JPS63132333A - Association coupling operating device - Google Patents

Abstract

PURPOSE:To speed up the coupling operation of a related data base by introducing the parallel processing mechanism of a macro level and a micro level, and a pipeline processing mechanism between processings. CONSTITUTION:A processing request of the coupling operation reaches a main control processor 7 from a host processor 5 through an LAN 6. The main control processor 7, receiving the processing request, requests an auxiliary control processor 20 of respective unit 8 to process it, through a connection 11. The auxiliary control processor 20 reads out an object attribute data from a unit corresponding data base storage 22, and starts the processing to load a unit corresponding association memory 21 all at once. The auxiliary processor 20 requests a broadcasting mechanism 23 to process a broadcast processing. The broadcasting mechanism 23 reads out the object attribute of the relation of another side of a coupling object from the unit corresponding data base storage 22, and transfers it to other unit through a bus 9.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Fields] The present invention provides a parallel processing mechanism at the macro level (between processes, between units) and at the micro level (word parallelism in associative memory) and a pipeline processing mechanism between processes. The present invention relates to an associative join operation device that executes join operations on relational databases at high speed.

[Prior Art] Join operations in relational databases are often used and are known as typical operations that require a heavy processing load.

There are two main processing methods for join operations: the nested loop method and the ordered join method.
(red merge) method. There are efficient methods for dedicated architectures, one of which is a method using associative memory. A conventionally known join operation device using this associative memory loads one relation to be joined into an associative memory, searches the associative memory for each tuple of the other relation in turn, and then satisfies the condition. The idea was to join between tuples that satisfy the condition. If the number of tuples in the relation to be processed is n, it would take 0 (n2) time without any modification, but with the conventional method, the search in the associative memory can be performed in word parallel, so it will take less time.
It is calculated that it will only take (n) hours. However, in reality, the loading process to the associative memory and the tuple joining process have a processing load that cannot be ignored.

[Problems to be Solved by the Invention] A conventional combination arithmetic device using an associative memory is shown in FIG. Further, FIG. 8 shows a time chart of the join operation processing in the conventional join operation device. In the conventional technology, the process of loading the target attribute data of one of the relations to be combined stored in the database storage 3 into the associative memory device 2, and the process of searching the associative memory device 2 and then combining the tuples that satisfy the two join conditions. Because there is only one internal bus or channel 4 that transfers data, and because there is only one control processor 1 connected to the host processor 5 via a line or LAN 6, the processing is serial and requires a lot of processing. It took time.

[Means for Solving the Problems] It is an object of the present invention to solve the processing load problem of load processing to associative memory and tuple join processing that conventional associative arithmetic units using associative memory have, by solving them at a macro level. The object of the present invention is to provide an associative connection arithmetic device that solves the problem by introducing a parallel processing mechanism and a pipeline processing mechanism.

In order to achieve the above object, the associative connection arithmetic device of the present invention has a sub-control processor that controls the operation of the unit, an associative memory device that connects a large number of associative memory LSIs so that they can be linked together, a database storage that stores a relational database, and a unit. A broadcasting mechanism that performs data broadcasting between units, and a plurality of internal buses or channels that interconnect these units, a broadcasting bus that connects each unit, and a host processor that controls communications and unit groups. It is composed of a main control processor that loads one relation in the database storage to the associative memory device in parallel by the number of units, and broadcasts the other relation in the database storage in parallel with the load processing, and then performs the load processing. As soon as the process is completed, the associative memory device is searched, and the combinations of doubles that meet the conditions are performed in parallel by the number of units, thereby parallelizing the processing at the macro/micro level. If the load process is delayed, the associative memory device search process and tuple join process are executed in the broadcast process and pipeline as soon as the load process is completed, thereby processing join operations in the relational database at high speed. It is characterized by.

[Operation] In the present invention, by adopting a plurality of unit configurations, the process of loading the target attribute data of one of the relations to be combined stored in the unit-based database storage into the unit-based associative memory device;
Since the process of combining tuples that satisfy the combination condition can be executed on a unit-by-unit basis, the drawbacks of the prior art (processing load problems of loading processing into associative memory and tuple joining processing) can be solved.

Note that by dividing into units, it is necessary to broadcast the target attribute data of the other relation to be combined to all units, but this is executed in parallel with the load process, and the broadcast process takes longer than the load process. In some cases, by performing search processing and pipeline processing, it is possible to apparently prevent the problem from appearing in the processing time.

[Example] FIG. 1 shows an entire example of the present invention. A specific example of a plurality of units 8 among them is shown in FIG.

FIG. 3 shows a time chart of the join operation process in the present invention. The progress of operations according to a specific example of the present invention is shown in FIGS. 4 to 6, focusing on the contents of the unit-based associative memory device 21 and the unit-based database storage 22. In FIGS. 4 to 6, the number of units is set to two for simplicity, but in reality, more units are prepared.

Hereinafter, one embodiment of the present invention will be described in detail.

The main control processor 7 coupled to the host processor 5 and the sub-control processors 20 in each unit 8 can be realized using, for example, a general-purpose multiprocessor. The unit-based associative memory device 21 is configured such that a plurality of associative memory LSIs can be linked together. The unit-based database storage 22 is composed of a large capacity memory or a magnetic disk. The broadcast bus arbiter 10 may be the same as a bus arbiter for a normal bus (for example, a vME bus).

The difference between the broadcast bus 9 and an ordinary bus (e.g. VME bus) is that in data transfer, the broadcast bus 9 operates like an ordinary bus (point-to-point) paint-
Not to-paint (point-'7-φ multipoint) port-to-multipotnts
There is a certain point.

In one embodiment of the present invention, each piece of data is assumed to have a fixed length. All numerical data becomes fixed length by converting it to the computer's internal format. For example, the variable length character string is
By using the "Pipeline String Search Algorithm Based on Associative Memory" published in Vol. Since it can be converted to a fixed length, the assumption of a fixed length does not impair generality to the data to be managed.

Further, it is assumed that the database data has already been equally divided and stored for every relation in the unit-corresponding database storage 22 of each unit.

First, a processing request for a join operation arrives from the host processor 5 via a line or LAN 6 to the main control processor 7 . Upon receiving the processing request, the main control processor 7 requests processing to the sub-control processor 20 of each unit 8 via the bus or the connection 11. In the specific example shown in FIG. 4, it is assumed that a join operation process with the condition (first attribute of relation R) - (second attribute of relation S) is requested.

Each sub-control processor 20 that has received a processing request from the main control processor 7 reads the target attribute data of one of the relations to be combined from the unit-based database storage 22 and starts the process of loading it into the unit-based associative memory device 21. do. The data is of fixed length (e.g. 4
(byte), the load process is performed by storing one value of target attribute data of one of the relations to be combined in one word of the unit-based associative memory device 21.

As shown in FIG. 3, each unit 8 executes a process of broadcasting target attribute data of the other relation to be combined to all units in parallel with the load process. That is, first, the sub-control processor 20 requests the broadcast mechanism 23 to perform broadcast processing by specifying the target attribute of the other relation to be combined. The requested broadcast organization 23 requests the broadcast bus arbiter 10 to use the broadcast bus 9. The broadcast bus arbiter 10 grants the right to use the bus to one of the broadcast mechanisms 23. The broadcasting mechanism 23 that has been given usage rights by the broadcasting bus arbiter 10 reads out the target attribute data of the other relation to be combined from the unit correspondence database storage 22 of its own unit, and sends the broadcast type to the other unit via the broadcasting bus 9. transfer. The broadcasting mechanism 23 of the unit other than the broadcasting mechanism 23 that broadcasts receives the data to be broadcast, and the unit corresponding database storage 2
Store in 2. When the broadcast is completed, the broadcast bus arbiter 10 grants the right to use the bus to any of the remaining broadcast mechanisms 23. The broadcasting mechanism 23 that has been given usage rights by the broadcasting bus arbiter 10 reads out the target attribute data of the other relation to be combined from the unit correspondence database storage 22 of its own unit, and sends the broadcast type to the other unit via the broadcasting bus 9. transfer. In this way, each unit completes broadcasting in sequence, thereby completing the entire broadcasting process. FIG. 5 shows the state at this point in time (the point at which the loading process and broadcasting process are completed) in a specific example. The first attribute value of relation R is unit-compatible associative memory device 2
1, and relation S is restored to its entire form by broadcast processing.

If the broadcast processing takes longer than the load processing, the broadcast processing and the search processing are processed in a pipeline. That is, since a part of the target attribute data of the other relation to be combined exists in the own unit, it is possible to start a search process for the unit-corresponding associative memory device 21. While the search is being performed, the broadcast process progresses, and the search process for the unit-based associative memory device 21 is continued using the broadcast target attribute data of the other relation to be combined. When the search process is too fast compared to the broadcast process and there is a possibility of overrun, it is controlled using the semaphore P, which is well known in the field of operating systems.
■This can be done using operations, etc. As described above, broadcast processing and search processing pipeline processing can be performed.

The search process for the unit-compatible associative memory device 21 is as follows:
Parallelism of associative memory (word parallel and bit parallel)
This makes condition pass/fail judgment extremely high-performance (several steps)
Can be executed. Generates a pair of tuples that meet the conditions as a result relation. This tuple joining process is performed in parallel for each unit. FIG. 6 shows the state of completion of tuple join processing according to a specific example.

When the tuple joining process is completed, the sub-control processor 20 notifies the main control processor 7 to that effect. After receiving notification of the completion of the combination process from all the sub-control processors 20, the main control processor 7 notifies the host processor 5 of the completion of the combination process via the line or LAN 6.

[Effects of the Invention] As explained above, according to the present invention, in nested rube type processing of a relational database join operation, the target attribute data of one of the relations to be joined stored in the database storage is stored in the associative memory device. The process of loading and the process of joining tuples that satisfy the join condition are
It can be executed in parallel for each of the multiple units provided, and furthermore, the process of broadcasting the target attribute data of the other relation to be combined to all units, which is required by dividing into units, is executed in parallel with the load process. If the broadcast processing takes longer than the load processing, by performing pipeline processing with the search processing, it is possible to superficially prevent it from appearing in the processing time, making it possible to execute the join operation at high speed.

[Brief explanation of the drawing]

FIG. 1 is an overall diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing a specific example of some devices of the present invention, and FIG. 3 is a time chart explaining the operation of an embodiment of the present invention. Figures 4 to 6
The figures up to the figure are schematic diagrams of the operation progress according to the specific example of the present invention, and FIG. 7 shows the prior art. FIG. 8 is a diagram showing a time chart of the operation of the embodiment of the prior art. 1... Control processor, 2... Associative memory device, 3
...Database storage, 4.Internal bus or channel, 5.Host processor, 6.Line or LAN, 7.Main control processor, 8.Unit, 9.Broadcast bus , 10... Broadcast bus arbiter, 11... Bus or connection, 20... Sub control processor, 211.・Unit compatible associative memory device, 22
...Unit compatible database storage, 23...
- Broadcast mechanism, 24...internal bus or channel. Applicant's Representative Patent Attorney Takehiko Suzue Figure 2 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8

Claims (1)

[Claims] A sub-control processor that controls the operation of the unit, an associative memory device that connects a large number of associative memory LSIs so that they can work together, a database storage that stores a relational database, a broadcasting mechanism that broadcasts data between units, and interconnects them. A plurality of units consisting of internal buses or channels, a broadcast bus connecting the units,
It consists of a main control processor that communicates with the host processor and controls a group of units, and loads one relation in the database storage to the associative memory device in parallel by the number of units, and loads all the other relations in the database storage in parallel with the loading process. By broadcasting the relation, as soon as the loading process is completed, the associative memory device is searched, and the tuples that meet the conditions are connected in parallel by the number of units, thereby parallelizing the processing at the macro/micro level. In addition, in the above procedure, if the completion of broadcast processing is delayed from the completion of load processing, as soon as the load processing is completed, the associative memory device search processing and tuple join processing are executed in the broadcast processing and pipeline, thereby updating the information in the relational database. An associative join operation device characterized by high-speed processing of join operations.