the POST issuing task from the "execution state" to the "executable state" and moves the POST receiving task from the "waiting state" to the "executable state". These are performed by updating the task status 301 in the TCB 300 of each task.

Next, move the POST issuing task from the "executable state" to the "suspended state" and register it in the processor waiting queue 400 (process 118), call the high-speed dispatcher (process 119), and allocate a processor to the POST receiving task.
"Execution state" (processing 106).

Registration of a task in the processor waiting queue 400 in process 118 means creating a processor request block PRB 500 pointing to the TCB 300 of this task and connecting it to the chain of PRBs in the processor waiting queue 400. At this time, exclusive control for updating the processor wait queue end pointer 402 is performed using the processor wait queue lock word 401.

The PO5T receiving task that has been given the right to use the processor is
It is possible to monopolize shared resources and resume processing.

[Effect of the invention] As described above, according to the method of the present invention, POST
In dispatch processing that follows processing and WAIT processing, task search processing for selecting the next task to be dispatched can be eliminated, and lock contention in the task management queue can be reduced.

That is, after POST processing, a processor is assigned to a task registered in the resource waiting queue, and after WAIT processing, a processor is assigned to a task registered in the processor waiting queue. Task search processing for searching from the task management queue can be omitted.

As a result, dispatch processing overhead in a multitasking multiprocessor system can be reduced, and the throughput performance of the entire system can be improved.

[Brief explanation of the drawing]

Figure 1 is a diagram showing the relationship among the task management queue, resource waiting queue, and processor waiting queue and the overall system configuration.
Fig. 2 shows the format of the task management queue, Fig. 3 shows the format of the resource waiting queue, Fig. 4 shows the format of the processor waiting queue, and Fig. 5 shows the dispatch method processing of the present invention. It is a figure showing a flow. 200...Task management queue 210...Task management queue header. 300...Task control block TCB (Task C
ontrol BLock). 400...Processor wait queue 410...Processor wait queue header. 500...Processor request block PRB (Pro
cessor Request f31ock). 600...Resource waiting queue 700...Wait control block WCB (w^IT
Control Block). 800...Lock request block LRB (Lock R
equest Block) Chrysanthemum/2 /Iji

Claims (1)

[Scope of Claims] 1. WAIT processing for transitioning task 2, which has failed to secure the right to exclusively use the shared resource because the shared resource was being exclusively used by task 1, to a waiting state, and the shared resource Task 3, which has finished using the shared resource, performs inter-task exclusive control of the exclusive usage rights of the shared resource through POST processing that transitions task 2 to an executable state and dispatch processing that allocates processor usage rights to tasks in the executable state. In a multiprocessor system using a multitasking method, after transitioning task 2 to an executable state by POST processing, execution of task 3 is interrupted, control information regarding task 3 is registered in a storage means, and task 3 is transferred to an executable state. A dispatch method that assigns the currently executing processor to task 2. 2. In the dispatch method according to claim 1, the task 4 that has failed to secure the exclusive right to use the shared resource is dispatched to the WAI
A dispatch method that places the task in a waiting state by processing T, retrieves control information related to task 3 registered in the storage means from the storage means, and assigns the processor that was executing task 4 to task 3.