JPH04360265A - Parallel transaction processing system for processing data distribution - Google Patents

Abstract

PURPOSE:To efficiently drive a system even when data are distributed to plural processors in one transaction by executing the processing of a reference unit on a processor storing data relating to the reference processing unit. CONSTITUTION:Transaction processing is decomposed into plural reference processing units, which are dynamically executed on plural processors in parallel with each other in accordance with the dependent relation of respective reference processing units. Each reference processing is executed by the processor including data relating to the reference processing. For instance a transaction 001 reads out the contents of records A, B and C and writes the sum of the contents in the record C. Respective reference processing units are allocated to processors #1 to #3 and executed and processing such as A + B + C relating to all records is executed by the processor #3 in which final data are to be written. Thus efficient execution can be attained.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a parallel transaction processing system that handles data distribution by decomposing a transaction into basic processing units and executing the basic processing units in processors in which data is stored. Regarding the arrangement of data processed by this transaction.

0002

[Prior Art] First, as a method for efficiently controlling the concurrent execution of transaction processing in a computer, (1)
"Transaction Processing System" (March 20, 1990)
(2) ``Deadlock detection method'' that detects deadlocks in parallel and performs partial rollback.
(filed on November 30, 1990, patent application No. 2-329658)
No.) was filed.

[0003] In application (1), in order to perform concurrency control flexibly and efficiently, transactions are broken down into basic processing units such as record locking, reading, content modification, and writing, and these basic processing units are parallelized (parallel). proposed a method to implement it. For efficient parallel processing, the data and control dependencies between basic processing units are graphed, and basic processing is suspended/interrupted according to the graph.
It is characterized by being able to resume (restart) and proceed with multiple transaction processing simultaneously. Since the optimal execution order is dynamically determined between basic processing units, processes that do not need to wait for locks can be preceded without having to wait until the lock that is the bottleneck for processing is released, and lock-dependent There is no need to include scheduling in consideration of relationships in the program. Furthermore, since graphs can be created for each application, versatility is also increased.

[0004] Furthermore, in application (2), as a method for detecting deadlock, a lock management table and a list of resource request source identifiers (transaction ID: TID in the case of transaction processing) are used to detect deadlocks during lock processing. How to investigate lock dependencies through communication, and by rolling back only the basic processing unit involved in the deadlock, instead of rolling back on a per-transaction basis when a deadlock is detected. We proposed a method to limit the In other words, once read data, data related to deadlock needs to be read again, but parts unrelated to deadlock can be used as is. For this reason,
Rollback processing can be made easier. Related parts and unrelated parts can be derived from the graph between basic processes.

[0005]

[Problem to be Solved by the Invention] The above invention has provided a means for efficiently executing transactions in parallel, but when actually executing transactions in parallel, the arrangement of data to be processed in a transaction is not taken into account. The problem remained that there was no. In other words, it was necessary to find out how to divide the data to be processed in a transaction by each parallel processing processor so that the transaction can be efficiently executed in parallel. When executing transactions in parallel without breaking them down into basic processing units, it is possible to process data stored across multiple processors within one transaction, since only transaction units can be allocated to processors. The problem was that it was not very efficient. That is, unless a transaction is broken down into basic processing units, the parallelism of processing for this transaction cannot be increased, and therefore this transaction cannot be processed at high speed.

SUMMARY OF THE INVENTION An object of the present invention is to provide a parallel transaction processing system that handles data distribution and can operate the system efficiently when data is distributed across multiple processors within one transaction. .

[0007]

[Means and effects for solving the problem] When transaction processing is decomposed into basic processing units and executed dynamically in parallel on multiple processors according to the dependencies of the basic processing units, the basic processing is Ensure that the process is executed on the processor where the related data resides.

For example, consider an example as shown in FIG. It is assumed that transaction 001 reads records A, B, and C, and writes the sum to record C. At this time, record A is sent to processor #1, and record B is sent to processor #2.
Assume that record C was on processor #3,
Each basic processing unit is assigned to and executed by a processor as shown in FIG. It is desirable that processing related to any record, such as A+B+C, be performed on processor #3, which ultimately writes the data.

[0009] In order to perform such allocation, a data arrangement dictionary regarding data arrangement is prepared. The data arrangement dictionary is a table that has entries for relevant data, and includes the identifier (or processor number) of the processor in which the data is stored. It is desirable that each entry can be accessed by indexing techniques such as hashing. That is, the data arrangement dictionary is a table. This table stores the processors with data and the entries for this data.

Furthermore, each processor has its own execution queue. Each basic processing unit obtains the processor number in which the data to be processed is stored from the data arrangement dictionary, and stores the data in the execution queue of the processor indicated by that number.

It is desirable that the arrangement of the dictionaries be changed depending on how the parallel processors are configured. Here, dictionary placement refers to where this dictionary should be placed in the system. Of course, the arrangement differs depending on the system configuration. For example, in the case of a shared memory type parallel processor, one data placement dictionary is placed on the shared memory, and in the case of a parallel processor without shared memory, each processor has a copy of the data placement dictionary. .

0012

【Example】

(Embodiment 1) The basic configuration of the present invention is shown in FIG.

[0013] Transaction 1 is input via the user interface and is processed by basic processing unit dependency graph 3.
Activate the input of the corresponding node in a graph prepared in advance (a method of creating a graph at the time a transaction is entered is also possible). The suspend/resume mechanism section 4 examines all inputs of nodes in the graph and checks whether the nodes can fire. here,
Firing is possible means that processing is possible because all nodes in the graph satisfy a certain condition. If firing is possible, the graph execution unit 5 is notified of the graph node. The graph execution unit 5 learns from the data arrangement dictionary 6 which processor stores the data related to the node of the graph instructed by the suspend/resume mechanism unit 4, and determines the basic processing unit of the processor. Put it in the execution queue (Figure 3). Each processor looks at its own queue and executes from the beginning, and notifies the graph execution unit when processing is complete. The graph execution unit activates the corresponding input of the node that inputs the execution result according to the basic processing unit dependency relationship graph 3. When executing a database access node in the graph, an access request is issued to the database search/update section 7 to access the database 8. Processing results are returned through the user interface section 2. If a lock is required for exclusive control, a lock request is issued to the lock mechanism section 9. The lock mechanism unit 9 uses a lock table and a TID list 10 according to the lock request.
operate.

FIG. 3 shows a configuration of the present invention on a parallel processor having a shared memory. The basic processing unit dependency relationship graph 3 and the data arrangement dictionary 6 are placed on the shared memory 21. Queue #1 corresponding to each processor
, #2, and #3 are also actually placed on the shared memory 21. The graph execution unit 5, suspend/resume mechanism 4, lock mechanism 9, lock table TID list 10, and user interface unit 2 are placed on one of the processors (processor #n in this example), but are not shown. do. A method of distributing and arranging these mechanical parts to each processor is also conceivable. The database search/update unit 7 is located on every processor in which data is stored (on disk).

Processor #n stacks executable basic processing units in a queue corresponding to each processor on the shared memory according to instructions from the basic processing unit dependency graph 3 and the suspend/resume mechanism section 4. Each processor #1,
#2 and #3 are the queues #2 and #3 corresponding to you on the shared memory.
1, #2, and #3, the basic processing unit to be executed next is fetched and executed.

(Embodiment 2) FIG. 4 shows the configuration of the present invention on a parallel processor that does not have a shared memory.

The basic processing unit dependency relationship graph 3 and data arrangement dictionary 6 are placed on all processors. Queues #1, #2, #3 corresponding to each processor, graph execution section 5, suspend/resume mechanism 4, lock mechanism 9, lock table TID list 10, and user interface section 2 are also placed on each processor. Also, the database search/update section 7 is placed on every processor in which data is stored (on disk).

Assuming that records A, B, and C are distributed and stored in different processors #1, #2, and #3, each basic processing unit of a transaction, readA
, readB, readC are different processors #1, #2
, #3 in parallel. The correspondence between each record and each processor is stored in a data arrangement dictionary 6 provided in each processor #1, #2, #3. Therefore, each processor can know which processor will perform the basic processing unit of the transaction for each record.

Each processor #1, #2, #3 proceeds with its processing according to its own dependency graph 3 and suspend/resume mechanism 4. When processing data in another processor using the data arrangement dictionary 6,
The basic processing unit is sent to the processor and queued on that processor. In order to start the next basic processing unit after a certain basic processing unit is finished, one possible method is to broadcast that information to all processors via the network.

FIG. 5 shows another embodiment of the invention. This is the case where there is only one queue instead of one queue provided for each processor. At this time, put the basic processing unit and processor number in the queue.

[0021] Each processor takes from the queue the basic processing unit that is the same as its processor number. Processing that does not specify a processor indicates an undefined processor number state and can be performed by any processor. In this embodiment, exclusive control is required for queue operations.

[0022]

[Effect of the invention] When data to be processed is distributed and stored in processors in parallel transaction processing,
By performing processing on the stored processor, unnecessary data movement can be avoided. When performing parallel processing on a transaction-by-transaction basis, if data to be processed within a transaction is distributed among multiple processors, it becomes difficult to perform processing on the processor where the data is stored. Efficient execution is possible by breaking down a transaction into basic processing units and executing each basic processing unit in the processor where the data is stored.

[0023] By having a data arrangement dictionary, it becomes possible to execute the basic processing unit on the processor in which the data is stored. The system can be configured efficiently by changing the placement of the dependency graph, data placement dictionary, and execution queue according to the processor configuration.

[Brief explanation of drawings]

FIG. 1 is a diagram showing a dependency relationship graph.

FIG. 2 is a diagram showing the basic configuration of the present invention.

FIG. 3 is a diagram showing a configuration in the case of memory sharing.

FIG. 4 is a diagram showing a configuration in the case of memory distribution.

FIG. 5 is a diagram showing another embodiment of the present invention.

[Explanation of symbols]

1 Transaction 2 User interface 3 Basic processing unit dependency graph 4 Suspend/resume mechanism 5 Graph execution 6 Data placement dictionary 7 Database search/update 8 Database 9 Lock mechanism

Claims (6)

[Claims] Claim 1: In a transaction processing system that decomposes a transaction into basic processing units and executes them in parallel in basic processing according to data and control dependencies between the basic processings, a processor in which data related to the basic processing units is stored. A parallel transaction processing system that handles data distribution characterized by executing the basic unit. 2. In the transaction processing system, identifiers of processors in which data is stored are managed in a table, and the table is used to determine a processor that executes a basic processing unit.
A parallel transaction processing system that handles data distribution as described. 3. The parallel transaction processing system for handling data distribution according to claim 2, wherein a table for managing identifiers of processors in which data is stored is placed in a shared memory in the transaction processing system. 4. A parallel transaction processing system for handling data distribution according to claim 2, wherein in said transaction processing system, a table for managing identifiers of processors in which data is stored is placed on all processors. . 5. The parallel transaction processing system for handling data distribution according to claim 2, wherein executable basic processing units are stored in an execution queue for each processor. 6. A parallel transaction processing system for handling data distribution according to claim 2, characterized in that executable basic processing units are stored in one execution queue provided in common to each processor.