JPS63229522A - Parallel sorting process method - Google Patents

Abstract

PURPOSE:To extremely decrease the data transfer frequency between processors and to improve the processing efficiency by sending the elements of data allocated evenly and previously by each processor to other corresponding processors with each other excluding the allocation amount of its own when a partial charge range is designated. CONSTITUTION:The processors 1-1-1-m receive information from a control processor 4 to divide the range of the data value into (m) partial ranges in the number equal to the number of processors. Then a partial range to which each allocated element belongs is discriminated and sent to the processor to which said partial range is allocated with only the element allocated to its own left as it is. At the same time, the element sent to its own processor via the similar processes carried out by other processors in parallel with each other is received. When the mutual allocation is ended among elements, each processor sorts its own allocated data. Thus a sorting process of whole data is completed. In such a way, the data sorting process is completed just with a single time of transfer of data at most for each element.

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] This is a processing method for improving the processing efficiency of parallel sorting in a computer system consisting of a plurality of processing devices.

Allocating a sharing range that divides the data value range into m pieces to m processing devices, distributing the elements of the data to be sorted by each processing device to each other according to the sharing range, and regarding the distributed data, Complete the sorting of all data by performing each sort individually.

With this method, data exchange between processing devices for sorting is reduced, and processing can be executed efficiently.

[Industrial application field]

The present invention relates to a parallel sort processing method in a computer system including a plurality of processing devices.

Sorting, which arranges data in order of size, is widely used in various fields of information processing, as is well known.

[Conventional technology]

FIG. 2 is a block diagram showing an example of the configuration of a computer system.

Processing devices 1-1 to 1-m are storage devices 2-1 to 1-m, respectively.
It is a device that can read out data held in 2-m and independently perform sorting processing, etc., and can also exchange information with each other through a communication path constituted by network 3.

The control processing device 4 is connected to the processing device 1 via the network 3.
-1 to 1-m, etc., performs necessary control related to the overall system.

If a large amount of data is sorted by distributing the data to each processing unit 1-1 to 1-m of such a computer system and processing it in parallel, it is possible to speed up the processing, and such parallel processing A typical method is the known pitonic merge method.

According to this method, data to be sorted that is distributed to a plurality of processing devices is first subjected to the following processing for each pair of two processing devices.

That is, as shown in Fig. 3, a set of two processing devices A and B processes the data distributed to each processing device (the figure shows an example in which four elements are distributed each) in parallel and individually. For example, after sorting A in ascending order and B in descending order (■ in the figure), compare the corresponding elements of the image processing device data, and sort the data so that, for example, the smaller element becomes A and the larger element becomes B. Move (■ in the diagram).

By sorting the data in, for example, ascending order in each processing device, it is possible to sort across processing devices A and B (■ in the figure).

The above processing is executed in parallel on all the other sets of processing devices for their own data, and then the results of processing devices A and B are processed, and the other processing devices C and D process other data. Compare and move the data (■ in the figure) between elements in the same way as above (■ in the figure), and use the results as described above in each pair of two processing devices. By sorting in this way, it is possible to sort across processing devices A to D.

In this manner, the data length is sequentially extended and the same processing is carried out, such as 4 processing units x 2, then 8 processing units x 2, and so on, thereby completing the sorting over all data.

[Problem that the invention seeks to solve]

The above method allows for parallel processing of sorting, but as can be easily considered from the above explanation, it is possible to move 1/2 of all data elements to compare data between processing devices, and if necessary depending on the comparison result. data movement is required repeatedly for each part sorting process. For this reason, the amount of data that must be exchanged between processing devices is relatively large (therefore, there is a problem that the processing speed cannot be increased).

[Means for solving problems]

FIG. 1 is a process flowchart showing the configuration of the present invention.

The figure shows the processing flow of each processing device, and 30 to 32 are processing steps.

[For production]

Each processing device receives information for dividing the data value range into m sharing ranges equal to the number of processing devices in processing step 30.

In processing step 31, the allocation range to which each element of data distributed to each processing device belongs is identified, and the element is sent to the processing device to which that allocation range is allocated, and only the elements allocated to itself are identified. leave. It also receives elements sent to itself through similar processing performed in parallel by other processing devices.

After the mutual distribution of the elements is completed, in processing step 32, each processing device sorts the data allocated in processing step 31, thereby completing the sorting of the entire data.

This processing method allows data transfer to be performed at most once for each element to complete the sorting.

〔Example〕

In the computer system shown in FIG. 2, data to be sorted is divided into approximately equal number of elements and stored in m storage devices 2-1.
~2-m shall be distributed in advance.

m processing devices 1-1 to 1-m (hereinafter referred to as PE, ~
For example, the PE (hereinafter referred to as CP) receives instructions sent from a control processing device (hereinafter referred to as CP) 4 to all the processing devices via the network 3, and the storage device 2 connected to each processing device.
Input data from -1 to 2-m and start sorting processing.

Here, for example, CP4 transmits each upper limit value (at, a2, ~a++-1) of m-1 assignment ranges from the smallest value to all processing units PE as information indicating the boundaries of assignment ranges. ,~
Notify PE.

Each of PE, ~PE, is given a number from 1 to m, and assigned to each processing device in numerical order starting from the smallest value, and each processing device receives m pieces from the above upper limit. The assignment range is recognized as follows, where X is the value of the data element.

PR, sharing range X≦a.

PIE, sharing range a,<x≦a2PE, , sharing range aM-z<X≦a+++-1P[! Fi's allocation range a, -, <x It is desirable from the viewpoint of parallel processing efficiency that the allocation ranges determined in this manner be determined so that approximately the same number of elements belong to each range.

Therefore, for example, if it is known that the values of each element are almost evenly distributed, first PEl-
PE notifies CP4 of the maximum and minimum values of the elements of the winter hair retention data, and CP4 knows the value range of the data currently being processed by taking the maximum and minimum values from among them. , the upper limit value may be calculated so as to divide this range into approximately m equal parts, and these values may be notified to PR, ~PH,.

However, in order to ensure that elements always belong to each sharing range almost equally, even when the distribution of element values is biased, for example, as described below, rni/ For each m1th element, i from 1 to m
It is desirable to change the value to -1, find m~1 pieces, and set them as the upper limit of the sharing range. Here, [x] indicates the smallest integer greater than X (an integer obtained by rounding up the decimal part of X), and n is the total number of data elements.

That is, the PEs, ~PE, sequentially determine each upper limit value at the CP4 by exchanging information with the CP4 as described below for the data to be sorted inputted from the storage devices 2-1 to 2-m. We will notify you.

FIG. 4 shows the flow of the upper limit value determination process, and the left side shows the process of CP4, and the right side shows the process executed in parallel in PEI-PR.

CP4 initializes variable j to 1 in processing step 10, sets variable k to r n / m 1, and performs PE, ~P
Start processing E.

Each processing device PE of PE, ~PE, performs processing step 2.
The variable j is initialized to 1 with 0, and the data inputted from the storage device 2-i are sorted and set as St for the current process.

In processing step 21, the intermediate value old is determined and notified to the CP4.

In processing step 11, CP4 calculates the intermediate value M from m Mi notified from PE, ~PE, and calculates the intermediate value M.
l” P R, to be notified.

In processing step 22, each PEi classifies the elements of Si according to M, and the set of elements smaller than M is 511, the set of elements equal to M is Siz, the set of elements larger than M is Si3, and Si+, the elements of Siz are Each number bL and bit are notified to CP4.

In processing step 12, CP4 totals bi, , big notified from all processing devices to obtain totals b1 and b2.

In processing step 13, compare bl, b2 and k,
For example, if ≦b1, 1, b; if <k≦bl+b2, 2, 'b, 10s; if <k, 3 is notified to all processing devices. In the case of 2, if there is a next upper limit value process, the variable j is incremented by 1, and rnj/m1-k-bz+1 based on the new value of j is set as the new value of k. In addition, in the case of 3, bI + from the value of the variable
Decrease b2.

In processing step 23, each PEi follows the notifications 1 to 3, and if it is 1, then Si+, if it is 3, then SI3 is the new Si
is set, the process returns to step 21, and the process is restarted from the process of calculating the intermediate value using this Si.

2, the value of M at that time is stored as the j-th upper limit value a, and if j=m-1, the upper limit value determination process is terminated.

If j<m-1, the process proceeds to step 24 to determine the next upper limit value, and only the elements larger than the value of M, which is currently set as the jth upper limit value, out of the sorted results of the original data are used for subsequent processing. Set them as the first Si as processing targets, add 1 to j, return to processing step 21, and set the new jth upper limit value, that is, rnj/m]th from the bottom for all data elements, as the current processing target. CF2 processing step 1 for
As you set the new value in step 3, rnj/l1 from the bottom.
ll-k-bz+1 search processing begins.

When m-1 upper limit values a, ~a□, are determined in this way, each PEi recognizes each assignment range as described above, sets the first assignment range as its own assignment range, and data Only the elements within its own sharing range are left, and the other elements are transferred via the network 3 to other processing devices PE, ~PE, in their respective sharing ranges, and elements transferred from other processing devices to itself. I will take it.

As a result, if the data collected in each PEi is individually sorted in parallel, the sorting of all data from PE to PE is completed.

FIG. 5 shows an example in which data of 15 elements of contents 1 to 15 are sorted in parallel by the processing method described above by three processing devices PEI, PH2, and PE3.

In the figure, the data status of each processing device is shown in the indicator columns PE, PE, PE3, and the processing of CF2 is shown in the left column.

Fig. 5 Ta) shows the upper limit value determination process, and in step (2), CF2 sets k to r15X 1/31 = 5. Each P
It is assumed that data such as D1 to D shown in step (2) are input to Ei.

In step ■, the sorting result of each PEi becomes as shown in SI NS3, and each intermediate value range, ~-1, shown in step ■ is sent to CF2, so as in step ■, their intermediate value M=8 is It is determined and notified to each PEi.

As in step ■, each PEi is classified using the intermediate value M=8, and the number of elements in each set shown in step ■ is CF
2 will be notified.

In CF2, obtain the total number of elements as in step ■, and k
Since k<bt is obtained compared to =5, 1 is notified to all PEi.

Therefore, as in step ■, in each PEi, Si1 is set as new Si, and as in step [phase], the intermediate value
Notify F2. Here, if the number of elements is even, a small value is set as the intermediate value by agreement.

In step ■, M=3 is determined in CF2, and the process using the intermediate value is performed in the same way as before in step @~0, so that b,+b2<k, so the value of k is 5-b, -b
, = 2, and 3 is notified to all PEi.

Therefore, as in step [phase], in each PEi, St is set as a new Si, and as in step [phase], the intermediate value is notified to CF2. ,.

In step 0, M・5 is determined in CF2, and the process using its intermediate value is performed in the same way as before in steps [phase] to [phase], resulting in b, +b, = k, so 2 is the total PEi
The current M=5 is determined to be the fifth largest element and becomes the first upper limit.

Next, by setting r15X 2 / 31 =10 and performing the same process as described above (not shown), 10 is determined as the second upper limit value.

FIG. 5 (bl) shows a situation in which each PEi mutually distributes data according to the upper limit values 5 and 10 determined above, and the resulting data is sorted in parallel.

Since the upper limit values are 5 and 10, 5 or less is the first processing device PE, 6 to 10 is PE2, and anything larger than 10 is the scope of PE3, and from the data on hand shown in step 0, the processing device of each place Transfer the elements in the assigned range like step @.

As a result, each processing device collects data as shown in step 0, and each processing device sorts the data in parallel to obtain sorted results arranged in ascending order across the three processing devices, as shown in step 0. Can be done.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, processing efficiency is significantly improved by significantly reducing the number of data transfers between processing units in parallel sort processing of a computer system having a plurality of processing units. It has industrial effects.

[Brief explanation of drawings]

Fig. 1 is a flowchart of processing showing the configuration of the present invention, Fig. 2 is a block diagram of an example configuration of a computer system, Fig. 3 is an explanatory diagram of the pitonic merge method, Fig. 4 is a flowchart of upper limit determination processing, FIG. 5 is an explanatory diagram of an example of sorting processing according to the present invention. In the figure, 1-1 to 1-m are processing devices, 2-1 to 2-m are storage devices, 3 is a network, 4 is a control processing device, and 10 to 1
3.20-24.30-32 indicate processing steps.

Claims (2)

[Claims] (1) In a computer system in which m processing devices are connected via a communication path, each processing device executes in parallel the sorting of data consisting of a plurality of elements distributed to each processing device. At , a predetermined notification indicating a sharing range obtained by dividing the value range of the data element into m pieces is received (30), and the data element held by the processing device is divided into m pieces of sharing range. The elements are mutually distributed to the respective processing devices assigned to handle the respective tasks (31), and a sorting process is performed on the distributed elements (3
2) A parallel sort processing method characterized by the following. (2) The predetermined notification indicating the allocation range has a predetermined magnitude relationship with ni/m, where the number of elements of the data is n and i is an integer taking a value from 1 to m-1, ni/
The value of each j_i-th size element from the smallest of the data elements, where j_i is the integer closest to m,
2. The parallel sort processing method according to claim 1, wherein said value is used as a value indicating a predetermined boundary between said m assignment ranges to identify said assignment range.