JPH1049387A - Method and device for buffer management - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the performance by effectively using a common memory and not performing common exclusive control as much as possible when hash- clusterd buffers are secured in the common memory. SOLUTION: This management device has a buffer management means 40 which assigns buffers to a free area by managing the free area in the common memory 11, a splitting means 20 which hashes records sent out by an arbitrary task and determines a hash cluster, and stores the records in the buffers in the hash cluster, an owner change means 30 which give the owner right of the buffers secured in the common memory to the task, and makes the task transfer the owner right of the buffers and forms a free area in the common area unless there is a free area in the common memory.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a buffer management method and apparatus, and more particularly, to a buffer management method and apparatus for hash clusters executed in parallel.

[0002]

2. Description of the Related Art In managing a buffer on a shared memory accessed by a plurality of tasks, various schemes have been proposed to avoid contention. For example,
According to JP-A-5-173953, a buffer is allocated to each processor (task), and contention is avoided without performing exclusive control. Further, since the buffer management is limited to one specific processor, there is no need to perform synchronization control for avoiding contention.

[0003]

The conventional buffer management device as described above has a disadvantage that the shared memory is wasted. That is, if a buffer is allocated for each processor, it is necessary to allocate a buffer for each processor in a cluster unit in the hash clustered buffer management used in the hash join process. This is because it increases by the number of buffers.

In a system that performs exclusive control each time a buffer is referenced and updated, the processing performance of the system may be significantly reduced. In other words, the original business process is postponed due to system overhead for exclusive control or waiting.

Accordingly, an object of the present invention is to effectively use a shared memory when a hash-clustered buffer used for hash join is secured on the shared memory.
An object of the present invention is to improve performance by minimizing shared exclusion control as much as possible.

[0006]

According to a buffer management method of the present invention, in a method of managing a buffer in a hash-clustered shared memory used by a plurality of tasks, the task that has secured the buffer has the ownership of the buffer. The task that holds and owns the ownership exclusively accesses the buffer, and when it is not possible to secure a new buffer on the shared memory, transfers the ownership of the buffer to another task that has already secured the buffer. And the task that has obtained the ownership is realized by exclusively accessing the buffer.

In a buffer management device according to the present invention, a buffer management device in a hash-clustered shared memory used by a plurality of tasks manages a free area in the shared memory and stores a buffer in the free area. Buffer managing means for allocating, split means for hashing a record sent by an arbitrary task to determine a hash cluster and storing the record in a buffer in the hash cluster, and assigning the ownership of the buffer secured in the shared memory to the task. Owner change means for transferring the ownership of the buffer to the task when there is no free area in the shared memory and creating a free area in the shared memory, and controlling the queuing between a plurality of tasks. And a queuing management unit.

The buffer management device according to the present invention further comprises an owner update request flag for requesting that a task that cannot secure a buffer because there is no free area in the shared memory transfers the ownership of the buffer to a task that has already secured the buffer. Wherein the owner changing means detects the owner update request flag and starts a process of creating a free area in the shared memory.

Further, in the buffer management device according to the present invention, when the buffer cannot be secured in the determined hash cluster, the split means instructs the owner changing means to release a buffer held by another task.

Further, the buffer management device of the present invention has a pointer for the buffer in the hash cluster for each task.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. In the figure, a buffer management device according to the present invention comprises a shared memory 11, buffers 10 and 14.
, Cluster 12, owner update request flag 13, splitting means 20, owner changing means 30, buffer managing means 40, searching means 50, queuing managing means 60, processing tasks 70 and 71, pointer 90 And 91, and databases 80 and 81.

The buffer 10 is for a processing task 70 created on the shared memory 11 included in the cluster 12.

The buffer 14 is for a processing task 71 created on the shared memory 11 included in the cluster 12.

The owner update request flag 13 is
Is a flag indicating whether or not a request to change the owner has been made.

The split means 20 hashes a key value in the data passed from the search means 50 and determines a cluster to be stored.

The owner changing means 30 changes the ownership of the buffer by referring to the owner update request flag 13 and the queuing management means 60.

The buffer management means 40 manages the shared memory 11 using a conventional technique, and secures a buffer.

The search means 50 searches the database using a conventional method.

The queuing management means 60 controls queuing between tasks using a conventional technique.

A processing task 70 and a processing task 71 are execution units for searching the database 80 and the database 81 in parallel, respectively. Each task has a pointer 90 and a pointer 91 which are pointers to a buffer which is the owner.
Contains.

The split unit 20 obtains data from the search unit 50, appropriately hashes key values, and determines a cluster to be stored. The splitting means 20 performs the processing task 70
In the case of the operation, the pointer 90 is checked, and if not, the pointer 91 is checked, and it is checked whether or not the pointer is set. If the pointer is not set, the buffer management means 40 is called to secure a buffer, set the pointer, and store the data in the buffer obtained from the pointer 90 or 91.

Further, the split means 20 calls the buffer management means 40 even when the buffer 10 has no free space and cannot store data, newly secures a buffer, sets a pointer 90, and stores the data. If a new free buffer cannot be created, the owner changing means 30 is called to request a change in the owner of the buffer secured by another processing task.

In the above series of processing, if there is no free space in the shared memory 11 and a buffer cannot be secured, the owner changing means 30 is called to change the owner of the buffer secured by another processing task. Call.

The owner changing means 30 calls the waiting management means 60 after setting the owner update request flag 13, and
Wait until the buffer owner gets permission to use it. When the use permission is obtained, the pointer 90 is set, and the process is returned to the split unit 20.

On the other hand, the split means 20 checks the owner update request flag 13 at an appropriate timing, and when there is an owner change request, calls the queuing management means 60 to cancel the queuing, deletes the pointer 90, and hands over the owner.

FIG. 2 is a flow chart showing the operation of the split means 20. Referring to FIG.
In step 1, the key value in the record passed from the search unit 50 is hashed to determine a cluster, and the process proceeds to step 202 (in this embodiment, the cluster 12 is determined).

In step 202, it is checked whether or not there is an owner change request by referring to the update request flag in the cluster determined in step 201. If there is an update request, the process proceeds to step 203. Otherwise, go to step 204.

In step 203, the owner changing means 30
And transfers the ownership of the buffer of the invoking task to another task, and proceeds to step 204.

In step 204, a pointer on the processing task corresponding to the cluster 12 determined in step 201 is checked to check whether the pointer is not NULL. Here, the pointer 90 is referred to when the splitting means 20 is operating in the processing task 70, and the pointer 91 is referred to when the splitting means 20 is operating in the processing task 71. If the buffer pointer is not NULL, the process proceeds to step 208. Otherwise, go to step 205.

In step 205, the buffer management means 40
To request a new buffer and go to step 206.

In step 206, it is checked whether the buffer requested in step 205 has been secured. If a buffer is secured, the process proceeds to step 208. If not, the process proceeds to step 207.

In step 207, another task calls the owner changing means 30 in order to obtain the owner's buffer, and proceeds to step 208.

In step 208, the record is transferred to the buffer indicated by the pointer, and the splitting means 20 ends. At this time, the split means 20 sets the processing task 7
When operating at 0, the pointer 90 is referenced, and when operating at the processing task 71, the pointer 91 is referenced.

FIG. 3 is a flowchart showing the operation of the owner changing means 30 described above. Referring to FIG.
In step 1, the type of owner change is determined. If the owner change type is the owner right surrender, the process proceeds to step 302. Otherwise, go to step 306.

In step 302, the contents of the pointer of the invoking task are copied to the pointer of the task waiting for surrender. Specifically, the owner changing unit 30 sets the processing task 70
, The contents of the pointer 90 are stored in the pointer 9
Copy to 1. If the processing task 71 is operating, the contents of the pointer 91 are copied to the pointer 90.

In step 303, the owner changing means 30
NUL sets the buffer pointer of the processing task where
Set to L, and proceed to step 304.

In step 304, the owner update request flag 13 in the cluster 12 is reset, and in step 305
Move on to

At step 306, the owner update request flag 13 is set, and the routine goes to step 307.

In step 307, the waiting management means 60 is called, waiting until the ownership is transferred, and when the waiting is released, the owner changing means 30 ends.

In step 305, the waiting management means 60 is called to cancel the waiting of the processing task requesting the release of the ownership, and the owner changing means 30 is terminated.

The search means 50 searches the database by a conventional technique, and transfers the record to the split means 20.

The buffer management means 40 manages the space in the shared memory 11 by a conventional technique to secure a buffer. If it cannot be allocated, it returns an error. If it can be allocated, it returns a pointer to a buffer.

Further, the queuing management means 60 controls queuing of the processing task by a conventional method (for example, semaphore control).

[0045]

As described above in detail, according to the present invention, the buffer is exclusively used by the processing task because the owner changing means does not operate until the shared memory is full. Therefore, it is not necessary to perform exclusive control every time a record is transferred to the buffer, and the split processing can be executed.

If there is no free space in the shared memory, the ownership is changed and the record is transferred to the buffer.
The shared memory can be used effectively.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of a split unit.

FIG. 3 is a flowchart showing the operation of the owner changing means.

[Explanation of symbols]

 10, 14 Buffer 11 Shared Memory 12 Cluster 13 Owner Change Request Flag 20 Split Means 30 Owner Change Means 40 Buffer Management Means 50 Search Means 60 Queue Management Means 70, 71 Processing Tasks 80, 81 Database 90, 91 Pointer

Claims (5)

[Claims] 1. A method for managing a buffer in a hash clustered shared memory used by a plurality of tasks, wherein a task that has secured a buffer holds the ownership of the buffer, and the task that holds the ownership is the task that holds the ownership. When the buffer is exclusively accessed and a new buffer cannot be secured on the shared memory, the task that has requested the transfer of the ownership of the buffer to another task that has already secured the buffer and has acquired the ownership is A buffer management method, wherein the buffer is exclusively accessed. 2. A buffer management device for managing a buffer in a hash clustered shared memory used by a plurality of tasks, wherein the buffer management means manages a free area in the shared memory and allocates a buffer to the free area. Split means for hashing a record sent by the task to determine a hash cluster and storing the record in a buffer in the hash cluster, and assigning and managing the task with the ownership of the buffer secured in the shared memory, and When there is no free area in the memory, there are owner changing means for transferring ownership of the buffer to the task to create a free area in the shared memory, and queuing management means for controlling queuing between a plurality of tasks. A buffer management device, characterized in that: 3. The buffer management apparatus according to claim 2, wherein a task that cannot secure a buffer because there is no free area in the shared memory requests the task that has already secured the buffer to transfer the ownership of the buffer. A buffer management device comprising an update request flag, wherein the owner change unit detects the owner update request flag and starts processing for creating a free area in the shared memory. 4. The buffer management device according to claim 2, wherein the split unit releases the buffer held by another task to the owner changing unit when a buffer cannot be secured in the determined hash cluster. A buffer management device. 5. The buffer management device according to claim 2, wherein a pointer pointing to a buffer in the hash cluster is provided for each task.