JPH0991261A - Parallel computer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the parallel computer which performs multinode simultaneous input/output processing at a high speed. SOLUTION: Calculation nodes 10 to 15 and input/output nodes 30 of the parallel computer are arranged like a matrix, and partial nodes out of plural calculation nodes 10 to 15 are provided with caches 50 to 52 as cache nodes 20 to 22, and input/output nodes 30 are provided with caches 53. In accordance with designation for the start of the parallel computer, one or more cache nodes 20 to 22 which can communicate with some of input/output nodes 30 are set, and one or more calculation nodes 10 to 15 which can communicate with cache nodes 20 to 22 are set. When calculation nodes 10 to 15 read out data from a secondary storage device 40, they transmit a data request message to communicatable cache nodes 20 to 22. Cache nodes 20 to 22 transmit data to calculation nodes 10 to 15 if data exists in caches 50 to 52, but cache nodes 20 to 22 read out data from the secondary storage device 40 to transmit it to calculation nodes 10 to 15 if data doesn't exist in caches 50 to 52.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention has a plurality of calculation nodes each having a local memory and a processor, and a plurality of input / output nodes each having a local memory, a processor and a secondary storage device, which are mutually connected by a network. Then, the present invention relates to a parallel computer in which each node executes a program in parallel.

[0002]

2. Description of the Related Art Parallel computers are roughly classified into those for business and those for science and technology. A parallel computer, which is generally called a business type, has a system configuration in which a secondary storage device such as a disk is connected to all the nodes including a processor, a memory, and a network communication device, since data processing is the main focus. On the other hand, in a parallel computer called scientific and technological computing, since data calculation occupies most of the processing, a calculation node that does not have a secondary storage device and a secondary storage device that only performs data calculation. And a parallel computer is configured by an input / output node that has an input / output node that performs input / output processing in addition to data calculation. The number of input / output nodes is generally smaller than the number of calculation nodes. By the way, when the scientific and technological calculations are executed on a parallel computer, it is often the case that the same program is executed almost simultaneously on all nodes in the parallel computer. Therefore, the same input / output processing occurs at the same time, and the input / output processing is likely to concentrate on one input / output node. When the processing concentrates on one node, the processing waiting time becomes longer as the number of nodes increases, and there is a problem that the input / output node which is the receiving side of the input / output processing request cannot receive the processing request. To do. Particularly in a massively parallel computer having more than 1000 nodes, the concentration of processing on a specific node affects the performance of the entire system.
As a method for avoiding concentration of processing, for example, Japanese Patent Laid-Open No. 3-2
As shown in 09531, there is disclosed a method of dynamically distributing a service process in which processing is concentrated to other nodes. However, the problem of processing concentration in the case where processing can be performed only by the input / output node to which the secondary storage device is connected as in the above problem has not been solved.

[0003]

When a program running on a plurality of computing nodes on a parallel computer simultaneously performs a read process on data on the same input / output node, the data is simultaneously written on the input / output nodes. A communication requesting a read process occurs. The input / output node receives a data read request and performs processing, but one input / output node can process only a request from one node at a time. If processing request communication is input from the N calculation nodes to the same input / output node at the same time, the time required for the input / output processing of each node is N / 2 times the average, and N times until the input / output processing of all nodes is completed. It will take time. In a massively parallel computer having more than 1000 computing nodes, the processing waiting time becomes 1000 times or more, and the influence on the overall computer performance becomes large. There is also a method in which a cache is provided on each computing node and the input / output data is cached here, but the same problem occurs because communication to the input / output node is required when the first input / output data is acquired. appear. A scientific and technical computing program, which is a typical application program of a parallel computer, has a format in which the same program is simultaneously executed on a plurality of computing nodes, so that the above problem is highly likely to occur. Therefore, an object of the present invention is to reduce the problem of the processing waiting time that accompanies the communication of the input / output processing from the programs executed in parallel on the parallel computer and to process the multi-node simultaneous input / output processing at high speed. To provide.

[0004]

In order to achieve the above object, the present invention has a plurality of calculation nodes, a plurality of input / output nodes, and a network for connecting the plurality of calculation nodes and the input / output nodes to each other. , The computing node has a processor, a memory for holding programs and data used by the processor, and a device for communicating with other computing nodes and input / output nodes via a network. A memory for holding a program and data used by a processor; a secondary storage device for holding a program and data used by the program; and a device for communicating with other computing nodes and input / output nodes via a network. In a parallel computer that has a cache node provided as a cache node in some of the plurality of compute nodes According to the designation at the time of starting the parallel computer, one or more cache nodes that can communicate with any of the input / output nodes are set, and one or more calculation nodes that can communicate with each of the cache nodes are set. When data read or write processing requests are simultaneously made to the same secondary storage device from one or more of the calculation nodes, the cache node requests the calculation node instead of the input / output node having the secondary storage device. I am trying to supply the data. Further, the computing node adds information to which data node a request enters the same data to the data request message to be transmitted, and the cache node follows the additional information in the data request message received from the computing node,
An identifier indicating whether or not there is a request for the same data corresponding to the calculation node communicating with the cache node and an identifier indicating whether or not the request is executed are set, and the data is supplied to all the calculation nodes that have requested the data. The data on the cache is held until the values of both identifiers match. Further, the computing node adds information on the next read or information on the next write to the data request message to be transmitted, in addition to the information on the current request,
When the data corresponding to the additional information in the data request message received from the computing node does not exist in the cache, the cache node transmits the data corresponding to the additional information to the secondary storage device via the input / output node. I'm trying to read ahead. Further, it has a plurality of calculation nodes having a communication area, a plurality of input / output nodes having a communication area, and a network interconnecting the plurality of calculation nodes and the input / output nodes, and the plurality of calculation nodes and the input / output nodes. Is the data to be transferred to any of the communication destination nodes, the identifier of the communication area assigned to the transfer destination user process that should use the data, and the data in the communication destination node to which the data to be communicated should be written. A transmission circuit for transmitting a message including information designating a position in the memory via the network, and a transmission circuit transferred from the source node of the plurality of calculation nodes and the input / output node via the network. A parallel circuit that has a receiving circuit that receives messages and can send data from one source node to multiple destination nodes simultaneously In a computer, a part of the plurality of calculation nodes is used as a cache as a communication area of the calculation node to be a cache node, and when the cache data of one of the caches is changed, the changed contents are the same at the same time. It sends it to other cache nodes that have cache data.

[0005]

[Operation] Since the number of the calculation nodes that transmit the I / O processing request is small for each cache node that caches data, the processing waiting time of the calculation node on the transmitting side is reduced, and the performance of the entire parallel computer is improved. To do.

[0006]

EXAMPLE An example of the present invention will be described in detail below.

FIG. 1 shows a configuration of a calculation node and an input / output node in a parallel computer according to the present invention, and FIG. 2 shows an embodiment of a procedure of a data reading process when the present invention is applied to a parallel computer. FIG. 3 is a flow chart, and FIG. 3 is a flow chart showing an embodiment of a processing procedure when data read processing requests are simultaneously issued from a plurality of calculation nodes. In FIG. 1, 10 to 15 are calculation nodes, 20 to 2
Reference numeral 2 is a calculation node having a cache, that is, a cache node, and 30 is an input / output node which has a cache and a secondary storage device and also performs input / output processing in addition to data calculation. Reference numeral 40 is a secondary storage device, 50 to 53 are caches, and 60 to 68 are networks that mutually connect calculation nodes and input / output nodes.

Next, the processing of each unit in FIG. 1 will be described with reference to the flowchart of FIG. Each node of the parallel computer is arranged in a matrix, but when the computer starts up, which compute node becomes the cache node, which cache node can communicate with which cache node, and which cache node is connected to which input / output node. It is specified in advance whether the node can communicate. In the example of FIG. 1, the calculation nodes 10 and 11 are cache nodes 20.
And the calculation nodes 12 and 13 are cache nodes 21
, The computing nodes 14 and 15 are designated to communicate with the cache node 22, and the cache node 2
0, 21 and 22 are designated to communicate with the input / output node 30.

When the calculation node 10 requests the data held in the secondary storage device 40, first, the calculation node 1
0 sends a data read request to the cache node 20 that may cache the data in the secondary storage device 40 (110). The cache node 20 receives the data read request from the computing node 10 (12
0), the cache node 20 checks whether the data requested by the calculation node 10 is stored in the internal cache 50 (130), and if the data is contained, transmits the data to the calculation node 10 (220). If there is no data, the cache node 20 sends a data read request to the input / output node 30 (140). The input / output node 30 receives a data request from the cache node 20 (1
50), it is checked whether the cache 53 in the input / output node 30 contains the requested data (160). If the data is included, the data is transmitted to the cache node 20 (190). 2 if there is no data in cache 53
The data is read from the next storage device 40 (170) and the data is cached (180). Then, the data is transmitted to the cache node 20 (190). The cache node 20 receives the data (200), puts the data in the cache 50 (210), and further stores the data in the compute node 1.
0 (220) and the computing node 10 receives the data. In the embodiment of FIG. 2, only the case where the calculation node 10 requests data is explained, but the calculation node 11
The same applies to the processes from # 15 to # 15.

Next, the processing in the case where processing requests are simultaneously issued from a plurality of calculation nodes will be described with reference to FIG. When the computing nodes 10 and 11 simultaneously request data held in the secondary storage device 40, the computing nodes 10 and 11 send a data request to the cache node 20 (300, 4).
00). When the cache node 20 receives the processing request (500), the processing is started from the previously received data read request. The embodiment of FIG. 3 shows the processing when the request from the calculation node 10 is accepted first. First, the cache node 20 checks whether or not the cache 50 has request data from the calculation node 10 (51).
0). If data exists, the data is calculated on node 1.
0 (520), and processing of the request from the computing node 11 is started (530). If the data does not exist, a data request is sent to the input / output node 30 (511),
After waiting for the response and putting the data in the cache 50 (512), the data is transmitted to the computing node 10 (52).
0). The computing node 10 receives the data (31
0), the processing on the side of the calculation node 10 ends.

Next, the cache node 20 starts processing a data request from the calculation node 11. The cache node 20 checks whether the cache 50 has data requested by the calculation node 11 (530). If the data exists, the data is transmitted to the calculation node 11 (54
0), the process of the cache node 20 ends. If the data does not exist, a data request is transmitted to the input / output node 30 (531), the response is waited for, the data is stored in the cache 50 (532), and then the data is transmitted to the calculation node 11 (540). ). The computing node 11 receives the data (410), and the processing on the computing node 11 side also ends.

In the embodiment of FIG. 3, the calculation nodes 10, 11 are
Has described the processing when the processing requests are issued at the same time. However, there is a difference in the time for issuing the processing requests between the calculation nodes 10 and 11, and for example, the calculation is performed while the cache node 20 is executing the processing for the calculation node 10. The processing is the same even when the processing request from the node 11 is received. Further, the embodiment of FIG. 3 has been described with respect to the calculation nodes 10 and 11 and the cache node 20, but the processing is the same for the calculation nodes 12 to 15 and the cache nodes 21 to 22.

According to this embodiment, a plurality of cache nodes play a role of relieving the concentration of processing requests to the input / output nodes, and by preparing a plurality of cache nodes, the processing requests for each cache node can be reduced. Since communication is reduced, the processing waiting time is shortened, and the performance of the parallel computer as a whole can be improved. For example, in the embodiment of FIG. 1, assuming that the processing time when a data read request is issued to a certain node is 1, if the calculation nodes 10 to 15 directly issue a process to the input / output node 30, the average is tripled, The processing time of 6 times is required until the processing of No. is completed, whereas if there is data in the caches 50 to 52, the processing time is 1.5 times on average and twice as long as the processing of all nodes is completed. become.

In the above embodiment, only one cache node is placed between the calculation node requesting the data and the input / output node, but it goes without saying that two or more cache nodes may be placed. Further, in the above embodiment,
Although the data read request process has been described, the same effect can be obtained in the case of a data write request or a process request for data on the cache. Further, in the above-described embodiment, the example in which the program that issues the data read request is executed only in the computing nodes 10 to 15 that do not have a cache is shown. However, the program that issues the data read request is in the cache nodes 20 to 22, It may be executed in the input / output node 30. In this case, if there is data in the cache, the processing wait time due to communication does not occur and the processing becomes faster.

In the embodiment of FIG. 1, the processing is performed only when the cache is on the calculation node, but as another embodiment, for example, it is shown in Japanese Patent Laid-Open No. 6-19856.
In a parallel computer in which multiple compute nodes (clusters) are connected by a network and the data area of a process can be assigned to the communication area of each compute node, a cache is used as the communication area of some compute nodes. By using the data area to be used, a cache node can be formed, and communication processing of data on the cache and cache management information can be efficiently performed. In this embodiment, since it is not necessary to copy the data in the cache and the cache management information to the communication area, the data copy time can be eliminated and the processing time in the node can be shortened. Further, when the communication function of the network has a function of simultaneously communicating with a plurality of calculation nodes, the efficiency of mutual communication between the calculation nodes can be improved. First, when the computer is started up, which computing node will be the cache node is specified, and at the same time, all the caching nodes are informed which computing node has the cache. Then, when a change occurs in the cache state, it immediately notifies other cache nodes using the simultaneous communication function. in this way,
No matter how many calculation nodes have caches in the parallel computer, since the operation is completed by one transmission process, the efficiency of communication between the calculation nodes can be improved.

As another embodiment, in the embodiment of FIG. 1, the cache in the cache node is provided with a cache access display flag 70 and a cache invalidation permission flag 75 as flags (identifiers) as shown in FIG. When the calculation nodes request the same data, it is possible to add a function of holding the cache data until the processing request from the other calculation nodes for the data stored in the cache is completed. In FIG. 4, the cache access display flag 70 is composed of a pair of an access node display flag 71 and an access completion display flag 72,
Each flag has an indication bit of the number of calculation nodes that may make a data request, and one bit corresponds to one calculation node. Access node display flag 7
1 indicates a calculation node requesting data in the cache data 80, and “1” is set in the flag of the calculation node requesting data. The access completion display flag 72 displays a calculation node that has already made a data request, and “1” is set in the flag of the calculation node that has already made a data request. Further, the cache invalidation permission flag 75 is the cache data 80.
It indicates whether or not the data in it may be invalidated, and "1" is set when the data may be invalidated.

Next, the processing procedure of this system will be described with reference to the flowchart of FIG.

First, some computational nodes requesting data from a cache node, when requesting data, send a data request message to which information on which computational node requests the same data is added (600). . When the cache node receives the first data request communication (605), it reads the information of the calculation node that issues the processing request for the same data from the data request communication message and sets it in the access node display flag 71 (61).
0). That is, "1" is set in the access node display flag corresponding to the calculation node that issues the processing request for the same data.
Stand up. Then, the data is transmitted to the calculation node (6
20), "1" is set to the access completion display flag corresponding to the calculation node that requested the data (630). Similarly, when the same data request communication is received from another computing node (640), the data is transmitted to the computing node (65).
0), "1" is set to the access completion display flag corresponding to the calculation node that requested the data (660). here,
The access node display flag 71 and the access completion display flag 72 are compared (670), and if the flags do not match, that is, if there are computing nodes that have not yet issued a data request, the next data request reception wait ( 68
0). If the flags match, it means that the data request communication has been entered from all the calculation nodes, so that the cache invalidation permission flag 75 is set to "1" (690), and the cache data 80 may be invalidated. Put in a state.

Although the above description has explained the cache access display flag, the cache invalidation permission flag, and the cache data relating to a set of calculation nodes requesting the same data, another set of calculation nodes requesting the same data. It goes without saying that a cache access display flag, a cache invalidation permission flag, and cache data are prepared for each. For example, a cache access display flag and a cache invalidation permission flag may be prepared for each tag of cache data. According to this embodiment, when a request for the same data is input from a plurality of calculation nodes, the data is not removed from the cache until the data is sent to all the calculation nodes that have issued the request. Communication of a data request does not occur in the input / output node, the communication between the cache node and the input / output node can be reduced, and the efficiency of inter-node communication can be improved.

In the above-described embodiment, the data request message is added with the information indicating from which calculation node the request enters the same data. However, as another embodiment, instead of this information, the information of the prefetching instruction of the data is added. It is also possible to add. In this embodiment, when the computing node sends a data request, it sends information about the next requested data together in a data request message. When the cache node receives the data request from the calculation node, if there is no data in the cache, the cache node requests the prefetch data instructed at the same time as the request data from the input / output node to prefetch the data. When there is data in the cache, the request data is transmitted to the calculation node, and when there is no prefetch data in the cache, the prefetch data is requested to the input / output node at the same time to prefetch the data. According to this embodiment, the cache node can read in advance the data required for the next data request from the calculation node from the input / output node, and can reduce the processing wait time due to a cache miss. The number of communications between the cache node and the input / output node can be reduced, and the performance of the parallel computer as a whole can be improved. In the above embodiment, the method of requesting further data at the time of data request has been described. However, the pre-read information of the next data is added at the time of data writing, and the next data is generated at the time of data request. It is also possible to add information related to data writing to be performed and information related to writing of next data when writing data.

[0021]

As described above, according to the present invention, it is possible to avoid the concentration of processing requests from a large number of calculation nodes to one input / output node and to shorten the processing waiting time. As a result, the performance of the entire parallel computer can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing connections of a calculation node, a cache node, and an input / output node in a parallel computer that is an embodiment of the present invention.

FIG. 2 is a flowchart showing a processing procedure in the embodiment of FIG.

FIG. 3 is a flowchart showing a processing procedure when a plurality of computing nodes simultaneously issue input / output processing requests in the embodiment of FIG.

FIG. 4 is a block diagram of a cache in which a cache access display flag and a cache invalidation permission flag are added in the embodiment of FIG.

5 is a flowchart showing a processing procedure in the embodiment of FIG.

[Explanation of symbols]

 10-15 Compute node 20-22 Cache node 30 I / O node 40 Secondary storage device 50-53 Cache 60-68 Inter-node network communication path 70 Cache access display flag 71 Access node display flag 72 Access completion display flag 75 Cache invalidation Permission flag 80 cache data

Claims (4)

[Claims] 1. A plurality of calculation nodes, a plurality of input / output nodes, and a network interconnecting the plurality of calculation nodes and the input / output nodes, wherein the calculation nodes are processors, and programs and data used by the processors. And a device for communicating with other computing nodes and input / output nodes via a network, and the input / output node includes a processor, a memory for holding a program and data used by the processor, and a program A parallel computer having a secondary storage device for holding data used by a program, and a device for communicating with other computing nodes and input / output nodes via a network, wherein a part of the plurality of computing nodes is provided. A cache is provided for each of the above compute nodes as a cache node, and the above-mentioned input / output is performed according to the specification when the parallel computer is started up. One or more of the cache nodes that can communicate with any of the input / output nodes of the node, and one or more of the compute nodes that can communicate with each of the cache nodes. A parallel computer, characterized in that, when a data read or write processing request is made to a storage device, the requested data is supplied from the cache node to the calculation node instead of the input / output node having a secondary storage device. . 2. The parallel computer according to claim 1, wherein the computing node adds information indicating from which computing node a request enters the same data to a data request message to be transmitted, and the cache node receives from the computing node. According to the additional information in the data request message, an identifier indicating whether there is a request for the same data corresponding to the calculation node communicating with the cache node and an identifier indicating whether the request has been executed are set, and the data is requested. A parallel computer, characterized in that the data is supplied to all the calculated nodes and the data in the cache is held until the values of both the identifiers match. 3. The parallel computer according to claim 1, wherein the computing node adds information about a next read or information about a next write to the data request message to be transmitted, in addition to the information about the current request, Prefetches the data corresponding to the additional information from the secondary storage device via the input / output node when the data corresponding to the additional information in the data request message received from the computing node does not exist in the cache. A parallel computer characterized by that. 4. A plurality of computing nodes having a communication area,
A plurality of input / output nodes having a communication area and a network interconnecting the plurality of calculation nodes and the input / output nodes are provided, and the plurality of calculation nodes and the input / output nodes transfer to any of the communication destination nodes. A message including data, an identifier of a communication area assigned to a transfer destination user process that should use the data, and information designating a position in the memory in the communication destination node where the data to be communicated is written A transmission circuit for transmitting the message via the network, and a reception circuit for receiving a message transferred from the transmission source node of the plurality of calculation nodes and the input / output node via the network, In a parallel computer capable of simultaneously transmitting data from a source node to a plurality of destination nodes, a part of the plurality of computing nodes When a cache is used as a communication area of the compute node and a cache is used as the communication area of the compute node, when the cache data of one of the caches is changed, the changed contents are simultaneously transmitted to another cache node having the same cache data. A characteristic parallel computer.