JPS6249557A - Data distributing system for parallel computer - Google Patents

Abstract

PURPOSE:To improve the data access efficiency with parallel computers by storing the data on the data number that is decided by an equation I into a cell of the i-th row and the j-th column with the dividing number of data defined as N. CONSTITUTION:Many cells are arrayed in a 2-dimensional grid form and the cells adjacent in four directions are connected to the parallel computers via communication channels. In such parallel computers, the data on the data number decide by an equation I is stored in a cell of the i-th line and the j-th column [i, j: positions in square grid of NXN set at value 0-(N-1)] when the dividing number of data is defined as N (=2L<2>+2L+1, where L means a natural number). The diagram shows an example where a piece of data is divided into five parts 0-4 for distribution to form a data distribution pattern. This pattern fills a (5X5) 2-dimensional square grid and any of these cells can give an access to the optional data within the distance '1'. Furthermore the minimum average access distance is obtained among the distribution patterns of the same number of divisions.

Description

[Detailed Description of the Invention] [Summary] This is a method of distributing a large amount of data on a parallel computer so that all the data exists within a certain distance from any cell.

[Industrial application field]

Recently, highly parallel computers have been attracting attention in order to improve computer performance. In particular, parallel computers in which cells are arranged in a two-dimensional grid are promising because of the ease of implementing communication paths between cells and the ease of handling two-dimensional data. FIG. 5 is a diagram showing an example of this type of parallel computer. In Figure 5,
1 is a host computer, 2 is an interface, 3 is a video controller, 4 is a camera, 5 is a display, 6-11 or 5-nm is a cell, C-BUS is a common bus, V-Bus is a video bus, VC -BUS indicates a video control bus, and P indicates a communication path. The cell 6 is composed of a microcomputer or a minicomputer.

A communication path P is provided between adjacent cells 6. In order to avoid a decrease in efficiency due to access to common data, it is desirable that each cell 6 independently hold information from all over the world and operate autonomously. However, in reality, a large amount of data, image data, etc. is too large for one cell 6 to hold. It would be convenient if such data could be divided, distributed and stored in a plurality of cells 6, and handled as if it were data within the own cell 6. Although this function can be realized by using a common bus, as the number of cells 6 increases, efficiency is inevitably lowered due to competition for the common bus. Therefore, it is desirable to realize data access using an adjacent communication network.

[Purpose of the invention]

The present invention is based on the above consideration, and an object of the present invention is to provide a method for distributing and arranging data so that data can be accessed efficiently in a parallel computer as described above.

[Means to achieve the purpose]

Therefore, in the data distribution arrangement method for a parallel computer of the present invention, in a parallel computer in which a large number of cells are arranged in a two-dimensional grid and connected to adjacent cells in four directions through communication paths, the number of data divisions is N. (However, N=2L" +
2L+1, where L is a natural number), the i-th row and j-th column (where i.

D = mod (i + (2L + 1), N) 0≦i, j
It is characterized in that data of a data number determined by ≦N-1 is stored.

[Embodiments of the invention]

Distributed data arrangement> It is assumed that the adjacent communication channels are connected in a two-dimensional square lattice pattern and are folded back at the top, bottom, left and right ends to form a torus-like connection. The arrangement pattern when distributing data must have the following characteristics.

■ Data placement can be determined independently for each cell.

■ Average access distance and maximum access distance are small.

■ The surface of the torus is not regularly filled with data.

As a method that satisfies the conditions (1) to (4), the present invention proposes a method of distributing data to cells within a distance of L from the own cell.

FIG. 1 shows an example of a data distribution arrangement pattern when one data is divided into five parts O to 4 and arranged. In this example, the maximum access distance is 1.0 and the average access distance is 0.8. This layout pattern is 5×
5, and any data can be accessed within a distance of 1 in any cell. Furthermore, the average access distance is the smallest among layout patterns with the same number of divisions.

FIG. 2 is a diagram showing an example of the maximum distance, the number of divisions, the average distance, and the arrangement pattern. As shown in FIG. 2, when the maximum distance is L, the number of divisions N is N.2L2+2L+1. Also, the average distance is.

When the divided data is numbered 0-N-1, the data number that each cell should have is as shown in the following equation.

D = mod (i+(2L+1) books j, N) 0≦
i, j ≦N-1 (11 The divided data exists within a distance from the center of every cell, but the location of the cell that has the data differs for each cell. This location , is obtained from the data number r that one owns and the data number m to be accessed.

mod (m-r, N)=mod(i+(2L+1)*
J, N) - ≦1+'j ≦t,, H+Ijl≦L
(2) That is, (t, j) that satisfies equation (2) indicates the relative position of the cell having the data to be accessed from the own cell.

<Data Access Method> There are two methods for accessing data distributed in each cell: a method that uses a common bus, and a method that uses adjacent communication channels. The method of using a common bus is effective when all cells require the same data at the same time. On the contrary,
The method of using adjacent communication channels is effective when each cell reads data randomly.

Distributed data is accessed from one cell as a virtual file image. Data held by one cell is requested by a plurality of other cells, but it is necessary to avoid a situation in which one cell is flooded with these requests at the same time as much as possible. For this reason, each fixed-length record is stored in a form that is sequentially assigned to a cell group one code at a time. Third
The figure shows the relationship between cells and records when the number of divisions N is 5, where Ci (i-0, 1,..., 4) is the cell that holds data with data number i, and Rj is the cell that holds data with data number i. The j-th record is shown respectively. By doing this, data concentration can be alleviated even when cells sequentially access data.

The data distributed storage module works in a multitasking environment and consists of the following tasks and libraries used by users to access data.

(a) User task UT (bl Relay task RT (C) Data management task DMT User task UT is a user program and requests data from data management task DMT via data access routine DAR. Relay Task RT is a task that relays messages and data to distant cells.

The data management task DMT is a task that stores and manages distributed data.

FIG. 4 shows the communication status of tasks in the case of reading. The data access routine DAR determines which cell has the data requested by the user task UT. If it has it, it requests the data from the data management task DMT of its own cell (■). If any of the cells adjacent in the four directions has it, a request is made to the data management task DMT of that cell (■). In other cases, a request is made to the relay task RT of the adjacent cell in the direction approaching the cell that has the data (■, ■). The relay task RT that has received the data request further requests data from the adjacent relay task RT or data management task DMT (■,
). When the data management task DMT receives a data request, it sends the requested data. At this time as well, if necessary, the process passes through the relay task RT.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, data can be distributed in a parallel computer configured by arranging cells in a two-dimensional grid without reducing processing efficiency.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of a data distribution arrangement pattern when one data is divided into five parts O to 4 and arranged.
Figure 2 shows an example of the maximum distance, number of divisions, average distance, and arrangement pattern, Figure 3 shows the relationship between cells and records when the number of divisions N is 5, and Figure 4 shows the relationship between cells and records when reading. FIG. 5, which is a diagram showing the communication status of tasks, is a diagram showing an example of a parallel computer in which cells are arranged in a two-dimensional grid. 1... Host computer, 2... Interface, 6-11 to 5-nm-cell, C-BUS...
Common bus, P...communication path.

Claims (1)

[Claims] In a parallel computer in which a large number of cells are arranged in a two-dimensional grid and connected to adjacent cells in four directions through communication paths, the number of data divisions is N (however, N=2L^2+2L+1 where L is a natural number), the i-th row and j-th column (where i
, j indicates the position in the N×N square grid and takes a value from 0 to N-1), 0=mod(i+(2L+1)*j,N) 0≦i, j≦N-1 A data distribution arrangement method in a parallel computer characterized by storing data of a data number determined by .