KR101101038B1 - Flash memory based database management system and method for merging pages for the same - Google Patents

Abstract

The database management system according to the present invention allocates a data storage space and a log storage space to a database including a plurality of nonvolatile blocks, and to each of the blocks, there is a request to update the original data stored in the data storage space. When the change record is stored in the log storage space in the same block as the data storage space in which the original data is stored, and the log storage space is insufficient, the change records stored in the log storage space corresponding to the original data are stored in the original data. It is applied to and merged to generate a new version of the data, and includes a database management unit for storing the generated new version of the data in the data storage space of another block. Therefore, by performing merge operation on a page-by-page basis for data that causes overflow of log storage space, there is an effect of preventing unnecessary data merge operation and improving write performance of the system, thereby preventing shortening the life of flash memory. It can work.

Flash memory, data storage, log storage, blocks, merger

Description

Flash memory based database management system and method for merging pages for the same}

The present invention relates to a flash memory-based database management system (DBMS) and a page merging method therefor. In particular, a merging that can improve the write performance of a system by reducing the occurrence of page merging in a technique of storing updates in a log form. It is about a method.

Flash memory is a non-volatile memory that does not erase stored data even when the power goes out. It is used in various fields because of its small size, low power, and strong durability. In recent years, due to the large capacity and low price, it is possible to store a large amount of data to replace the disk. As a result of the large capacity of the flash memory, a database management system (DBMS) using a flash memory whose operation speed is faster than that of a disk has been actively studied.

The most important feature of flash memory to efficiently design flash memory based DBMS is update operation of data. Flash memory has pages and blocks, which are read and write units, respectively, and have a characteristic of erasing the block to which the page belongs and writing new contents to the block for arbitrary page update. This characteristic is inefficient compared to the disk performing the update by simply overwriting the location of the data to be updated.

Therefore, a storage technique of a flash memory database management system has been studied to solve such an inefficient update method. Conventional Patent Registration No. 900189 "Database Management System and Method Based on Non-Volatile Memory", when an update operation on original data stored in an arbitrary data page (minimum unit of data storage on flash memory) occurs, The change record for the update is stored in a log storage space in the same block as the data storage space in which the original data is stored.

In this case, when the overflow occurs because the log storage space is full, the block performs a block unit merge operation.

The merge operation reads all the change records (stored in the log storage space) containing the original data stored in the arbitrary data storage space and related update information into the main memory, and reflects the change records in the original data to obtain the latest version. After creating the data, it is stored in the empty storage space of flash memory. Block-by-block merging is a method in which the merge is performed by reflecting each corresponding change record on all original data stored in the data storage space in the block, and the latest version of the merged data is stored in the data storage space of the new block. .

That is, as shown in FIG. 1, if an update operation on one of the original data stored in the data storage space of the block A before the merge occurs, the log storage space of the block A is already full and overflow occurs. As a result, a block-by-block merge operation is performed, and the merge operation is performed by applying change records (stored in the log storage space) related to each of the original data stored in all data storage spaces in the block A. FIG. Thereafter, the merged new version of data is sequentially stored in the data storage space of the empty block B, and when the merge operation of all the original data of the block A is completed and the storage to the block B is completed, the block A is dirty. Is erased and the merge operation is completed.

However, the merging method in this manner is inefficient because each data stored in the block is performed according to whether or not the log storage space of the block overflows regardless of the incidence of writing change records in the log storage space. In other words, when there is data in which a frequent update occurs in a block, a large number of change records related to this data are stored in the log storage space, whereby a lot of data may cause unnecessary merging regardless of its update rate.

In addition, when performing block-by-block merging, there may be a problem in that storage space may be wasted as the actual number of data generated by the merge may be smaller than the total number of data storage spaces that can be stored in the block.

An object of the present invention is to provide a database management system and a page merging method that can reduce unnecessary data merge operations and prevent waste of storage space.

The database management system according to the present invention allocates a data storage space and a log storage space to a database including a plurality of nonvolatile blocks, and to each of the blocks, there is a request to update the original data stored in the data storage space. When the change record is stored in the log storage space in the same block as the data storage space in which the original data is stored, and the log storage space is insufficient, the change records stored in the log storage space corresponding to the original data are stored in the original data. It is applied to and merged to generate a new version of the data, and includes a database management unit for storing the generated new version of the data in the data storage space of another block.

The database manager may convert the original data into a dirty state in the data storage space of the previous block when storing the generated new version of data in the data storage space of the other block.

Preferably, the other block is a block having empty data storage space among blocks that are in use or a block that is not in use.

After the database manager stores the new version of data in the data storage space of the other block, if an update request for one original data among other original data stored in the data storage space of the previous block occurs, Apply and merge change records stored in the log storage space of the previous block corresponding to the original data to the one original data to generate a second new version of the data, and generate the second new version of the data; It is preferable to store in the data storage space of another block, and to convert the one original data into the dirty state in the data storage space of the previous block.

The database manager may delete the block if the number of original data which is converted into a dirty state in the data storage space of the previous block is a predetermined number or more.

When the update request for the data of the new version occurs, the database manager keeps the data of the new version as it is and logs a change record related to the data of the new version in the same block as the data storage space in which the data of the new version is stored. Save to storage space.

The database management system may further include a database buffer configured to store, in-memory data, original data accessed by a computing device among the original data, and an update request of the computing device to the in-memory data. And a buffer manager configured to update the in-memory data and store the updated in-memory data in the database buffer, and to store an update record regarding the updated in-memory data in another portion of the database buffer.

In the page merging method according to the present invention, when an update request for the original data stored in the data storage space in an arbitrary block occurs, an overflow occurs in the log storage space in the same block as the data storage space in which the original data is stored. A second step of judging whether a change occurs in the log storage space, and a second step of applying and merging all change records corresponding to the original data for which update is required to the original data to generate a new version of data. And a third step of storing the generated new version of data in a data storage space of another block.

If the overflow does not occur in the first step, the original data required for updating is maintained and the change record corresponding to the original data is stored in a log storage space in the same block as the data storage space in which the original data is stored. It is desirable to.

The third step may include determining whether there is an empty data storage space for storing the new version of data among the in-use blocks, and if there is an in-use block having empty data storage space, the data storage space of the block. Storing the new version of data and storing the new version of data in a data storage space of an empty block that is not in use if there is no empty data storage space.

After the third step, it is preferable to convert the original data of the previous block corresponding to the new version of data stored in the other block into a dirty state that is no longer used.

Further, after the third step, if an update request for one original data among other original data stored in the data storage space of the previous block occurs, the update is stored in the log storage space of the previous block corresponding to the corresponding original data. Applying and merging change records to the one original data to generate a second new version of the data, storing the generated second new version of data in the data storage of the other block, and then the data of the previous block. In the storage space, the one original data is switched to a dirty state.

If the number of original data converted to a dirty state in the data storage space of the previous block is a predetermined number or more, it is preferable to erase the block.

Further, after the third step, if an update request for the new version of data stored in the data storage space of the other block occurs, the new version of the data is kept as it is and the change record related to the new version of the data is recalled. The new version of the data is stored in the log storage space in the same block as the data storage space.

The present invention has an effect of improving the write performance of a system by preventing a merge operation of unnecessary data by performing a merge operation on a page-by-page basis of data that causes an overflow of a log storage space, thereby shortening the lifespan of a flash memory. There is an effect that can be prevented.

Hereinafter, a database management system and a page merging method according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

Figure 2a is a block diagram showing a database management system according to an embodiment of the present invention, Figure 2b is a diagram showing the structure of a database of a database management system according to an embodiment of the present invention.

As shown in FIG. 2A, a database management system according to an exemplary embodiment includes a database 201, a database manager 203, a database buffer 205, and a buffer manager 207.

The database 201 includes a plurality of blocks 210, 220, 230, 240, as shown in FIG. 2B.

The database manager 203 allocates a plurality of data storage spaces and a plurality of log storage spaces to each of the plurality of blocks in the database 201. That is, taking the first block 210 of FIG. 2B as an example, the data storage space is allocated to four pages 211, 212, 213, and 214, and the log storage space is allocated to four sectors 215, 216, 217, and 218. ) Is assigned.

Further, when there is a request for updating the original data stored in the data storage space, the database manager 203 stores a change record in the log storage space in the same block as the data storage space in which the original data is stored. If the log storage space is insufficient, a new version of the data is generated by applying and merging change records stored in the log storage space corresponding to the original data to the original data, and creating the new version of the data in the data storage space of another block. Save it.

The database buffer 205 stores the original data accessed by the computing device among the original data as in-memory data. In this case, the database buffer 205 may be a volatile memory having a shorter access time than the database 310. For example, the database buffer 205 may be dynamic RAM or static RAM.

The buffer manager 207 updates the in-memory data and stores the in-memory data in the database buffer 205 when there is a request for updating the in-memory device. 205) in another part.

At this time, the database manager 203 may write a change record related to the in-memory data as the data storage space in which the original data corresponding to the in-memory data is stored only when the in-memory data is released from the database buffer 205. It can be stored in the log storage space in the block.

For example, if the database manager 203 has a request to update the contents of the original data stored in the second page 212 in the first block 210, the original data is maintained as it is, and the change associated with the original data is maintained. A record (log) is stored in the first sector 215 of the log storage spaces in the first block 210. In this case, it is preferable that the change record is sequentially stored in a storage area among log storage spaces. Also, the change record preferably includes address information of original data corresponding to the change record.

As described above, as the update request occurs for the content of the original data, the database manager 203 sequentially maintains the corresponding original data and stores only the change record in the storage area of the log storage space.

In this case, if an overflow occurs in which the log storage space for storing the change record is insufficient in the first block 210, the database manager 203 may enter the log storage space in the first block 210. A new version of the data is generated by reading the original data corresponding to the change record causing the overflow and all the change records stored in the corresponding log storage space and performing a merging operation.

Referring to FIG. 3, it is assumed in FIG. 3A that the A block is the first block 210 and the B block is the second block 220, and the data storage spaces and the log storage space of the A block. Assuming that it is full and block B is empty, a change record (log) should be stored in the log storage space when an update request for the original data DP1 stored in the data storage space of block A occurs. The log storage is already full, so overflow occurs.

Accordingly, the database manager 203 reads the original data DP1 that caused the overflow and all corresponding change records (stored in the log storage space) to perform a merge operation to generate a new version of the data DP1 '. It is.

In this case, the database manager 203 should allocate another block capable of storing the new version of data, and the method of allocating a block having an empty data storage space among the blocks being used and not in use There is a way to allocate an empty block.

First, it is determined whether there is a block having empty data storage space among the used blocks, and if there is a block, the page located at the top of the free data storage space of the block is allocated and the new version of data is stored. The first page of the data storage space of the empty block is allocated to store the new version of data.

After storing the new version of the data, the original data stored in the data storage space of the previous block is switched to the dirty state.

In this embodiment, a method of storing the new version of data in another empty block that is not in use is used. That is, as shown in (b) of FIG. 3, the first page of the data storage space of the empty block B, which is not in use, is allocated to store the generated new version of data DP1 ', and then the original of the block A. The data DP1 switches to the dirty state.

In addition, after storing the new version of the data in the data storage space of the other block, if the update request for one of the original data of the other original data stored in the data storage space of the previous block occurs also at this time Since the log storage space of the block is full, an overflow occurs.

Thus, a second new version of data is generated by performing a merge operation by reading the one original data corresponding to the change record causing the overflow and all corresponding change records from the log storage space of the previous block. To store in another block, and to convert the one original data of the previous block into a dirty state.

That is, as shown in (c) of FIG. 3, when an update request for another original data DPN-1 stored in the data storage space of block A occurs, the log storage space is already full, and thus the database management unit 203 ) Reads the original data DPN-1 and all corresponding change records (stored in the log storage space), merges them, and generates a second new version of data DPN-1 'to store the data storage space of block B. Are stored sequentially, and DPN-1 of block A is switched to a dirty state.

On the other hand, when an update request is generated for the new version of data stored in the other block, the database manager 203 keeps the new version of data in the same block as the data storage space in which the new version of data is stored. It is preferable to sequentially store the change records generated in the log storage space.

That is, as shown in (d) of FIG. 3, when an update request for the new version of data DP1 'stored in the data storage space of the block B occurs, the change record LS0 for this is written to the log storage space of the block B. To save.

In addition, when the number of original data converted to a dirty state from the previous block is a predetermined number or more, it is preferable to erase the corresponding block. That is, when the number of original data that has been changed from the block A to the dirty state is a predetermined number or more, the block A is erased to make an empty block.

4 is a flowchart illustrating a page merging method according to an embodiment of the present invention.

First, when an update request for original data stored in a data storage space in an arbitrary block occurs, it is determined whether an overflow occurs in the log storage space in the same block as the data storage space in which the corresponding original data is stored (S401). If an overflow occurs in the log storage space, the original data corresponding to the change record causing the overflow and the corresponding change records (stored in the log storage space) are read and merged. Generate a new version of the data (S403).

After that, it is determined whether there is an empty data storage space for storing the generated new version of data among other blocks in use (S405), and if there is a data storage space for storing, the empty data in the corresponding other used block. The new version of the data is stored in a storage space (S407), and if there is no other block in use having an empty data storage space for storing data, the new version of the new version is stored in the data storage space of another empty block that is not in use. Data is stored sequentially (S409).

Thereafter, the data that causes the overflow in the previous block in which the overflow occurs, that is, the original data corresponding to the data of the new version, is switched to the dirty state (S411).

For example, as shown in (a) of FIG. 3, assuming that block A is full of N data storage spaces and M log storage spaces, and block B is an empty block, If an update request for the original data DP1 stored in the data storage space occurs, an overflow occurs because the log storage space of the current block A is full.

Then, the original data DP1 and all related change records (stored in the log storage space of the block A) are read and merged to generate a new version of the data DP1 ', and the generated new version of the data DP1' is stored as an empty block B. In the first data page of the data storage space of the old block A, the old version of the data DP1 is switched to a dirty state that is no longer used.

Then, after determining whether an update request for the data of the new version has occurred (S413), if an update request for the data of the new version occurs, a log storage space of the same block as the data storage space in which the data of the new version is stored. The change record for the update is stored in step S415.

If the update request for the new version of the data does not occur, an update request for one original data among other original data stored in the data storage space in the previous block occurs, so that the log storage space in the previous block occurs. In step S401, a determination is made whether an overflow has occurred.

For example, suppose that an update operation on another original data DPN-1 occurs again in the previous block A, resulting in an overflow in the log storage space, as shown in FIG. Then, as in the above, the original data DPN-1 and all related change records (stored in the log storage space of the block A) are merged to generate a second new version of the data DPN-1 ', which is an empty block. In the second data page of the data storage space of In B, the old version of the data DPN-1 of the previous block A is switched to the dirty state.

Alternatively, assume that an update operation has occurred for a new version of data DP1 'stored in the data storage space of block B. Then, as shown in (d) of FIG. 3, the change record LS0 of the update occurring in the log storage space of the block B is stored.

In an embodiment of the present invention, a method of storing a new version of data in an empty block is taken as an example. However, when there is an empty data storage space in a valid block (live block) being used, the new version of the data is stored in the data storage space of the block. Data can be saved.

In addition, as the merge operation occurs as described above, the number of dirty data in the previous block increases, and when the number of dirty data exceeds a certain number, the operation for erasing the corresponding block is performed. You can do that.

In order to prove the validity of the present invention, the merge operation of block unit and merge unit of page unit are compared in flash memory DBMS. In this experiment, we use the log records related to the write operation extracted by creating a TPC-C benchmark trace set with a 1GB database, a 20MB buffer pool, and 100 virtual users, as input data.

The log record calculation statistics of the experiment trace are shown in Table 1 below.

calculate Number of occurrences ratio(%) insertion 673,077 47.11 delete 46,136 3.22 update 709,287 49.65 total 1,428,500 100.0

Table 1. Log Record Statistics for Experiment Traces

Table 2 shows a block consisting of 15 8KB data pages and 16 512B size log sectors, and the delete condition of the block (in case of page-by-page block consolidation) is set to '0' data pages in the block with live status. Experimental results when set.

Block Merger: Number of Block Mergers 47,896 Page Mergers: Number of Page Mergers 549,961

Table 2. Comparison of the number of merge operations between block unit merge and page unit merge

From Table 1, it can be seen that page-by-page mergers generated about 11 times more merges than block-based mergers. However, block-based merging is the fact that merging occurs for all pages that are in a live state in a block, so in fact, page-by-page merging is more efficient.

If the pages in the block are all live, the actual data unit merge reduces the number of data page merge operations by about 23.45% compared to the block unit merge. Thus, it can be seen that page-by-page merge that prevents unnecessary data page merges is more efficient with regard to write performance.

Table 3 compares the number of deletions of page merges and block merges.

Block unit merge: block delete count 47,896 Merging Pages: Block Deletion Count 35,305

Table 3. Comparison of the number of block deletions between block unit merge and page unit merge

The number of block deletions is about 26.28% less than that of page merges. From this, it can be seen that the page-by-page merge method has an effect of reducing the occurrence of block deletion and extending the life of the flash memory.

While the present invention has been described using the preferred embodiments, these embodiments are illustrative and not restrictive. Those skilled in the art to which the present invention pertains will appreciate that various changes and modifications can be made without departing from the spirit of the present invention and the scope of rights set forth in the appended claims.

FIG. 1 is a diagram illustrating a state before and after a block unit merging of conventional In-Page-Logging composed of N data pages and M log sectors.

2A is a block diagram illustrating a database management system according to an embodiment of the present invention.

2B is a diagram illustrating a database of a database management system according to an embodiment of the present invention.

3 is a diagram illustrating a merging method of data page units in a block according to a database management method according to an exemplary embodiment.

4 is a flowchart illustrating a page merging method according to an embodiment of the present invention.

Claims (14)

A database comprising a plurality of nonvolatile blocks; Allocate a data storage space and a log storage space for each of the blocks, and when there is a request for updating the original data stored in the data storage space, the log storage space in the same block as the data storage space in which the original data is stored. If the log storage space is insufficient, if the log storage space is insufficient, the change records stored in the log storage space corresponding to the original data are applied to the original data and merged to generate a new version of the data, and the generated new version of the data. Management unit to store the data in the data storage space of another block Flash memory based database management system comprising a. The method of claim 1, The database manager, when storing the generated new version of the data in the data storage space of the other block, the original data in the data storage space of the previous block to the dirty state (dirty), characterized in that the flash memory-based database Management system. 3. The method of claim 2, The other block is a flash memory-based database management system, characterized in that the block having a free data storage space of the block in use or not in use. 3. The method of claim 2, After the database manager stores the new version of data in the data storage space of the other block, if an update request for one original data among other original data stored in the data storage space of the previous block occurs, Apply and merge change records stored in the log storage space of the previous block corresponding to the original data to the one original data to generate a second new version of the data, and generate the second new version of the data; And storing the one original data in a dirty state in a data storage space of another block, and converting the one original data from a data storage space of a previous block into a dirty state. The method of claim 4, wherein And the database manager deletes the block when the number of original data which is converted to a dirty state in the data storage space of the previous block is a predetermined number or more. The method of claim 4, wherein When the update request for the data of the new version occurs, the database manager keeps the data of the new version as it is and logs a change record related to the data of the new version in the same block as the data storage space in which the data of the new version is stored. Flash memory based database management system, characterized in that for storing in the storage space. The method of claim 1, The database management system A database buffer which stores original data accessed by a computing device among the original data as in-memory data; A buffer manager configured to update the in-memory data and store it in the database buffer if there is an update request of the computing device for the in-memory data, and store an update record regarding the updated in-memory data in another portion of the database buffer. Flash memory-based database management system further comprises. A first step of determining whether an overflow occurs in a log storage space in the same block as the data storage space in which the original data is stored when an update request for the original data stored in the data storage space in any block occurs; A second step of generating a new version of the data by applying and merging all change records corresponding to the original data that is required to be updated to the original data when an overflow occurs in the log storage space; A third step of storing the generated new version of data in a data storage space of another block; Page merging method comprising a. The method of claim 8, wherein in the first step If the overflow does not occur, the original data for which the update is requested is kept as it is, and a change record corresponding to the original data is stored in a log storage space in the same block as the data storage space in which the original data is stored. How to merge pages. The method of claim 8, wherein the third step, Determining whether there is an empty data storage space for storing the new version of data among the blocks in use; If there is a block in use with empty data storage space, the new version of the data is stored in the data storage space of the block; if there is no empty data storage space, the new version of data is stored in the data storage space of the unused empty block. Steps to Page merging method comprising a. The method of claim 8, wherein after the third step, And converting the original data of the previous block corresponding to the new version of data stored in the other block into a dirty state which is no longer used. The method of claim 11, wherein after the third step, When an update request for one original data among other original data stored in the data storage space of the previous block occurs, change records stored in the log storage space of the previous block corresponding to the corresponding original data are stored in the one original data. Apply and merge to generate a second new version of the data, store the generated second new version of data in the data storage of the other block, and then store the one original data in the data storage of the previous block. Page merge method characterized in that the switch to the dirty (dirty) state. The method of claim 12, And deleting the corresponding block if the number of original data converted to a dirty state in the data storage space of the previous block is a predetermined number or more. The method of claim 11, wherein after the third step, When a request for updating the data of the new version stored in the data storage space of the other block occurs, the data of the new version is stored and the change record related to the data of the new version is maintained while the data of the new version is maintained. A page merging method, characterized by storing in log storage space in the same block as the space.

Images