KR101057468B1 - Buffer Consistency Management Method in Multiserver Database Management System Using Shared Disk Model - Google Patents

Abstract

The present invention relates to a method of managing buffer consistency in a multi-server database management system (DBMS) using a shared-disk model. The technical problem to be solved is a page that causes a buffer consistency problem. Multiple server database management systems and buffers using a shared disk model that can be selectively deleted from the server processor's buffers, reducing the unnecessary work for efficient use of the buffers, and maintaining consistency among the buffers of multiple server processes. It is to provide a consistency management method.

To this end, in the multi-server DBMS using the shared disk model according to the present invention, the buffer consistency management method includes information on the modified buffer page when the server process writes out the buffer page to disk during a transaction. A transaction execution step of storing page information (page-info) in a page-info list in the server process, and page information in a coherency volume if the server process acquires a volume lock during transaction execution. page-info) is read and stored in the page-info list of the server process at the end of the transaction after the volume lock acquisition and transaction execution or volume lock acquisition steps that match the contents of the buffer with the database. Transactions that store modified page-info in a coherency volume Characterized in that it comprises a termination step.

Multi-server database management system (DBMS), shared-disk, buffers, consistency, pages

Description

Buffer consistency management method for multi-server database management system using shared-disk model}

The present invention relates to a buffer consistency management method in a multi-server database management system (DBMS) using a shared-disk model.

Most companies today offer a variety of services online by allowing their databases to be accessed over the Internet. For example, a flight schedule stored in the airline's database can be viewed through the Internet. When a ticket is booked, the reservation information is stored in the airline's database. In order to provide such a service, a web server and a database management system (DBMS) are used in conjunction with each other. In this case, a shared-disk model is used to quickly process a large number of incoming requests. There is a need to use a multi-server DBMS.

As shown in Fig. 1, a multi-server DBMS is a DBMS having a plurality of server processes. At this time, the server processes operate on one machine or a plurality of machines, and shared-memory, shared-nothing, and shared disks (depending on whether or not the plurality of server processes share memory and disks with each other). Shared-disk) model.

In the shared-memory model, as shown in FIG. 2, all server processes share a disk and a memory. Thus, while fewer disk accesses and lower communication costs, if a particular server process performs a wrong operation and stores it in shared memory, all server processes are affected, which makes it less reliable and difficult to scale to multiple machines. There are disadvantages.

In the shared-nothing model, as shown in FIG. 3, each server process has a separate memory and disk. Therefore, there is no shared resource among server processes, so it is easy to expand to multiple machines, and even if a certain server process does a wrong operation and saves it in memory, it does not affect the memory of other server processes. However, since there is no shared resource, the cost of database update is high when the communication costs between server processes and the modifications made by a particular server process are reflected on all disks.

In the shared-disk model, as shown in Figure 4, each server process has independent memory and only shares disks with each other. At this time, a method such as a network file system (NFS) or a storage area network (SAN) is used to share disks among a plurality of machines.

The shared-disk model, like the shared-nothing model, is highly reliable because server processes do not share memory, and the overhead of sharing disks on multiple machines incurs memory sharing. Relatively small compared to the overhead, it is easy to scale to multiple machines compared to the shared-memory model.

Multi-server DBMS used for web service must satisfy the following conditions. First, Web services must be available to users 24 hours a day, so they must be reliable. Second, in order to process many user requests quickly, it is advantageous for server processes to run on multiple machines, so it must be easy to scale to multiple machines. Third, the recent web services support the updating of contents by the user, so the updating of the database should be easy.

Therefore, of the three models of multi-server DBMSs, the shared-disk model is less reliable and scalable than the shared-memory model, or the shared-nothing model, which requires expensive database updates. Suitable for use in services.

However, the shared-disk model has a disadvantage in that, because server processes do not share buffers with each other, if multiple server processes access the same page, each server process needs to load a new page from the public disk. .

Also, even if a specific server process modifies the database, the modified contents are not reflected in the buffers of other server processes. When other server processes query, the query is executed on the contents before the modification, reflecting the exact state of the database. There is a buffer consistency problem.

As shown in FIG. 5, the buffer consistency problem indicates that there is a problem in the consistency of database information. When server processes P ₁ and P ₂ read the same page p, each server process reads a buffer page corresponding to page p. It is held in each memory. Subsequently, if server process P ₁ modifies page p, the contents of buffer page p of server process P ₂ differ from those of page p in the database. Thus, if process P ₂ reads or modifies the buffer page p, it will cause a database consistency problem.

Assuming that all pages in the database have been modified to solve the buffer consistency problem, there are ways to delete all the pages that the server process reads or writes from the buffer and read the modified pages into the new buffer. Even if pages that are not modified and do not cause buffer consistency problems are deleted from the buffer and read back from the database, there is a problem that greatly reduces the performance of the multi-server DBMS.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems. When a server process loads a page of a database into a buffer or stores a page in a database, a page modified by another server process, that is, a page that causes a buffer consistency problem, is generated. Multiple server DBMSs using a shared-disk model that selectively removes only from the server processor's buffers, thereby reducing unnecessary work and streamlining the use of buffers and maintaining consistency among the buffers of multiple server processes. And the purpose of this is to provide a buffer consistency management method using the same.

In addition, it uses a shared-disk model that manages buffer consistency by using minimal information and a small amount of storage for pages modified by a particular server processor, thereby increasing the performance of multi-server DBMSs. The purpose is to provide a multi-server DBMS and a buffer consistency management method using the same.

In order to solve the above problems, according to the present invention, in the buffer consistency management method in a multi-server database management system using a shared-disk model, the server process buffer page during the transaction Performing a transaction for storing page information (page-info) having information on the modified buffer page in a page-info list in the server process when writing to a disk; If the server process acquires a volume lock during the transaction, the volume lock reads the page-info page in the coherency volume where the page information is stored and matches the contents of the buffer with the database. At the end of the transaction after performing the acquisition step and the transaction execution step or volume lock acquisition step And a transaction termination step of storing the modified page information (page-info) stored in the page-info list of the existing server process in the coherency volume. A buffer consistency management method is provided in a multi-server database management system that uses.

Here, the server process maintains a last access timestamp indicating information on when the database was last accessed, and a page-info page in the coherency volume. Each page information (page-info) stored in the web page maintains a page identifier (page ID) for identifying a modified page and a timestamp indicating when the page is modified, and the coherency volume is a timestamp counter used to assign a timestamp value to a timestamp of the page-info or the last access timestamp. It is characterized by maintaining.

The page information page stored in the coherency volume may store page information in a circular list having a predetermined size N in the system.

The transaction ending step may include reading a timestamp counter from the coherency volume, and last access time stamp of the server process after the reading step. All page information (page-info) in the page-info list of the server process after an update step of updating a timestamp to a result of performing the read step and the update step. And a delete step of deleting them.

In addition, when the transaction is terminated after the transaction is performed, if a commit process is performed to actually reflect the database update performed in the transaction to the database, between the update step and the delete step. The latest access timestamp updated in the update step is added to a timestamp of all page information in the page information list of the server process. Write page and write page-infos in the page-info list of the server process to a page-info page in a coherency volume And further increasing the timestamp counter by one.

In addition, the volume lock acquiring step may include: the page information stored in a page-info page in a coherency volume and a last access timestamp of the server process. The timestamp of the page-info is compared to the last access timestamp, and if the timestamp of the page-info is later, the corresponding page information (page-info). And deleting the page pointed to by the buffer.

In addition, in order to solve the case where some of timestamps of page-infos stored in the circular list of the coherency volume are overflowed in the volume lock obtaining step, the overflow occurs. The maximum value of the timestamp is used as the maximum value T _{timestamp_max} smaller than the maximum expression value T _max of the integer type that is the representation of the _timestamp , and the time of the page-infos. The timestamp that overflows among the stamps is added to a predetermined value T _add preset in the system.

의 값을 가지고, 상기 T_add는,

의 값을 가지는 것을 특징으로 한다. In addition, the T _{timestamp_max} is,

_Has a value of T _add ,

It characterized by having a value of.

Further, in the volume lock obtaining step, some of the page information (page-info) stored in the circular list of the coherency volume is overwritten by the newly stored page information (page-info). If the last access timestamp is less than the timestamp of the first page-info of the circular list of the coherency volume, the server process is responsible for all of the buffers in the buffer. By deleting pages from the buffer, all pages in the buffer are matched with the database.

In addition, when the timestamp of all page information (page-info) stored in the circular list of the coherency volume overflows in the volume lock obtaining step, the most recent access time of the server process If the last access timestamp is greater than the timestamp of the page-info of the circular list, the server process deletes all pages in the buffer from the buffer, thereby removing the pages in the buffer. It is characterized by matching all pages to database.

delete

As described above, according to the multi-server DBMS using the shared-disk model according to the present invention and the buffer consistency management method using the same, only a minimum amount of information on pages modified by a specific server processor is used, By using the storage space and data structure to deal with the buffer consistency problem, the processing speed is increased to improve the performance of the multi-server DBMS.

In addition, when a server process attempts to read a buffer page that loads a page into the buffer or reads a page into the database, the server processes the buffer consistency problem by selectively deleting only pages that have already been modified by another server process from the server processor's buffer. By eliminating buffer pages that do not cause consistency problems, the unnecessary work of reloading from the database is reduced, which makes the use of the buffer more efficient and maintains consistency among the buffers of multiple server processes.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. First, it should be noted that the same components or parts in the drawings represent the same reference numerals as much as possible. In describing the present invention, detailed descriptions of related well-known functions or configurations are omitted in order not to obscure the gist of the present invention.

FIG. 6 is a diagram illustrating a multi-server DBMS using a shared-disk model configured to enable consistent management of buffers according to the present invention, and includes a timestamp counter included in a coherency volume. A counter is a data structure used to assign a timestamp value to a timestamp or last access timestamp of page information.

Multi-server DBMS using a shared-disk model, as shown in Figure 6, consists of a plurality of server processes, databases existing on the existing system, and here is a characteristic volume of the consistent configuration of the present invention (coherency volume) is additionally included.

Hereinafter, a buffer consistency management method according to the present invention will be described in detail.

7 is a block diagram illustrating a buffer consistency management method according to the present invention, and FIG. 8 is a block diagram showing detailed steps performed in a transaction termination step of a server process in the buffer consistency management method according to the present invention.

9 is an exemplary view showing a buffer matching time of a multi-server DBMS according to the present invention, Figure 10 is a coherency volume of page information (page-info) at the end of a transaction of the multi-server DBMS according to the present invention A diagram showing an algorithm to store in.

For buffer consistency management in a multi-server DBMS, the server process newly buffers up-to-date pages from the database by deleting those pages in the buffer before accessing the modified pages from other server processes in the buffer. Should be read as. Here, in the following description, the server process deletes pages modified in another server process from the buffer as 'matching the buffer contents with the database'.

The timing of matching the buffer contents with the database is closely related to the database's concurrency control model, because buffer consistency issues arise from the execution of multiple server processes. The buffer consistency problem occurs at a specific point in time in a plurality of server processes executed according to the concurrency control model, so it is necessary to find this point and match the contents of the buffer with the database.

In the present invention, lock-based protocols are used as the concurrency control method. In the lock-based method, the unit of lock is a volume. A volume is a logical storage unit consisting of one or more files. A database consists of one or more volumes.

There are two types of volume locks: read volume locks and write volume locks. A read volume lock is a lock that is acquired when a read operation is performed on a volume, which is compatible with the read volume lock and compatible with the write volume lock. I never do that. Thus, transactions with read volume locks on the same volume can be executed concurrently, but not with transactions with write volume locks.

The write volume lock is a lock obtained when a write operation is performed on the volume. The write volume lock is not compatible with both the read volume lock and the write volume lock. Therefore, a transaction with a write volume lock cannot run concurrently with a transaction with a read volume lock or a write volume lock on the same volume.

In this concurrency control model, after a server process modifies a volume's contents and terminates a transaction, buffer consistency issues arise when another server process accesses the same volume. Therefore, the time to match the contents of the buffer with the database must be after the server process has finished the write transaction for a particular volume, but before other server processes have access to that volume. At this time, since the end of the write transaction of another server process is difficult to know, the contents of the buffer should be matched with the database when the volume lock immediately before the volume is acquired.

A buffer matching time point of the multi-server DBMS according to the present invention may be given as shown in FIG. 9.

After the server process P ₁ a, modify the contents of the volume V, and end the transaction, when the server process P ₂ can access the volume V buffer page corresponding to the page, the server processes P ₁ changes to the buffer of the server process P ₂ has If present, a buffer consistency problem will occur. At this time, server process P ₂ obtains the volume lock for volume V before accessing volume V. Therefore, at the time of obtaining the volume lock, the contents of the buffer must be matched with the database. If server process P ₂ matches the contents of the buffer with the database at the beginning of a transaction rather than at the point of acquiring a volume lock, modifications made by server process P ₁ during interval i are not reflected in the buffer of server process P ₂ . This results in buffer consistency problems.

The buffer consistency management method according to the present invention comprises a process as shown in FIG.

Each server process should perform a task as shown in FIG. A process of storing page information (page-info), which is information about, in a page information list.

At this time, the timestamp of the page information (page-info) is left blank to store the last access timestamp to be allocated to the server process at the end of the transaction. The reason for doing this at the end of a transaction rather than during a transaction is that if you modify a page multiple times in a transaction, it is inefficient to write the page identifier to the page-info list every time the page is modified. In order to reduce the overhead of determining modified pages.

In addition, the transaction termination step of FIG. 7 is a process of storing the page information (page-info) in the page-info list of the server process in the coherency volume at the end of the transaction, FIG. An algorithm representing the work to be done at the city.

In more detail, the transaction termination step may include up to five steps, as shown in FIG. 8.

The read step reads the value of the timestamp counter of the coherency volume, and the update step reads the value of the last access timestamp of the server process. The step of updating the timestamp counter value read in the read step.

On the other hand, the write step is a timestamp value of all the page-info stored in the page-info list of the server process as the last access timestamp value. The step of updating, and the increase step stores the page information (page-info) of the page-info list to the page information page (coherency volume) of the page-info page, and time A step of increasing the value of the stamp counter by 1 is performed.

Finally, the delete step is to delete all page information (page-info) of the page information list (page-info list) of the server process.

At this time, the write step and the increase step are performed when the transaction is terminated after the transaction is executed and when the database update performed by the transaction is actually reflected in the database, that is, when the commit process is performed. Is performed only in all cases and all other steps are performed in any case.

The volume lock obtaining step of FIG. 7 is a step of accessing a page-info page in a coherency volume and matching the buffer contents with a database. To do this, we search the circular list of coherency volumes from behind, find page-info that has a timestamp greater than the last access timestamp of the server process, and find the corresponding content. Delete from server process.

At this time, the reason for scanning the circular list of the coherency volume from behind is to reduce the scanning range. That is, a page having a timestamp value larger than a last access timestamp value because page-info pages of a page information page are stored in a timestamp order. You only need to scan and delete the page-info.

Since multiple server processes access coherency volumes, concurrency control on coherency volumes is required. Coherency volumes are accessed at the end of a transaction and at the time of obtaining a volume lock. At the end of a transaction, a coherency volume is updated to obtain a write volume lock on the coherency volume and to obtain a volume lock. At this time, the read volume lock for the coherency volume is obtained because only the coherency volume needs to be read.

Meanwhile, the present invention stores only a limited number of page-infos in a coherency volume and uses an integer timestamp of limited size for efficient buffer consistency management. Problems caused by wrap around, problems caused by overflow of timestamp values of page-info stored in the prototype list, and information of page-info stored in the prototype list. An overflow of the timestamp value may cause a problem of not finding a wrap around of the prototype list.

Examples of the problems that may occur as described above and how to solve these problems will be described.

First, we define the notations used for explanation.

t _{last_access} is the last access timestamp value of the server process that is performing the task of matching the contents of the buffer to the database, and t _head is the page information stored in the circular list of coherency volumes. Is a timestamp value of a head, that is, a first page information (page-info), and t _tail is a _tail (page-info) stored in a circular list of coherency volumes. tail, that is, the timestamp value of the last page information (page-info).

In addition, T _max is a maximum value of a timestamp, N is a maximum number of page-infos stored in a circular list of coherency volumes, and is set in the system.

11 is an exemplary diagram illustrating a problem due to a wrap around of a prototype list according to the present invention, and FIG. 12 is a timestamp value of page-info stored in the prototype list according to the present invention. FIG. 13 is a diagram illustrating a problem due to an overflow of a circular list, and FIG. 13 is a circular list due to an overflow of a timestamp value of page-info stored in a circular list according to the present invention. This is an exemplary view showing a problem of not finding a wrap around.

In addition, FIG. 14 is a problem of not finding a wrap around of a prototype list due to an overflow of a time stamp value of page information stored in the prototype list according to the present invention. FIG. 15 is a diagram illustrating an algorithm for matching a content of a buffer with a database when a server process acquires a volume lock, according to an embodiment of the present invention.

delete

In a circular list of coherency volumes, page-info, which some server processes have not yet matched the database, may be overwritten by newly stored page-info. In this case, some server processes fail to match the buffer page corresponding to the overwritten page-info, even though the database must be matched, which is caused by the wrap around of the prototype list. It is defined as. Only a maximum of N page information (page-info) is stored in the circular list of the coherency volume. This occurs because the page information is stored in the circular list of limited size.

Hereinafter, the problem caused by the wrap around of the circular list of the page-info page according to the present invention will be described with reference to the example shown in FIG.

After server process P ₁ performs transaction T ₁₁ , when server process P ₂ performs transaction T ₂₁ and modifies two pages, performs transaction T ₂₂ and modifies N-1 pages, transaction T ₂₁ <P ₂ , 2>, the page-info of the page modified by, is overwritten by <p _{N + 1} , 3>. At this time, the t _{last_access} server process P ₁ at the time when m ₂ is 1, so match the buffer page server process P ₁ a that corresponds to the page information _{(page-info) <p 2} , 2> while performing the transaction T ₁₂ and the database It should not match, though it must.

This can happen if the t _{last_access} of the server process is smaller than the t _head of the circular list. Since at least N pages have been modified since t _{last_access} , many modifications _assume that most of the buffer pages have been modified. Resolve all pages in the buffer by matching them with the database.

Meanwhile, when a timestamp value of page information stored in the circular list overflows, page information of the circular list of the coherency volume is timestamped. ) Will not be sorted in order. In this case, it is contrary to the assumption that page infos are arranged in timestamp order in the circular list of coherency volumes, so matching the contents of the buffer with the database when obtaining volume locks is normal. A problem occurs that is not performed. This problem is defined as a problem due to an overflow of a timestamp value of page information stored in a circular list. This problem occurs because there is a maximum size (T _max ) of values that can be represented by an integer type used as a timestamp.

Hereinafter, the problem caused by the overflow of the timestamp value of the page information (page-info) stored in the circular list according to the present invention will be described with reference to the example shown in FIG.

Page-info <p _N-1 , 0>, <p _N-2 , 1> is recorded later in time than page-info <p _N-2 , T _max > The timestamp value is 0, 1, which is less than T _max .

At this time, a method of determining whether or not the timestamp value overflows is as follows. The timestamp value of each page-info is compared with t _head while the circular list of coherency volumes is scanned from the tail. If the timestamp value of the page information (page-info) is smaller than t _head , it is determined that the corresponding timestamp has overflowed.

를 사용한다. 그리고,오버플로우(oveflow)된 타임스탬프(timestamp)들에는 일정한 값

를 더하여 t_{last_access}와 비교한다. 이때, T_{timestamp_max}와 T_add는 다음 세 가지 조건을 만족시키는 값이다. 첫째, 타임스탬프(timestamp)의 최대값에 T_add를 더한 값이 오버플로우(overflow) 되지 않아야 한다. 이를 위해서 T_{timestamp_max} + T_add가 T_max보다 작아야 한다. 둘째, 타임스탬프(timestamp)의 최소값(=0)에 T_add를 더한 값이 어떠한 타임스탬프(timestamp) 값보다 커야 한다. 셋째, 타임스탬프(timestamp) 값이 자주 오버플로우(overflow) 되는 것을 방지하기 위해서 T_{timestamp_max}를 최대화시켜야 한다. To solve this problem, do not use T _max as the maximum value of the timestamp, but smaller than that.

Use And a constant value for overflowed timestamps.

Add to compare t _{last_access} . At this time, T _{timestamp_max} and T _add satisfy the following three conditions. First, the maximum value of the timestamp plus T _add should not overflow. To do this, T _{timestamp_max} + T _add must be smaller than T _max . Second, the minimum value of timestamp (= 0) plus T _add must be greater than any timestamp value. Third, T _{timestamp_max} should be maximized to prevent the timestamp value from overflowing frequently.

FIG. 13 and FIG. 13 do not find a wrap around of the prototype list due to an overflow of a timestamp value of page information stored in the prototype list according to the present invention. Each is explained by 14.

In the case of FIG. 13, after server process P ₁ performs transaction T ₁₁ , server process P ₂ performs transactions T ₂₁ and T ₂₂ and outputs a coherency volume of more than N page-info. In the prototype list. To be the server process P ₁ at the time point m to match the buffer page for every page while performing T ₁₂ modified on the server process P ₂ with the database, the server process P ₁ a t _{last_access} any page information in the circular list (page -info) does not match the contents of the buffer because it is larger than the timestamp value.

This problem occurs _when t _{last_access} of the server process is larger than the timestamp value of all the page- _infos of the circular list of the coherency volume. If this problem occurs, match all the buffer pages to the database as in the case of the wrap around of the prototype list.

However, this method does not handle the case where the overflowed timestamp value becomes as large as t _{last_access that} does not overflow. 14 is an exemplary view illustrating such a case.

개의 트랜잭션을 수행하였다면, 이때 저장되는 페이지 정보(page-info)들의 타임스탬프(timestamp) 값은 오버플로우(overflow)된 후 다시

이 된다. 시점 m에서 서버 프로세스 P₁이 T₁₂를 수행하면서 서버 프로세스 P₂에서 수정된 모든 페이지들에 대한 버퍼 페이지를 데이터베이스와 일치시켜야 하는데, 서버 프로세스 P₁의 t_{last_access}가 t_tail과 같은 값을 갖기 때문에 버퍼 내용을 데이터베이스와 일치시키지 못한다. In Figure 14, after the server process P ₁ performs the transaction server process P ₂ is T ₁₁

If the number of transactions is executed, the timestamp value of the page-infos stored at this time overflows and then again.

Becomes At time m, server process P ₁ performs T ₁₂ and must match the buffer page for all pages modified in server process P ₂ with the database, because t _{last_access} of server process P _{1 has} the same value as t _tail. The contents of the buffer do not match the database.

However, such a case as in FIG. 14 is unlikely to occur in an actual application. If you use an integer of 4 bytes as a timestamp, in an environment where 5.4 transactions per second are executed, this problem requires that a specific server process not perform a transaction for about 12.6 years before performing another transaction. For reference, if 5.4 transactions per second are executed, about 170 million transactions are executed per year, which is equivalent to the 170 million transactions of the Korea Stock Exchange. Therefore, this problem is considered not to occur in a real environment.

The problems described so far can be categorized as follows using the size relationship of t _{last_access} , t _head , and t _tail .

In other words, (t _head <= t _{last_access} <= t _tail ) is normal.

If (t _head > t _tail ), the problem occurs due to overflow of timestamp value of page-info stored in the circular list, and (t _{last_access} <t _head < = t _tail ) is a problem caused by wrap around of the circular list. In addition, when (t _head <= t _tail <t _{last_access} ), the wrap around of the prototype list is wrapped around due to the overflow of the timestamp value of the page-info stored in the prototype list. This is the case when you don't find).

The above-described problems can be solved by dividing the cases as described above and performing the operation as described above with respect to each case.

15 is a diagram illustrating an algorithm for matching a content of a buffer with a database when a server process acquires a volume lock according to the present invention.

Using a multi-server DBMS using a shared-disk model, the performance of the buffer consistency management method proposed by the present invention and the conventional method are compared.

Here, the conventional method is to invalidate the pages in the volume in the buffer each time a volume lock is obtained. As DBMS, Odysseus: a High-Performance ORDBMS Tightly-Coupled with IR Features, "In Proc. 21st (Whang, K. Lee, M., Lee, J., Kim, M., and Han, W.). Int'l Conf. On Data Engineering (ICDE), pp. 1104-1105, Demonstration Paper, Apr. 2005.7. The data uses the databases and transactions defined in the Transaction Processing Performance Council (TPC), the TPC Benchmark C Standard Specification, Revision 5.9, June 2007.

In this experiment, when the database size is 790MB and the buffer size is 80MB, the number of processes is increased from 2 to 32 and the performance change is measured accordingly. FIG. 16 shows the results of the experiments of the present invention and the conventional method, and the method according to the present invention is improved by 1.2 to 2.9 times as compared to the conventional method. In this case, the number of I / O operations increases as the number of server processes increases, since the number of pages modified by other server processes increases, so the number of pages invalidated in the buffer increases for buffer consistency management. to be.

As described above with reference to the drawings illustrating a multi-server DBMS using a shared-disk model according to the present invention and a buffer consistency management method using the same, but by the embodiments and drawings disclosed herein The invention is not limited, and of course, various modifications may be made by those skilled in the art within the technical scope of the present invention.

1 is a diagram illustrating a multi-server DBMS.

2 is a diagram illustrating a multi-server DBMS using a shared-memory model.

3 is a diagram illustrating a multi-server DBMS using a share-nothing model.

4 is a diagram illustrating a multi-server DBMS using a share-disk model.

5 is an exemplary diagram illustrating a buffer consistency problem in a multi-server DBMS.

6 is a diagram illustrating a multi-server DBMS using a shared-disk model according to the present invention.

7 is a block diagram illustrating a buffer consistency management method according to the present invention.

8 is a block diagram illustrating a second process of the buffer consistency management method according to the present invention.

9 is an exemplary view showing a buffer matching time of a multi-server DBMS according to the present invention.

10 is a diagram illustrating an algorithm for storing page information (page-info) in a coherency volume at the end of a transaction of a multi-server DBMS according to the present invention.

11 is an exemplary diagram illustrating a problem due to wrap around of a circular list according to the present invention.

FIG. 12 is an exemplary diagram illustrating a problem due to overflow of a timestamp value of page information (page-info) stored in a circular list according to the present invention.

FIG. 13 illustrates an example in which a wrap around of a prototype list is not found due to an overflow of a timestamp value of page information stored in the prototype list according to the present invention. 1.

FIG. 14 illustrates an example in which a wrap around of a prototype list is not found due to an overflow of a timestamp value of page-info stored in the prototype list according to the present invention. 2.

15 is a diagram illustrating an algorithm for matching a content of a buffer with a database when a server process acquires a volume lock according to the present invention.

Claims (10)

In the method of managing buffer consistency in a multi-server database management system using a shared-disk model, When the server process writes out a buffer page to disk during the execution of a transaction, page-info having information about the modified buffer page is returned to a page-info list in the server process. Transaction execution step to store in, If the server process acquires a volume lock during the transaction, the volume lock reads the page-info page in the coherency volume where the page information is stored and matches the contents of the buffer with the database. Acquisition stage and After performing the transaction performing step or the volume lock obtaining step, the modified page-infos stored in the page-info list of the server process at the end of the transaction are converted into the coherency volume. A buffer consistency management method in a multi-server database management system using a shared disk model, characterized in that it comprises a transaction termination step of storing in a database. The method of claim 1, The server process maintains a last access timestamp indicating information about when the database was last accessed. Each page information (page-info) stored in the page information page (page-info page) in the coherency volume, the page ID (page ID) for identifying the modified page and when the page is modified Maintain a timestamp that indicates The coherency volume is a timestamp counter used to assign a timestamp value to a timestamp of the page-info or the last access timestamp. A method of managing buffer consistency in a multi-server database management system using a shared disk model, characterized by maintaining a timestamp counter. The method of claim 1, The page information page (page-info page) stored in the coherency volume (shared page), characterized in that for storing the page information (page-info) in a circular list having a size (N) preset in the system A method for managing buffer consistency in a multiserver database management system using a model. The method of claim 1, The transaction termination step, A read step of reading a timestamp counter in the coherency volume; An update step of updating a last access timestamp of the server process to a result of performing the read step after the read step; And a deleting step of deleting all the page information in the page information list of the server process after the updating step. A method of managing buffer consistency in a multiserver database management system. The method of claim 4, wherein When the transaction is terminated after the transaction is performed, if a commit process is performed to actually reflect the database update performed in the transaction to the database, between the update step and the delete step, The last access timestamp updated in the update step is recorded in a timestamp of all page-info in the page-info list of the server process. Write phase and Store page information in the page information list of the server process in a page information page in a coherency volume and store a timestamp counter. A method of managing buffer consistency in a multi-server database management system using a shared disk model, characterized by further increasing the increment by one. The method according to any one of claims 1 to 5, The volume lock obtaining step, Comparing a last access timestamp of the server process with a timestamp of the page-infos stored in a page-info page in the coherency volume If the timestamp of the page-info is later than the last access timestamp, the page indicated by the page-info is deleted from the buffer. Buffer consistency management method in a multi-server database management system using a shared disk model, characterized in that it further comprises a process. The method of claim 6, In order to solve a case where some of timestamps of page-infos stored in the circular list of the coherency volume overflow. As the maximum value of the timestamp, a value T _{timestamp_max} smaller than the maximum expression value T _max of an integer type that is an expression of a _timestamp is used as the maximum value, and the page-infos Buffer consistency in a multi-server database management system using a shared disk model, which adds a predetermined value (T _add ) to a system that overflows among timestamps. How to manage. The method of claim 7, wherein The T _{timestamp_max} is,

Has the value of, T _add is,

A buffer consistency management method in a multi-server database management system using a shared disk model, characterized in that the value of. The method of claim 6, The last access timestamp of the server process due to the problem that some page information (page-info) stored in the circular list of the coherency volume is overwritten by the newly stored page information (page-info). ) Is smaller than the timestamp of the first page information (page-info) of the circular list of the coherency volume, And the server process deletes all the pages in the buffer from the buffer so that all the pages in the buffer match the database. The method of managing buffer consistency in a multi-server database management system using a shared disk model. . The method of claim 6, When the timestamp of all page information (page-info) stored in the circular list of the coherency volume has overflowed, the last access timestamp of the server process is determined. Is greater than the timestamp of the last page-info of the prototype list, And the server process deletes all the pages in the buffer from the buffer, thereby matching all the pages in the buffer with the database. .

Images