JPH0944462A - Data transfer processor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve an effective transfer data ratio and to raise the transfer performance (throughput) of inter-cluster network by synthesizing plural transfer data of the same destination and making the data into one inter-cluster transfer data. SOLUTION: When the destination cluster of the first transfer data transmitted from the plural arithmetic processors 110 and 111 within a cluster 1 via the network 130 within a raster is the same plural pieces of the first data are synthesized in the transfer width direction or the transfer length direction of the transfer form of an inter-cluster network 3, the second transfer data to which a synthetic indentification data is added is generated in a synthetic part 143 and the data is transferred to the inter-cluster network 3. The second transfer data is decomposed in the transfer width direction or the transfer length direction of the transfer form of the inter-cluster network 3 by the synthetic identification data in a decomposition part 147 and the data is transferred to destination arithmetic processors 210 and 211.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data transfer processing device of an information processing system in which a plurality of clusters are connected by an inter-cluster network, and particularly, transfer data is arranged in a cluster and mediates an intra-cluster network and an inter-cluster net. The present invention relates to a data transfer processing device that synthesizes and disassembles data.

[0002]

2. Description of the Related Art The first prior art is generally shown in FIG.
The configuration shown in FIG.

This information processing system includes a plurality of arithmetic processing devices, a shared memory shared by the arithmetic processing devices, a plurality of arithmetic processing devices, and a plurality of intra-cluster networks connected to the shared memory and data transfer processing devices. An example of a case where data is transferred from the arithmetic processing units 510 to 511 in the cluster 5 configured by one cluster and the inter-cluster network connected to the plurality of clusters to the arithmetic processing units 610 to 611 in the cluster 6 The operation will be described below.

Processors 510 to 51 in the cluster 5
1 is a first data receiving unit 54 in the data transfer processing device 540 for transmitting data via the intra-cluster network 530.
Send to 1.

The first data receiving section 541 of the data transfer processing device 540 sends the received data to the first data converting section 542. The first data conversion unit 542 converts the data received from the first data reception unit 541 into the transfer data format of the intercluster network 7 and sends it to the second data transmission unit 544.

The second data transmitter 544 sends the data received from the first data converter to the inter-cluster network 7. The inter-cluster network 7 sends the transfer data to the second data receiving unit 545 in the cluster 6.

Data transfer processor 640 in cluster 6
The second data receiving unit 545 of the second data converting unit 54 transfers the transfer data received from the inter-cluster network 7.
Send to 6. The second data conversion unit 546 converts the transfer data format of the inter-cluster network into the first data format and sends it to the first data transmission unit 548. The first data transmission unit 548 transfers the data to the arithmetic processing devices 610 to 611 via the intra-cluster network 620.

When the transfer between the arithmetic processing devices 510 and 511 competes with each other, arbitration is performed in the intra-cluster network 530 to transfer the data to the data transfer processing device 540 in order. The data transfer processing device 540 also sequentially transfers the data sent from the intra-cluster network 7 to the inter-cluster network 7.

The second conventional technique is disclosed in Japanese Patent Laid-Open No.
9-165172 publication ”. This has a plurality of processors and a memory connected to each processor, and each processor is connected by a serial data line to enable mutual data transfer.

Each processor accesses a memory connected to another processor through this serial data line.

The shared data on the memory is coded for each type, and each type of data is divided into a plurality of units as needed, and each unit is further divided into its subordinate elements.

Each processor sends the type (code), unit number, and element number of the shared data to the other processor, and the other processor that receives these information sends the corresponding data to the shared data. You are accessing.

[0013]

In the above-mentioned first prior art information processing system, in the data transfer width from the arithmetic processing device to the data transfer processing device and the data transfer width from the shared memory to the data transfer processing device, The latter is set larger.

Therefore, when data is transferred from a plurality of arithmetic processing units to the same cluster, the data transfer processing unit receives the transfer data of the arithmetic processing units sent in order from the intra-cluster network and the inter-cluster network. Only a part of the data portion of the transfer data format is used, and the transfer data of the arithmetic processing unit is sent to the inter-cluster network in the order of being sent from the intra-cluster network.

As a result, there is a drawback that the transfer performance of the inter-cluster network deteriorates.

Further, when the inter-cluster transfer is performed by using the block data having a fixed block length, one block data is occupied only by the transfer data of one arithmetic processing unit and only a part of the block data is used. There is a drawback that the transfer performance of the network is degraded.

The above-mentioned second prior art information processing system can be configured when the number of processors is small and shared data is small, but the number of processors is 1
In the case of a large number such as 28 units, it is not realistic to connect the processors to each other.
There is a drawback that the amount of hardware for connection becomes huge. In addition, when sharing a huge amount of data, the element that is the transfer unit is not one word but 64 words to 2 words.
It is easier to handle and transfer efficiently if the transfer unit is about 56 words. When the device is improved in this way, there is a drawback that the inter-group transfer efficiency is reduced when the actual amount of data to be transferred is small.

[0018]

A first data transfer processing device of the present invention includes a plurality of arithmetic processing devices, a shared memory shared by the plurality of arithmetic processing devices, and a data transfer processing device. An information processing system including a plurality of clusters having an intra-cluster network connecting the plurality of arithmetic processing devices, the shared memory, and the data transfer processing device, and an inter-cluster network connecting the plurality of clusters to each other. A data transfer processing device,
(A) Transferring a plurality of first data to an inter-cluster network when the destination clusters of the first transfer data sent from a plurality of arithmetic processing units in the cluster via the intra-cluster network are the same A synthesizing means for synthesizing in the transfer width direction of the format and adding synthetic identifiers to generate synthetic transfer data; and (b) the synthetic transfer data received from the inter-cluster network by the synthetic identifier in the synthetic transfer data. Disassembling means for disassembling in the transfer width direction of the transfer format of the inter-network and transferring the first transfer data to the destination arithmetic processing device.

A second data transfer processing device of the present invention includes a plurality of arithmetic processing devices, a shared memory shared by the plurality of arithmetic processing devices, a data transfer processing device, and the plurality of arithmetic processing devices. The data transfer processing device of an information processing system comprising: a plurality of clusters having an intra-cluster network connecting the shared memory and the data transfer processing device; and an inter-cluster network connecting the plurality of clusters to each other. And (a) if the destination clusters of the first transfer data sent from the plurality of processing devices in the cluster via the intra-network are the same, the plurality of first data are transferred to the inter-cluster network. And (b) the combination means for generating combined transfer data by combining in the transfer length direction of the transfer format and adding a combination identifier. Decomposing means for decomposing the composite transfer data received from the inter-star network in the transfer width direction of the transfer format of the inter-cluster network by the composite identifier in the composite transfer data, and transferring the first transfer data to the destination arithmetic processing unit. , Is provided.

[0020]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention. The information processing system shown in FIG. 1 includes clusters 1 and 2 and an intercluster network 3 that is connected to the clusters 1 and 2 and transfers data.

The cluster 1 includes the arithmetic processing units 110 to 11.
1, a shared memory 120 shared by them, a data transfer processing device 140, and an intra-cluster network 130 that connects them.

The cluster 2 includes the arithmetic processing units 210 to 21.
1, a shared memory 220 shared by them, a data transfer processing device 240, and an intra-cluster network 230 connecting them.

The data transfer processing device 140 also includes a first data receiving unit 141 and a first data receiving unit 141 that receive transfer data and the like from the intra-cluster network 130.
Receiving the transfer data from the same destination detecting unit 142 for detecting whether the destination cluster of the preceding transfer data and the destination cluster of the succeeding transfer data are the same, and receiving the same destination detection signal as the transfer data sent from the same destination detecting unit 142 A data synthesizing unit 143 and a data synthesizing unit 14 for synthesizing the preceding transfer data and the subsequent transfer data in the transfer width direction or the transfer length direction of the transfer data format of the inter-cluster network 3.
A second data transmission unit 144 for transmitting the inter-cluster transfer data sent from the inter-cluster network 3 to the inter-cluster network 3.
A second data receiving unit 145 that receives the inter-cluster transfer data via the inter-cluster network 3 and the inter-cluster transfer data sent from the second data receiving unit 145 to detect whether or not the data is combined. A combined data detection unit 146, a combined data decomposition unit 147 that receives the inter-cluster transfer data and the combined data detection signal sent from the combined data detection unit 146, and decomposes the transferred data before combining in the data combining unit 143 The decomposed transfer data sent from the data decomposer 147 is processed by the arithmetic processing unit 11 via the intra-cluster network 130.
0-111 and the 1st data transmission part 148 sent to the shared memory 120.

The data transfer processing device 240 has the same structure as the data transfer processing device 140.

Next, the arithmetic processing unit 110 is replaced by the arithmetic processing unit 2
The operation of the data processing apparatus will be described by taking the case where the arithmetic processing apparatus 111 transfers data to the arithmetic processing apparatus 211 as an example.

Data of a transfer data format (hereinafter referred to as first data) sent from the arithmetic processing unit to the first data receiving section of the data transfer processing unit via the intra-cluster network and the data transfer data format of the inter-cluster network. 2 (hereinafter referred to as second data) is shown in FIG.

First, the first data is the arithmetic processing unit 110.
And 111 to the intra-cluster network 130. The intra-cluster network 130 sends the first data received from the plurality of arithmetic processing devices to the data transfer processing device 140 according to a predetermined rule. Here, it is assumed that the first data of the arithmetic processing devices 110 and 111 are sequentially supplied to the data transfer processing device 140.

In the data transfer processing device 140, the first data receiving section 141 receives the first data of the arithmetic processing devices 110 and 111 in order. First data receiving unit 141
First sends the first data from the arithmetic processing unit 110 and then the first data from the arithmetic processing unit 111 to the same destination detecting unit 142.

The same destination detecting unit 142 monitors the destination cluster field of the first data and turns on the same destination signal when the preceding first data and the following first data are the same destination. .

The first data of the arithmetic processing unit 110, the first data of the arithmetic processing unit 111 and the same destination signal (“on”) are sent to the data synthesizing unit 143 in order.

The data synthesizing unit 143 is the same destination detecting unit 1
42, the data 1 field of the second data is the data field of the first data of the arithmetic processing unit 110, and the data 2 field is the first data of the arithmetic processing unit 111. Synthesize the data fields of.

The destination cluster and the destination ID field are also combined. In addition, the composite identifier is added to the composite identification field of the second data.

Then, the second data is sent to the second data transmission section.

The second data transmitter 144 sends the second data received from the data synthesizer 143 to the inter-cluster network 3.

The inter-cluster network 3 transfers the second data to the second data receiving unit 145 in the cluster 2.

The second data receiving section 145 receives the second data from the inter-cluster network 3 and sends it to the combined data detecting section 146. The combined data detection unit 146 monitors the combined identifier field in the second data, and if they are combined, sets the combined data detection signal to “ON” and sends the combined data detection unit 147 together with the second data to the combined data decomposing unit 147. send. The combined data decomposing unit 147 receives the combined data detection signal and the second data, and when the detection signal is “ON”,
The second data is "destination cluster field + destination I
The data is decomposed into the first data composed of "D1 field + data 1 field" and the first data composed of "destination cluster field + destination ID2 field + data 2 field" and sequentially sent to the first data transmission unit 148.

The first data transmission unit 148 sequentially transmits the first data transmitted from the combined data decomposition unit 147 to the intra-cluster network 230.

The intra-cluster network 230 transfers the first data sent from the first data transmission unit 148 to the arithmetic processing units 210 and 211 according to the destination ID of the first data.

In the above example, the transfer from the arithmetic processing unit in one cluster to the arithmetic processing unit in another cluster is shown, but the arithmetic processing unit in one cluster shares the data in another cluster. It goes without saying that the present invention can also be applied to transfer from a memory or a shared memory in one cluster to a shared memory in another cluster.

[0041]

As described above, by combining a plurality of first data into second data and decomposing the second data into a plurality of first data, the data utilization rate of the second data can be improved. By enabling the improvement, there is an effect that the performance improvement of the data transfer of the inter-cluster network which the conventional device has can be achieved.

Further, by using the inter-cluster network, there is an effect that the amount of hardware for connection can be reduced as compared with the case where the processors are connected to each other.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a format diagram showing an intra-cluster / inter-cluster transfer data format.

FIG. 3 is a block diagram showing an embodiment of a conventional technique.

[Explanation of symbols]

1, 2, 5, 6 clusters 3, 7 inter-cluster network 110, 111, 210, 211 arithmetic processing unit 510, 511, 610, 611 arithmetic processing unit 120, 220, 520, 620 shared memory 130, 230, 530, 630 Intra-cluster network 140, 240, 540, 640 Data transfer processing device 141, 541 First data receiving unit 142, 242 Same destination detecting unit 143 Data combining unit 144, 544 Second data transmitting unit 145, 545 Second data Receiver 146 Combined data detector 147 Combined data decomposer 148, 548 First data transmitter 542 First data converter 546 Second data converter

Claims (2)

[Claims] 1. A plurality of arithmetic processing devices, a shared memory shared by the plurality of arithmetic processing devices, a data transfer processing device, the plurality of arithmetic processing devices, the shared memory, and the data transfer processing device. In the data transfer processing device of the information processing system, comprising: a plurality of clusters having an intra-cluster network connecting to each other; and an inter-cluster network connecting the plurality of clusters to each other,
(A) Transferring a plurality of first data to an inter-cluster network when the destination clusters of the first transfer data sent from a plurality of arithmetic processing units in the cluster via the intra-cluster network are the same A synthesizing means for synthesizing in the transfer width direction of the format and adding synthetic identifiers to generate synthetic transfer data; and (b) the synthetic transfer data received from the inter-cluster network by the synthetic identifier in the synthetic transfer data. And a disassembling means for disassembling in the transfer width direction of the transfer format of the inter-network and transferring the first transfer data to the destination arithmetic processing device. 2. A plurality of arithmetic processing devices, a shared memory shared by the plurality of arithmetic processing devices, a data transfer processing device, the plurality of arithmetic processing devices, the shared memory, and the data transfer processing device. In the data transfer processing device of the information processing system, comprising: a plurality of clusters having an intra-cluster network connecting to each other; and an inter-cluster network connecting the plurality of clusters to each other,
(A) Transferring a plurality of first data to an inter-cluster network when the destination clusters of the first transfer data sent from a plurality of arithmetic processing units in the cluster via the intra-cluster network are the same A synthesizing unit for synthesizing in the transfer length direction of the format and adding synthetic identifiers to generate synthetic transfer data, and (b) the synthetic transfer data received from the inter-cluster network by the synthetic identifier in the synthetic transfer data. And a disassembling means for disassembling in the transfer width direction of the transfer format of the inter-network and transferring the first transfer data to the destination arithmetic processing device.