JPH07114515A - Decentralized memory computer with network for synchronous communication - Google Patents

Abstract

PURPOSE:To provide the decentralized memory computer with the network for synchronous communication which is superior in communication efficiency. CONSTITUTION:This computer is provided with plural node computers N11 and N12 which have processors, communication means, and local memories for communication, a node connection network which connects those and interchanges data among them. Further, synchronous memories for variables and synchronous communication means that should be synchronized among all the processors are provided. Further, the respective synchronous communication means are connected to a network different from this node connection network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a distributed storage memory parallel computer.

[0002]

2. Description of the Related Art A vector computer such as a super computer has been conventionally used to execute a scientific and technological program written in a language such as Fortran. In recent years, however, a parallel computer has come to be used. Came.

This parallel computer includes a distributed memory parallel computer capable of increasing the number of node computers according to the amount of data handled by a program, and is particularly suitable for scientific and technological computation. The configuration of this distributed memory parallel computer is
As shown in FIG. 2, a plurality of node computers having a processor, a communication unit, and a local memory used for communication,
It has a node connection network for connecting these and exchanging data mutually. This node computer is almost the same as an ordinary small computer except for communication means. The node connection network is a network for connecting several to several thousand node computers, and various topological networks such as hypercubes, meshes, tori, and trees have been developed and used.

[0004]

A so-called single program multiple data (hereinafter referred to as SPMD) method is known as a method for performing parallel calculation by a plurality of node computers using such a node connection network. There is.

As an example, as shown in FIG. 3, in this program, as data, a real number s and a real number array a (i), b having 1000 elements of two are provided.
(I), a (i) is calculated by referring to b (i) in the do loop of S1, and d of S1 is calculated in the do loop of S2.
The sum of the array a (i) calculated in the o loop is calculated. In order to perform this calculation in parallel and at high speed, the arrays a (i) and b (i) are divided into 500 arrays each, and SP
An example of a program executed by the node computers N11 and N12 by the MD method is schematically shown in FIGS. 4A and 4B (for the sake of explanation, although not perfect in details as described later). If you create a program like this,
In the program shown in FIG. 3, it can be expected that the arrays, each of which has 1,000 arrays, are halved in each node computer, and that if they are simultaneously calculated, they can be executed at high speed.

By the way, a used in the examples given here
Variables such as (i) and b (i) that are distributed and stored in a plurality of node computers are called multivalues (poly), and be careful when referencing such multivalues in a program. Requires. This is because if the programs of FIG. 4 are executed at the same time, the node computer N11 will run on the node computer N11.
Since b (501) corresponding to b (1) on 12 does not exist, the expected calculation result cannot be obtained or an error occurs when i = 500, and the node computer N12 is on the node computer N11. A above (50
Since a (0) corresponding to 0) does not exist, similarly, the expected calculation result cannot be obtained or an error should occur. Therefore, from the node computer N11 to the node computer N12, a (500) on the node computer N11
From the node computer N12 to the node computer N11
, B (1) on the node computer N12 must be sent.

When referencing the multiple values on each node computer in this way, it is necessary to communicate the data such as the boundary dividing the array between the adjacent node computers. In the example of FIG.
Since the one-dimensional array is divided, only one piece of data needs to be communicated between the nodes, but in general scientific calculation, the array is often two-dimensional or more. Must communicate and exchange data.

The data communication A caused by the program that refers to the multiple values generally has the following features. A1. Communicate blocks containing multiple data such as array elements. A2. Basically, it is one-to-one communication between nodes, and when viewed as a whole node computer group, it is data communication at the boundary, and the nodes that communicate are often adjacent to each other, and local communication Is.

On the other hand, the program d in FIG.
In the o-loop statement S12 and the do-loop statement S22 of the program in FIG. 4B, s is a simple variable, and a value is assigned each time calculation is performed, so it is necessary to communicate each time it is executed. . Such a variable is called a single value (mono), and is generally used by copying it on the node computer, and the single value data exists in the copy on all connected node computers. ing. Therefore, when calculating by referring to a single value, it is possible to refer to the copy, so it is not necessary to communicate, but when a new value is assigned to a single value, it must be copied to all node computers. I have to.

The data communication B resulting from substituting a new value for this single value generally has the following features. B1. There is relatively little data to communicate. B2. This is wide area communication in which all connected node computers participate synchronously.

In the case of a simple program as shown in FIG. 4, data communication A and data communication B are do loop statements S11 and S21, which are programs for multivalued reference, and programs newly assigned to single values. Since the certain do loop statements S12 and S22 are separate loops, no problem occurs. However, in a generally executed program, this data communication A and data communication B are present in the same loop, and so on. This often occurs, and since communication with completely different properties is mixed on one node connection network,
It was causing a drop in communication efficiency.

An object of the present invention is to provide a distributed memory computer having a synchronous communication network with excellent communication efficiency.

[0013]

As shown in FIG. 1, a distributed memory computer having a synchronous communication network according to the present invention includes a processor, a plurality of node computers having communication means and a local memory used for communication, and In a distributed memory computer having a node connection network for connecting these and exchanging data with each other, each has a synchronous memory for variables that need to be synchronized among all processors, and a synchronous communication means 112, Further, each synchronous communication means is connected by a network different from the node connection network.

As the connection topology of the node connection network, as described above, there is relatively little data to be communicated, and all connected node computers are in wide area communication in a synchronous manner. It is preferable to use.

As described above, the connection topology of the synchronous communication network is one for communicating blocks including a plurality of data such as array elements, and basically, one-to-one communication between nodes. It is preferable to use a tree because it is a local communication.

In the present invention, the synchronous memory preferably includes a data memory and a time stamp memory indicating the last rewriting time, and further preferably includes a data memory and a memory for storing the contents of synchronization processing for the data.

[0017]

In the present invention, since two networks are used, it is possible to simultaneously perform local communication and wide area communication having different properties.

[0018]

【Example】

<< Node Connection Network >> FIG. 6 shows a node connection network used in the embodiment of the present invention. This network includes node computers N1, N2, N3, N4, the same number of nodes C1, C2, C3, C4 as the node computers, communication paths L1, L2, L3, L4 between the nodes, and communication between the nodes and the computers. It is composed of paths NL1, NL2, NL3, NL4 and is mesh-connected. In this configuration, the node computers directly connected to the communication nodes, such as C1 and C2 or C1 and C4, directly communicate with each other and communicate with each other, for example, C1 and C4 or C2 and C3. Node computers that are not directly connected to each other communicate with each other by the intervening communication node passing data to another communication node. The characteristics of this network are as follows. Communication between adjacent node computers is the fastest. When a node computer is connected to a square mesh with N sides, the longest communication distance (time) is about 2 (N-1) times the communication distance between two node computers.

<< Synchronous Communication Network >> FIG. 7 shows a synchronous communication network used in the embodiment of the present invention. This network is composed of two tree networks to which roots are connected, the input side tree is composed of synchronous communication merge nodes M1, M2, M3 and M4, and the output side tree is a synchronous communication fork node F1. , F2
It is composed of F3 and F4. The four node computers have respective merge nodes and input communication paths I1, I
2, I3, I4, and each fork node,
The output communication paths O1, O2, O3, and O4 are connected to perform synchronous communication in which all node computers participate. For example, when all node computers configured by this network add the data of each node computer, (1) Each node computer wants to add data to the merge node connected via the synchronous memory. Is output. (2) In the synchronous communication network, the merge node M1 adds the data of the node computers N1 and N2, and at the same time, the merge node M2 adds the data of the node computers N3 and N4. (3) Next, the merge node M3 adds the results of the merge node M1 and the merge node M2. (4) Copy the result to the fork node F1. (5) Fork node F1 to fork nodes F2, F3
Send data to. (6) The fork node F2 sends data to the node computers N1 and N2, and the fork node F3 sends data to the node computers N3 and N4. As a result, after a certain time, each node computer
All addition results can be obtained.

(Merge Node) FIG. 8 shows the structure of the merge node used in the present invention. This sync merge node
Queues Q1, Q connected to two inputs MI1, MI2
2, a comparator COMP for comparing data types, and a calculator P for calculating data, and the calculation result is MO.
Is output from. As shown in FIG. 9, the synchronous communication data used in the present invention has a time stamp TM indicating the rewriting time of the data, an operation instruction OP which is the content of the processing performed by all the node computers, and each node computer. It is composed of the data DATA of the. The operation of the merge node used in the present invention will be described below. First, the data output from each node computer is input to the queue Q1 or Q2 in the merge node, and the comparator COMP waits for the time stamp TM of Q1 and Q2 and the operation instruction OP to match, and then to match. Then, the arithmetic instruction is sent to the arithmetic unit P, the arithmetic unit P performs arithmetic according to the instruction, and the result is output from the MO.

(Fork Node) FIG. 10 shows the structure of a fork node used in the present invention. In this fork node, the copy device D in the node copies the input data and outputs it from two outputs FO1 and FO2.

(Characteristics) Such a tree-type synchronous communication network has the following characteristics. The communication is synchronized, and all node computers require the same time for communication. In the case of a node computer having N square meshes on one side, the maximum communication time of this synchronous communication network is 2
It is about 2 * LOG ₂ N times the communication between two node computers.

[0023]

As described above, according to the present invention, when referring to multivalued data, the communication between adjacent node computers can be the fastest. Even if a node computer is connected to a square mesh with N sides, the maximum communication distance (time) is at most 2
It is about 2 (N-1) times the communication distance between two node computers, and when a new value is assigned to a single value, the communication is synchronized and all node computers have the same communication. It takes time. In the case of a node computer having N square meshes on one side, the maximum communication time of this synchronous communication network is 2
It can be satisfied that the communication between two node computers is about 2 * LOG ₂ N times at the same time. From this fact, (1) For multiple block data, such as array elements, which requires a relatively large packet Can be designed and optimized. (2) Communication between adjacent node computers is fast,
It is possible to efficiently perform local communication. (3) All connected node computers can participate in a wide area communication in which they participate in a synchronous manner, and overall communication and reduction can be executed at high speed. (4) It is suitable for communication with relatively small amount of data to be communicated and can be optimized for small communication packets. It is possible to provide a distributed memory computer having a synchronous communication network having the two effects described above.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining the principle of the present invention.

FIG. 2 is a block diagram for explaining a conventional example.

FIG. 3 is an example of a program for explaining the problems of the present invention.

4A and 4B are examples of programs for explaining the problems of the present invention.

FIG. 5 is a table for explaining the problems of the present invention.

FIG. 6 is a schematic configuration diagram showing a part of an embodiment of the present invention.

FIG. 7 is a schematic diagram of another part of the configuration showing an embodiment of the present invention.

8 is a block diagram showing the configuration of the synchronous communication merge node of FIG. 7. FIG.

FIG. 9 is a table showing a structure of data used in the synchronous communication of the present invention.

10 is a block diagram showing a configuration of the synchronous communication fork node of FIG. 7. FIG.

Claims (5)

[Claims] 1. A distributed memory computer having a processor, a communication means, and a plurality of node computers having a local memory used for communication, and a node connection network for connecting these and exchanging data with each other. A synchronization memory for variables that need to be synchronized with each other and a synchronization communication means 112, and each synchronization communication means is connected by a network different from the node connection network. A distributed memory computer having a synchronous communication network. 2. A distributed memory computer having a synchronous communication network according to claim 1, wherein a mesh is used as a connection topology of the node connection network. 3. A distributed memory computer having a synchronous communication network according to claim 1, wherein a tree is used as a connection topology of the synchronous communication network. 4. A distributed memory computer having a synchronous communication network according to claim 1, wherein the synchronous memory includes a data memory and a time stamp memory indicating a last rewriting time. 5. A distributed memory computer having a synchronous communication network according to claim 1, wherein the synchronous memory includes a data memory and a memory for storing contents of synchronous processing for the data. .