JP2003203061A - Parallel computer system, and program for selecting calculation node in parallel computer system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a parallel computer system capable of improving the reliability of a system by excluding a calculation node being in a situation that the abnormal termination may be generated during the calculation, from an object in advance. <P>SOLUTION: In this parallel computer system 10 comprising a master computer 20 and a plurality of calculation nodes 30, the calculation node selected from the plurality of calculation nodes refers a shared file 25 of the master computer to process a program in parallel, and the selection of the calculation node is performed on the basis of whether the calculation node can refer the shared file of the master computer or not. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a parallel computer system and a program for selecting a calculation node in a parallel computer system, and more particularly to a parallel computer system for scientific and technological calculation and a program for selecting a calculation node in a parallel computer system.

[0002]

2. Description of the Related Art A science and technology calculation program (analysis program), which has a huge amount of calculation, may be processed in parallel in a parallel computer system.

In a parallel computer system, when a master computer distributes a job to a calculation node (slave computer), it is necessary to confirm whether or not the calculation node can be calculated. This is to improve the reliability of the system by excluding in advance the calculation nodes that may cause “abnormal termination” during the calculation.

In the current method of scheduling a computing node, the validity of the computing node is confirmed by the response of the computing node. Therefore, when a failure occurs in the computing node, it may be impossible to execute the analysis program even if the response is good depending on the kind of the failure.

Here, the reason why the analysis program cannot be executed is that the calculation node cannot read the parameters of the file in the shared file of the master computer, or the memory capacity is large to execute the job. This is because there is a shortage.

In that case, the calculation is redone, but since the analysis program generally takes time, re-execution causes a great loss of time. Further, since the analysis program is operated under the schedule management program, it may not be possible to operate with a commercially available analysis program or the like.

[0007]

In a parallel computer system, it is desired to improve the reliability of the system by excluding in advance the calculation nodes that are in a situation that causes abnormal termination during calculation. .

The following parallel computer system is described in Japanese Patent Laid-Open No. 2001-22601. To secure the number of nodes required by the first job waiting for execution in a parallel computer system that is composed of multiple nodes that are computers and a memory that is shared between the multiple nodes and that is connected to each other via a network , Means for registering as a target job a job group in which the total number of nodes secured by the currently executing jobs among the plurality of currently executing jobs satisfies the number of requested nodes of the first job, The total number of nodes is compared with a value obtained by multiplying the number of required nodes of the first job by a weight predetermined by the system depending on the job operating status, and the system is preset to the number of required nodes of the first job depending on the job operating status. If the value multiplied by the weight is larger than the total number of nodes of the target job, the target job is temporarily stopped or temporarily suspended. It is characterized in that a means for preferentially schedule the beginning of the job.

An object of the present invention is to provide a parallel computer system for excluding beforehand from a target a calculation node in a situation that causes an abnormal termination during calculation, and a program for selecting a calculation node in the parallel computer system. .

[0010]

[Means for Solving the Problem] [Means for Solving the Problem] will be described below by using the numbers and symbols used in the [Embodiment of the Invention]. These numbers and signs are
Although added to clarify the correspondence between the description in [Claims] and the description in [Embodiment of the Invention], the technology of the invention described in [Claims] It should not be used to interpret the scope.

The parallel computer system (10) of the present invention is
Master computer (20) and multiple computing nodes (3
0) and a parallel computer system (10) comprising:
The selected computing node (30) of the plurality of computing nodes (30) is the master computer (2).
0) refers to the shared file (25), the program is processed in parallel, and the calculation node (30) selects the shared file (25) of the master computer (20). ) Can be referred to or not.

[0012] In the parallel computer system (10) of the present invention, the selection of the computing node (30) is based on whether or not mounting of a part of the master computer (20) by the computing node (30) is normal. Is done by or.

In the parallel computer system (10) of the present invention, the master computer (20) selects the computing node (30).

The parallel computer system (10) of the present invention is
Master computer (20) and multiple computing nodes (3
0) and a parallel computer system (10) comprising:
The selected calculation node (30) of the plurality of calculation nodes (30) processes a program in parallel, and the selection of the calculation node (30) is performed by the memory (31) of the calculation node (30). ) Free space.

The parallel computer system (10) of the present invention is
Master computer (20) and multiple computing nodes (3
0) and a parallel computer system (10) comprising:
The selected computing node (30) of the plurality of computing nodes (30) is the master computer (2).
0) refers to the shared file (25), the program is processed in parallel, and the calculation node (30) selects the shared file (25) of the master computer (20). ) Can be referred to and the free capacity of the memory (31) of the calculation node (30).

In the parallel computer system (10) of the present invention, the master computer (20) selects the computing node (30).

In the parallel computer system (10) of the present invention, the selection of the computing node (30) is further performed by the load factor of the CPU (32) of the computing node (30).

In the parallel computer system (10) of the present invention, the selection of the calculation node (30) further includes the load factor of the CPU (32) and the CPU (32) of the calculation node (30). It is performed based on the calculation ability.

A program for selecting a calculation node in the parallel computer system (10) of the present invention is for selecting a calculation node in a parallel computer system (10) including a master computer (20) and a plurality of calculation nodes (30). A program comprising: (a) the calculation node (3
0) confirms whether or not the shared file (25) of the master computer (20) can be referred to, and (b) according to the confirmation result of (a), among the plurality of calculation nodes (30). From the shared file (25) of the master computer (20) to select the computing node (30) that processes the program in parallel.

In the program for selecting a computing node in the parallel computer system (10) of the present invention, in (a), a part of mounting of the master computer (20) by the computing node (30) is normal. Or not.

In the program for selecting a computing node in the parallel computer system (10) of the present invention, the above (a) and (b) are performed by the master computer (20).

A program for selecting a calculation node in the parallel computer system (10) of the present invention is for selecting a calculation node in a parallel computer system (10) including a master computer (20) and a plurality of calculation nodes (30). (C) the calculation node (3
0) the step of confirming the free capacity of the memory (31) of the calculation node (30), and (d) the parallelization of the program from the plurality of calculation nodes (30) according to the confirmation result of (c). The computing node (30) to process
Is a program for causing a computer to execute the step of selecting.

The program for selecting a computing node in the parallel computer system (10) of the present invention is for selecting a computing node in a parallel computer system (10) including a master computer (20) and a plurality of computing nodes (30). A program comprising: (e) the calculation node (3
0) whether or not the shared file (25) of the master computer (20) can be referred to, and (f) the memory (31) of the computing node (30) of the computing node (30). From the plurality of calculation nodes (30), the step of confirming the free capacity and (g) the confirmation results of the above (e) and (f),
The shared file (2) of the master computer (20)
5) The step of selecting the calculation node (30) that processes the program in parallel with reference to 5) is a program for causing a computer to execute.

A program for selecting a calculation node in the parallel computer system (10) of the present invention, further comprising:
(H) a step of confirming a load factor of the CPU (32) of the calculation node (30),
The confirmation result of the above (h) is used for the selection of 0).

A program for selecting a calculation node in the parallel computer system (10) of the present invention, further comprising:
(I) A step of confirming the load factor of the CPU (32) of the calculation node (30) and the calculation capacity of the CPU (32) is included, and the selection of the calculation node (30) includes confirmation of the above (i). The result is used.

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of a parallel computer system of the present invention will be described with reference to the accompanying drawings.

FIG. 1 is a block diagram for schematically showing the parallel computer system of this embodiment. FIG. 3 is a flowchart showing an execution procedure of parallel calculation by the parallel computer system of this embodiment. FIG. 4 is a detailed flowchart of the “computing node selection step” in the flowchart of FIG.

First, with reference to FIG. 3, a procedure for executing parallel calculation by the parallel computer system of this embodiment will be described.

As shown in step S1, the master computer 20 divides the analysis calculation program to be executed into a plurality of (N) jobs as preprocessing. As many calculation nodes 30 as the number of divided jobs are required.

Next, as shown in step S2, the master computer 20 selects N calculation nodes 30. In this selection step, the master computer 20 needs to confirm whether the calculation node 30 is in a state capable of calculating the job. This is because the calculation nodes 30 that may cause “abnormal termination” during the calculation of the job are excluded in advance. The contents of step S2 will be described later.

Next, as shown in step S3, the master computer 20 sends a job execution command to the calculation node 30 selected in step S2. Note that, here, only the execution command is transmitted to the calculation node 30.
Is for reading the program body from the shared file 25 of the master computer 20.

Next, as shown in step S4, the computing node 30 executes the received job. The received program includes a "file read" command, and the "shared file 25 of the master computer 20" is described as the read destination. The calculation node 30 executes the command and executes the job (analysis calculation) while referring to the parameters in the shared file 25 of the master computer 20.

Next, the step of selecting the calculation node 30 (step S2) will be described with reference to FIGS.

The parallel computer system 10 that executes an analytical calculation program for science and technology generally needs a shared disk space 25 for reading parameters and writing calculation results. This is performed by mounting a part of the master computer 20 on the calculation node 30 as described later (see the chain double-dashed line M in FIG. 1).

In the present embodiment, first, the master computer 20 confirms whether or not the mount is normally performed, thereby determining whether or not the computing node 30 can execute the job of the analysis program. As a guideline (first step S11 in FIG. 4).

The master computer 20 confirms that the calculation node 30 can read the parameters of the file in the shared file 25 of the master computer 20, and then sends the job of the analysis program to the calculation node 30. This avoids many situations where the computing node 30 cannot execute the job of the analysis program.

The calculation node 30 is the master computer 2
In order for the master computer 20 to determine whether or not the shared file 25 of 0 can be read, the master computer 20 determines whether or not the file 25 of the master computer 20 is recognized as a part of the computing node 30. Check.

That is, the master computer 20 confirms whether or not the shared disk space 25 of the master computer 20 is mounted on the computing node 30 (steps S11 and S12).

Here, "mounted on the calculation node 30" means that the actual data (parameters of the analysis program, etc.) is a shared file 25 accessible to the calculation node 30 in the HDD 22 of the master computer 20.
Exists (stored) in the calculation node 30 and does not exist (is not stored) in the calculation node 30, the calculation node 30 does not have the shared file 25 accessible to the calculation node 30 in the HDD 22 of the master computer 20. It means being aware.

The master computer 20 determines what file system the computing node 30 has (what kind of file device the computing node 30 has).
With a command to inquire of the computing node 30.

As a result of the master computer 20 executing the command, if it is confirmed that the directory of the shared file 25 of the master computer 20 exists (is mounted) in the compute node 30, the compute node 30 becomes the master computer. The master computer 20 determines that the analysis program using the parameters in the shared file 25 of 20 can be executed (step S12-Y). The confirmation is made by the master computer 2 when the master computer 20 executes the command.
The shared file 25 directory of 0 is the calculation node 30
It can be done by returning as a response from.

As shown in FIG. 2, a "network" file exists in one of a plurality of file devices existing in the computing node 30, and the shared disk 25 of the master computer 20 can be seen in a part of the "network". If it is connected (connected), the mount of the compute node 30 is normal.

The master computer 20 executes step S
As a result of the determination of 12, when it is determined that the mount of the calculation node 30 is not normal, the calculation node 30 is excluded from the selection target of the selection step S2 of FIG. 3 (step S
12-N).

If the computing node 30 is normal in the parallel computer system 10, the shared file 25 in the master computer 20 is automatically mounted when the computing node 30 is started. .

Shared file (directory) in the hard disk (HDD) 22 of the master computer 20
If it is confirmed that 25 is mounted on the computing node 30, the computing node 30 and the master computer 20 in the network of the parallel computer system 10
It can also be confirmed that the communication with is normal. This is because the mounting is performed on the network of the parallel computer system 10.

Next, the master computer 20 inquires about the free memory capacity of the calculation node 30, and based on the free memory capacity obtained as a result of the inquiry, the calculation node 30 can execute the job. It is determined whether or not (second step S14 in FIG. 4).

Some analysis calculation programs require a large amount of memory. The master computer 20
The computing node 30 is inquired about the free space of the memory (reference numeral 31 in FIG. 2), and the free space of the memory 31
It is determined whether the calculation node 30 can execute the job (steps S14 and S15).

The master computer 20 determines whether or not there is no problem (is sufficient) in the free space of the memory 31 found by inquiry, according to the load or the amount of calculation of the job to be executed by the calculation node 30. (Step S15).

The master computer 20 executes step S
As a result of the determination of 15, when it is determined that the mount of the calculation node 30 is not normal, the calculation node 30 is excluded from the selection target of the selection step S2 of FIG. 3 (step S
15-N).

As a result, the master computer 20
As a whole, the parallel computer system 10 can eliminate the obstacle that makes the program inexecutable after the program is started.

It should be noted that in the above embodiment, "confirmation of whether mounting is normally performed on the computing node 30".
(First step S11, S12) and "Query free memory capacity" (second step S14, S15)
You only have to do one or the other. With only one of them, it is possible to obtain the effect of eliminating the computing node 30 that cannot execute the job. Needless to say, if both of them are performed, the above effect is great.

Next, "confirm whether mount is normally performed on the computing node 30" (first step S11,
S12) and "Inquiry about free memory capacity" (second
As a result of performing at least one of steps S14 and S15), when there are a plurality of calculation nodes 30 satisfying the condition (or those conditions), the job is executed by the following method (third steps S16 and S17). The calculation node 30 to be shared is determined.

The master computer 20 inquires the OS of each computing node 30 about the load factor (percentage of idle state) of each CPU (reference numeral 32 in FIG. 2) of the plurality of computing nodes 30 satisfying the above conditions. Then, the master computer 20 sends C to each computing node 30.
If the load factor of PU32 is calculated | required, it will sort from the one with the low load factor. The master computer 20 can select an appropriate calculation node 30 according to the jobs to be shared based on the result of the sorting.

As a result of the sorting, the master computer 20 sets the N or less calculation nodes 30 from the one having the lowest load factor as the selection target in the selection step S2 of FIG. 3 (step S17-Y). On the other hand, as a result of the sorting, the master computer 20 excludes the calculation nodes 30 that are not within the Nth from the selection target of the selection step S2 of FIG. 3 (step S17-N).

When a certain CPU 32 is executing a program with a large load, the memory 31 to be used generally has a large capacity. Therefore, the above step (second step) of "query for free memory capacity" is performed. Thus, the computing node 30 having the CPU 32 is often certified as being incapable of executing a job. However, although the capacity of the memory 31 to be used is small, the load on the CPU 32 is large, and it may not be suitable for job sharing. Therefore, in the third step, it is effective to determine the load of the job by obtaining the “CPU load”.

Further, the master computer 20 does not inquire the load of the CPU 32 of the calculation node 30 with respect to the specific scheduling software that controls the execution of the analysis program in the calculation node 30, but rather with respect to the calculation node 30. It takes the form of inquiring directly to the OS. Therefore, even if a plurality of programs are being executed by a plurality of application software in the calculation node 30, the load on the CPU 32 can be accurately grasped.

Here, in the plurality of calculation nodes 30,
In order to deal with the case where the CPUs 32 having different abilities are mixed, in step S16, a value obtained by multiplying the load factor of each CPU 32 by the spec value (clock frequency etc.) of the CPU 32 is obtained, and the values are sorted. Appropriate evaluation of ability can be realized.

According to the parallel computer system of this embodiment, since it is not executed under the control of a special program, it can be applied to a wide range of programs.

[0059]

According to the parallel computer system of the present invention,
The reliability of the system can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram schematically showing an embodiment of a parallel computer system of the present invention.

FIG. 2 is a block diagram showing a schematic configuration of a computing node in one embodiment of the parallel computer system of the present invention.

FIG. 3 is a flowchart showing an execution procedure of parallel calculation in one embodiment of the parallel computer system of the present invention.

4 is a flowchart showing in detail a "computing node selection step" in the flowchart of FIG.

[Explanation of symbols]

10 Parallel computer system 20 master computer 22 hard disk 25 Master computer shared files 30 compute nodes 31 memory 32 CPU

Claims (15)

[Claims] 1. A parallel computer system comprising a master computer and a plurality of computing nodes, wherein the selected computing node among the plurality of computing nodes refers to a shared file of the master computer, A parallel computer system that processes programs in parallel, and the selection of the calculation node is performed depending on whether or not the calculation node can refer to the shared file of the master computer. 2. The parallel computer system according to claim 1, wherein the selection of the calculation node is performed depending on whether or not mounting of a part of the master computer by the calculation node is normal. 3. The parallel computer system according to claim 1, wherein the selection of the calculation node is performed by the master computer. 4. A parallel computer system comprising a master computer and a plurality of calculation nodes, wherein the selected calculation node of the plurality of calculation nodes processes a program in parallel, The parallel computer system is selected according to the free memory capacity of the computing node. 5. A parallel computer system comprising a master computer and a plurality of computing nodes, wherein the selected computing node among the plurality of computing nodes refers to a shared file of the master computer, A parallel computer system in which a program is processed in parallel, and the selection of the calculation node is performed depending on whether or not the calculation node can refer to the shared file of the master computer and a free space of a memory of the calculation node. 6. The parallel computer system according to claim 5, wherein the selection of the calculation node is performed by the master computer. 7. The parallel computer system according to claim 1, wherein the selection of the calculation node is further performed by a load factor of a CPU of the calculation node. 8. The parallel computer system according to claim 1, wherein the selection of the calculation node further includes a load factor of the CPU of the calculation node and a calculation capacity of the CPU. A parallel computer system based on. 9. A program for selecting a computing node in a parallel computer system including a master computer and a plurality of computing nodes, comprising: (a) whether the computing node can refer to the shared file of the master computer. And (b) from the plurality of computation nodes according to the confirmation result of (b) above,
A program for selecting a calculation node in a parallel computer system for causing a computer to execute the step of selecting the calculation node that processes a program in parallel by referring to a shared file of the master computer. 10. The program for selecting a calculation node in a parallel computer system according to claim 9, wherein the (a) is performed depending on whether or not mounting of a part of the master computer by the calculation node is normal. A program for selecting computing nodes in a parallel computer system. 11. The program for selecting a computing node in a parallel computer system according to claim 9 or 10, wherein (a) and (b) are for selecting a computing node in the parallel computer system performed by the master computer. program. 12. A program for selecting a computing node in a parallel computer system comprising a master computer and a plurality of computing nodes, comprising the step of: (c) checking the free capacity of the memory of the computing node of the computing node. , (D) selecting a calculation node that performs parallel processing of a program from among the plurality of calculation nodes according to the confirmation result of (c), and selecting a calculation node in a parallel computer system For programs. 13. A program for selecting a computing node in a parallel computer system comprising a master computer and a plurality of computing nodes, comprising: (e) whether the computing node can refer to the shared file of the master computer. And (f) checking the free space of the memory of the computing node of the computing node, and (g) confirming the result of the (e) and (f) of the plurality of computing nodes. A program for selecting a computing node in a parallel computer system for causing a computer to execute the step of selecting the computing node that processes a program in parallel by referring to a shared file of the master computer. 14. A program for selecting a computing node in the parallel computer system according to claim 9, further comprising (h) a step of confirming a CPU load factor of the computing node. A program for selecting a calculation node in a parallel computer system, wherein the confirmation result of (h) is used to select the calculation node. 15. A program for selecting a calculation node in the parallel computer system according to claim 9, further comprising (i) a CPU load factor of the calculation node and a calculation of the CPU. A program for selecting a calculation node in a parallel computer system, comprising a step of confirming an ability, wherein the selection result of (i) is used for the selection of the calculation node.