JPH0782427B2 - Sort processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention relates to a sorting method in the field of information processing,
That is, the present invention relates to a processing method for arranging a large amount of data in a fixed order based on the value of a specific field.

[Conventional technology]

FIG. 6 is a block diagram showing a conventional sorting method. In the figure, 27 is a CPU (central processing unit), 28 is a memory, and 29 and 30 are data on the memory 28.
9 indicates an unsorted record group and 30 indicates a sorted record group.

FIG. 7 is a diagram showing a format of records to be sorted in the sort processing method of FIG. In the figure, 31 is m
It is a record of a format consisting of a KEY part of bytes and a non-KEY part of n bytes. Here, m is a natural number and n is 0 to a natural number.

In the conventional sort processing method shown in FIG.
With an ordinary computer consisting of CPU 27 and memory 28,
The sort process is executed as follows. That is, after the unsorted record group 29 is read into the memory 28 from an input / output device (not shown) such as a disk device, it is shaped and the data of the unsorted record group 29 is created in the memory 28. The unsorted record group 29 is an array of records 31 in the format shown in FIG. For the unsorted record group 29, a sort program (not shown) existing in the memory 28 is executed, and the sorted record group 30 is created in the memory 28. As an algorithm of the sort program, for example, "My Compute" issued by CQ Publisher
r ", 1985, No. 18, pp. 84-85.

[Problems to be solved by the invention]

In the conventional sort processing method as described above, the sort processing is performed by software on an ordinary computer composed of one CPU 27 and one memory 28. When the number of records in the unsorted record group 29 is N, it is necessary to perform O (Nlog ₂ N) comparison operations even if the fastest algorithm such as quick sort when sorting by software is used. It has been known. Therefore, when the number of records N becomes large, there is a problem that the time required for the sorting process becomes very long.

The present invention has been made to solve the above problems, and an object of the present invention is to obtain a sort processing method capable of performing sort processing more quickly.

[Means for solving problems]

According to the sort processing method of the present invention, all records to be sorted are distributed as partial record groups in a plurality of CPU local memories each having a local memory, and each CP
U sorts the partial records in its own local memory, transfers the sorting results of some CPUs among multiple CPUs to the common memory via the bus, and then transfers them to the common memory. The sort result and the sort result of some other CPU not transferred to the common memory are merge-sorted.

[Action]

In the sorting method of the present invention, a plurality of CPUs each having a local memory, for example, two CPUs out of four CPUs sort the partial record groups each in its own local memory. Then, each sort result is transferred to the common memory by performing the fetch / store operation via the bus, and the sort result transferred to the common memory and some other parts not transferred to the common memory are transferred.
If you do merge sort with the sorting result of CPU,
The frequency of access to the common memory is lower than that when four CPUs perform fetch / store operations all at once (sufficiently twice), so memory contention / bus contention to the common memory is less likely to occur. , Sort processing can be executed at high speed.

〔Example〕

FIG. 1 is a block diagram showing a sort processing method according to an embodiment of the present invention. In the figure, 1 is a common memory, 2 is a common bus, and 3,4,5,6 are CPU (I) and CPU, respectively.
(II), CPU (III), and CPU (IV). 7,8,9,10 are local memory (I), local memory (II), local memory (III), local memory (IV),
These are local memories of the CPUs (I) 3 to (IV) 6. Whether each of the CPUs (I) 3 to (IV) 6 accesses its own local memory or the common memory 1 is divided according to the memory address range. Each CPU (I)
The programs executed by 3 to (IV) 6 are normally stored in the local memory. 11, 12, 13, and 14 are interrupt interface signals for notifying an interrupt, and 15 is an interrupt control device, which is used by one CPU to generate an interrupt to another CPU.

2, 3, 4, and 5 are respectively the first
It is a figure for demonstrating the content of the sort process in the sort process system of the figure. 2 to 5, 16,17,1
8,19 are unsorted partial record groups, 20,21,22,23,
24 and 25 are sorted partial record groups, and 26 is a sorted record group which is the final result.

Next, the sort processing of the sort processing method according to the embodiment of the present invention shown in FIG. 1 will be described with reference to FIGS. In FIG. 2, the unsorted partial record groups 16 to 19 are loaded into the local memories (I) 7 to (IV) 10 from an external input / output device (not shown). Each of the unsorted partial record groups 16 to 19 is divided into four partial record groups in advance so that all the records to be sorted have substantially the same number of records.
When the unsorted partial record group 16 is generated in the local memory (I) 7, the CPU (I) 3 executes a sort program to perform the sort processing, and generates the sorted partial record group 20. . The quick sort algorithm described above is used as the algorithm in the sorting program. Similar to CPU (I) 3, each CPU (II) 4, CPU (I
In II) 5 and CPU (IV) 6, the same operation is simultaneously performed in parallel. When the number of records is N, the quick sort algorithm performs an average of 1.39 Nlog ₂ N comparison operations. Compared to the case of sorting the total number of records N,
Here about Sorts the number of records so an average of 1.39 It is sufficient to perform the comparison operation once. If you find this ratio, And the time to sort the entire record It can be executed in the following times.

When the above partial sort is completed, the odd-numbered CPU
The (I) 3 and the CPU (III) 5 respectively transfer the sorted partial record groups 20 and 22 to the common memory 1 as shown in FIG. That is, each of the CPUs (I) 3 and (III) 5 fetches sorted data from the local memories (I) 7 and (III) 9 and stores the sorted data in the common memory 1. This means that the two CPUs access the common memory 1 only once out of the two memory accesses, even if only for the data access. The access ratio to 1 is further reduced, and therefore, there is a feature that memory contention / bus contention for the common memory 1 between two CPUs is unlikely to occur.

If four CPUs simultaneously transfer the sorted data to the common memory 1, memory contention / bus contention for the common memory 1 is likely to occur and the transfer speed becomes slower. As for the time required for the above data transfer processing, if the same unit (time required for one comparison operation) as for quick sort is used, It is estimated that Now, when the CPU (I) 3 finishes such an operation, the C (via the interrupt interface signal 11, the interrupt controller 15 and the interrupt interface signal 12)
Tell PU (II) 4 to that effect. In response to this, the CPU (II) 4 starts processing of the contents shown in FIG. That is, the partial sort result of the sorted partial record group 20 of the CPU (I) 3 stored in the common memory 1 and its own sorted partial record group 21 stored in the local memory (II) 8.
The partial sort results are sequentially read, the 2-way merge sort processing is executed, and the partial sort result of the sorted partial record group 24 is stored in the common memory 1. For 2WAY merge sort processing, for example, "My Com" issued by CQ publisher.
puter ", 1985, No. 18, p. 91. Also, as for CPU (IV) 6, as shown in FIG.
The same processing as (II) 4 is simultaneously executed in parallel. In this case, when considering only data access, CPU (II) 4
Since the CPU and the CPU (IV) 6 each access the common memory 1 twice in three times, the memory contention is larger than that shown in FIG. The time required for this process is approximately It is estimated that

In this way, when the CPU (II) 4 and the CPU (IV) 6 have completed the above processing, the CPU (II) 4 and the CPU (IV) 6 respectively receive the CPU
(I) Notify 3 to that effect. CPU (I) 3 is for each CPU (I
When both I) 4 and CPU (IV) 6 know that the above processing has been completed, the processing shown in FIG. 5 is started. That is,
The merge sort result of the sorted partial record group 24 generated by the CPU (II) 4 on the common memory 1 and the CPU (IV)
The merge sort result of the sorted partial record group 25 generated in 6 is sequentially read, and the 2-way merge sort process is executed to generate the final merge sort result of the sorted record group 26. The processing time for this is estimated to be approximately the number N of records. And when all the above processing times are summed up, Becomes

On the other hand, when the sorting process is performed by the conventional sorting process method, the processing time of 1.39 Nlog ₂ N is required. For example, if N = 12 ¹⁴ = 16384 records, log ₂ N = 14, And the respective processing times are as follows.

1.39Nlog ₂ N = 19.46N That is, when the sorting process is performed by the sorting method of the present invention, the sorting process is performed by the conventional sorting method, It can be seen that the sorting process can be executed twice as fast.

In the above embodiment, the case where four CPUs are used has been described, but it is obvious that the present invention can be easily applied to the case where eight CPUs are used, for example.

Further, in the above embodiment, the CPU (I) 3 is adapted to perform the final sort processing shown in FIG.
(II) 4 to (IV) 6 may do this, or a CPU separate from these may be further connected to the common bus 2 and executed by that CPU.

In the above embodiment, the sort process shown in FIG. 2 is performed using the quick sort algorithm.
Other sort algorithms may be used.

Further, in the above embodiment, as shown in FIG.
3 and CPU (III) 5 transfer the sorted data to the common memory 1, but each CPU (I) 3 to (IV) 6
All four CPUs transfer the sorted data to the common memory 1, and the sort processing as shown in FIG.
You may do it above.

〔The invention's effect〕

As described above, according to the present invention, in the sort processing method, all records to be sorted are distributed as partial record groups in the local memory of each CPU having a local memory, and each CPU allocates its own local memory. Sorts the partial records contained in it, transfers the sorting results of some CPUs out of multiple CPUs to the common memory via the bus, and transfers the sorting results transferred to the common memory to the common memory. Since the sort results of some other CPUs that have not been transferred are merge-sorted, memory contention / bus contention for common memory is less likely to occur, and the sorting process It has an excellent effect that the speed can be greatly improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a sort processing method according to an embodiment of the present invention, FIG. 2, FIG. 3, FIG. 4, and FIG.
FIG. 6 is a diagram for explaining the contents of sort processing in the sort processing method of FIG. 1, FIG. 6 is a block configuration diagram showing a conventional sort processing method, and FIG. 7 is a sort processing method of FIG. It is a figure which shows the format of the record which should be sorted in. In the figure, 1 ... Common memory, 2 ... Common bus, 3 ...
6 ... CPU (I) to CPU (IV), 7 to 10 ... Local memory (I) to local memory (IV), 12 ... 14 ... Interrupt interface signal, 15 ... Interrupt control device, 16-19
...... Unsorted partial record group, 20 to 25 …… Sorted partial record group, 27 …… CPU, 28 …… Memory, 29 …… Unsorted record group, 26,30 …… Sorted record group. In the drawings, the same reference numerals indicate the same or corresponding parts.

Claims (1)

[Claims] 1. A plurality of CPUs each having a local memory
And a device having a common memory accessible to each CPU and a bus connecting the common memory and each CPU, all records to be sorted are distributed in a partial record group in the local memory of each CPU. The above CPs
U sorts the partial record group in its own local memory, transfers the sorting result of some CPUs of the plurality of CPUs to the common memory via the bus, and A sort processing method characterized in that the sort result transferred to the memory and the sort result of some other CPUs not transferred to the common memory are merge-sorted.