JPH11312149A - Load distribution control system and its device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To keep the balance of loads among computers by adding load information which indicates the load state of the self computer to a transmission message at the time of an inter-node communication processings among the optional computers and exchanging the load information between the computers. SOLUTION: Load information is transmitted from a load monitor demon 35 on the computer 11 to the computer 11a by a certain opportunity and is stored in plural pieces of load information file 34. Then, a scheduler demon 33 on the computer 11a determines by which one of the computers 11 a job or a process 31, which reaches a reception queue 32 is executed based on the pieces of load information, queues the job or the process 31 in the execution waiting queue 37 of the applying computer 11, successively takes-out it by a job execution demon and exchanges the pieces of load information between the respective computers 11. Thus, the balance of the loads among the respective computers 11 is kept.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to execution control of processes and jobs in a parallel computer system, and more particularly to a load distribution control method for a computer.

[0002]

2. Description of the Related Art As an example of a conventional method, Japanese Patent Laid-Open Publication No.
A method in which each computer exchanges load information with another computer via a shared memory as in 160884 to balance loads among the computers.

Regardless of the above, when a scheduler computer schedules a job or a process based on the load information of a computer, the load information is transmitted only from each computer to the scheduler computer.

[0004]

The method of exchanging load information of a computer via a shared memory is applicable only to a system having a shared memory, and therefore has a limited scope.

When a scheduler computer schedules a job or process based on load information, it is difficult to accurately determine the load applied to the computer when the job or process is executed before execution. In some cases, the decision was made, which caused load imbalance between computers.

In the case where the load imbalance occurs, the load information is transmitted from the computer in which the load imbalance occurs to the scheduler computer in order to correct the load imbalance. In addition to the above, it causes local load and overhead of the network to the scheduler computer, and the accuracy of the sent load information is gradually lost by spending time on these processes.
It has been difficult to achieve accurate load imbalance correction using the scheduler computer.

[0007]

When load imbalance occurs between computers, in order to correct the load imbalance, when performing inter-node communication processing between arbitrary computers, the load of the own computer is included in a transmission message. Load information indicating the state is added.

The computer that has received the transmission message and the load information compares the load on its own computer with the load on the message transmission side, and moves the load (unexecuted job or process) as necessary. .

That is, when the load on the computer on the message sending side is high and the load on the computer on the message receiving side is low, load information indicating that the load is high is added to the transmission message, and the unexecuted job of the message sending side computer is added. Alternatively, the process is moved to the message receiving computer.

Conversely, when the load on the computer on the message sending side is low and the load on the computer on the message receiving side is high,
Load information indicating that the load is low is added to the transmitted message,
Move the unexecuted job or process of the message receiving computer to the message transmitting computer.

The level of the load is based on a relative comparison of the load between the two computers. If the difference between the loads is small, the two computers can be regarded as having substantially the same load. Less than a move. Therefore, a difference allowable value k is introduced as a criterion as to whether the difference in load between the two computers can be regarded as minute or not, whereby the message receiving computer has a higher load on its own computer. Is lower, lower, or the same.

When an unexecuted job or process is moved, the fact of the movement and the latest load information of both computers are transmitted to the scheduler computer, and the scheduler computer obtains the load information of the computer, thereby Scheduling based on

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram showing a system configuration of a parallel computer system to which the present invention is applied. Reference numeral 11 denotes a computer, and a plurality of computers 11 are connected to a network 12 to configure a system. Although FIG. 1 shows four computers, the number of computers 11 is arbitrary.

FIG. 2 is a block diagram showing a detailed configuration of the computer 11. As shown in FIG. 2, the computer 11 has a CPU
21, a memory 22, and an external storage device 23, each of which is connected to a bus 24 and exchanges data through the bus 24.

An embodiment will be described with reference to FIG.

The computer to which the processes and jobs requested by the user arrive collectively arrives at the computer 11a.

The load information is sent from the load monitoring daemon 35 of the computer 11 to the computer 11a at a certain timing (for example, at a certain time interval or when there are no more jobs and processes to be executed). 34
Therefore, the scheduler daemon 33 on the computer 11a determines on which computer 11 the job or process 31 arriving at the reception queue 32 is to be executed, based on the information, and determines the execution queue of the computer. 3
By queuing a job or a process 31 at 7, a computer to be executed is assigned.

When a job or process 31 is queued in the execution queue 37 on the computer 11, the job or process 31 is sequentially taken out and executed by the job execution daemon.

(The process and the job 31 to which the computer to be executed is assigned will be referred to as a process and a job 311.) However, the scheduler daemon 33 uses the information of the load monitoring daemon 35 as information for determining when assigning the computer. If time is required from the time when the load is measured to the time when the load is determined, the determination may not be made accurately, and as a result, a job or process may be assigned to an inappropriate computer, thereby balancing the load between the computers. Is not maintained.

The following description will continue on the case where the load imbalance occurs.

When a load imbalance occurs, any two computers can be classified into a computer with a relatively small load and a computer with a relatively large load. However, when the difference between the loads is very small, the meaning of this classification becomes insignificant. Therefore, the loads are quantified, and the upper limit of the difference that can be considered that the loads are almost the same is introduced as the difference allowable value k. k is defined in units of load information to be compared, and is related to how much difference is allowed at the time of comparison. Therefore, k is determined by a system administrator or the like according to operation.

Here, when an inter-node communication process is performed between any two computers in response to an execution request from the AP or the like, the message transmitting computer includes load information indicating the load status of its own computer in a message to be transmitted. Add.

The message receiving computer refers to the added load information and compares it with its own load status. The following cases can be considered as comparison results.

Case A: A case where the load is smaller on the message transmitting computer than on the receiving computer (the difference in load exceeds the allowable difference value k).

Case B: A case in which the load on the message transmitting computer is larger than that on the receiving computer (the difference in load exceeds the allowable difference value k).

Case C: The load difference between the message sending computer and the receiving computer is very small. (If the load difference is less than or equal to the difference allowable value k) In case C,
Strictly speaking, although a load imbalance occurs, the difference is within the reference value k, and the load between the two computers is regarded as substantially the same.

In cases A and B, it is necessary to move the load from a computer with a large load to a computer with a small load in order to correct the imbalance of the load. The movement of the load is realized by moving the job or process 311 on the execution queue 37 of the computer with a large load onto the execution queue 37 of the computer with a small load as follows.

In the following description, the message transmitting computer is referred to as a computer S, and the message receiving computer is referred to as a computer R.

(I) Case A When a message is issued from the computer S, the processing is performed in the following procedure. (FIG. 7,
(See case A in FIG. 8) (a) The computer R compares the load between its own computer and the computer S, and if the difference exceeds k, it is necessary to move the load, and the fact is reported. I do. In this case, the computer S is notified that the unexecuted job or process is to be moved. If the difference is within k, as in case C, no load shift occurs.

(B) The computer S determines whether an unexecuted job or process can be executed at that time, and responds to the computer R with the determination result.

If the result of the determination in (c) and (b) is that the computer R can be moved, the computer R transmits the unexecuted job or process 3
11 is separated from the execution queue 37 of its own computer, and the job or process execution request received from the scheduler daemon 33 is sent to the computer S. The computer S executes the execution queue of its own computer based on the received execution request An unexecuted job or process is queued at 37.

(D) Both the computer S and the computer R report the load on their own computers to the scheduler daemon 33 in order to notify that the load on their own computers has changed.

(Ii) Case B When a message is issued from the computer S, processing is performed in the following procedure. (FIG. 7,
(See case B in FIG. 8) (a) The computer R compares the load between its own computer and the computer S, and if the difference exceeds k, it is necessary to move the load, and the fact is reported. I do. In this case, the computer S is notified that the unexecuted job or process is to be moved. If the difference is equal to or smaller than k, as in the case C, the load does not move.

(B) The computer S disconnects the unexecuted job or process 311 to be moved from the queue 37 of its own computer and sends the job or process execution request received from the scheduler daemon 33 to the computer R. Queues an unexecuted job or process 311 in the execution wait queue 37 of its own computer based on the received execution request content.

(D) Both the computer S and the computer R report the load on their own computers to the scheduler daemon 33 in order to notify that the load on their own computers has changed.

(Iii) Case C: No load shift occurs. (See Case C in FIGS. 7 and 8)

[0038]

According to the present invention, even in a computer system having no shared memory, the scheduler computer can transfer the process to a computer with a low load by using the accurate load information of each computer.

[Brief description of the drawings]

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

FIG. 2 is a configuration diagram of a computer shown in the configuration example of FIG. 1;

FIG. 3 is a system configuration diagram showing an embodiment of the present invention.

FIG. 4 is a diagram showing exchange of information between a message transmitting computer and a message receiving computer.

FIG. 5 is a diagram showing exchange of information between a message transmitting computer and a message receiving computer.

FIG. 6 is a diagram showing exchange of information between a message transmitting computer and a message receiving computer.

FIG. 7 is a flowchart of a message transmitting computer.

FIG. 8 is a flowchart of a message receiving computer.

[Explanation of symbols]

11: Computer 12: Network 21: CPU 22: Memory 23: External Storage Device 231: Load Information File 24: Bus 31: Job or Process Requested by User 311: Job or Process Requested by User, Scheduler A computer to which a computer is assigned 32: Job or process reception queue 33: Scheduler daemon program 34: Load information file 35: Load monitoring daemon 36: Job execution daemon 37: Execution waiting process or job queue

Claims (2)

[Claims] 1. A parallel computer system in which a plurality of computers are connected via a network, a scheduler computer having a scheduler daemon for determining which computer should execute a process and a job received from a user, In a system composed of computers that simply execute processes and jobs (hereinafter simply referred to as computers), when performing inter-node communication (communication processing between computers) such as message passing between computers, transmission is performed. By exchanging load information between computers by adding the load information of the own computer to the message,
A load distribution control method characterized by maintaining the load balance between computers. 2. In a parallel system in which a plurality of computers are connected by a network, and a plurality of computers perform inter-node communication, a device for adding load information of the own computer to a transmission message from the own computer to each computer. And a device for recognizing load information of another computer added to the received message, and a device for moving a job or process waiting for execution assigned to a computer with a high load from the load information to a computer with a low load. A load distribution control device characterized by effectively correcting a load imbalance by performing the above operation.