JPH10171673A - Network using method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform the parallel calculation in a large scale and at the low cost by accepting many tasks when the load of a computer where a program is resident is light and then decreasing the tasks when the computer load is heavy. SOLUTION: A task server 120 is resident at every computer 110 in a LAN and consists of a load monitoring part 121, a task management part 122 and a task execution part 123. The part 121 monitors the load of the computer 110. The part 122 permits the acceptance of tasks sent via the LAN after waiting by 1W second if the calculated load 1 is not more than 3. If the load 1 is more than 3, the task acceptance requests are disregarded. The W is decided based on the performance of the computer 110. If the load 1 is larger than 4, the part 123 selects the optional one of tasks under execution to discontinue its execution and shifts this task to another computer.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for utilizing a network, in particular, an entire local area network (LAN) as one loosely coupled parallel computer.

[0002]

2. Description of the Related Art There is known a technique called a workstation cluster in which a plurality of general-purpose workstations are connected and used as one parallel computer. With this technology, large-scale calculations that previously required a supercomputer can now be performed at low cost with a workstation cluster.

However, the construction of a workstation cluster requires a high-speed and dedicated network. Further, the workstation itself is more expensive than a general personal computer, and it can be said that the workstation is still suitable for special use in view of the construction cost.

On the other hand, it is used in ordinary offices and the like.
The LAN is constructed by connecting a large number of inexpensive personal computers and one or several server computers with low-cost cables such as Ethernet cables.

[0005] The main uses of LANs are disk sharing, printer sharing, and CPU sharing. By sharing disks,
Information stored on the disk of the server computer can be shared throughout the office. By sharing printers, a small number of printers (often connected to a server computer) can be used by many people. CPU
By sharing, the CPU of the server computer can be used from a personal computer or the like.

[0006] CPU sharing is a remote procedure call
(RPC, remote procedure call) or remote method invocation (RMI, remote method call)
This is done by calling a program on the server computer using such a technique. An example of CPU sharing is a system that runs a database server on a server computer and searches a database in response to an inquiry from a client.

[0007] Each of the methods uses only the resources of a small number of server computers on the LAN.

The genetic algorithm (GA) has been applied to combinatorial optimization problems in various fields. For example, GA
Bus schedule diagram creation device (JP-A-8-22740)
No. 9). GAs require a lot of CPU power, but recent improvements in the performance of personal computers have made it possible to handle real problems, such as creating bus schedules, on personal computers.

The GA is suitable for distributed processing because it processes many individuals in parallel, but requires a dedicated machine for parallel computation.

[0010]

The performance of personal computers has been remarkably improved, and the performance of personal computers has surpassed that of lower models of workstations. This is the result of a desire for a user-friendly graphical user interface (GUI), and a high-speed CPU is indispensable for realizing the GUI.

However, in the GUI, a large amount of CPU is always used.
Power (the amount of computation processed by the CPU per unit time) is not required. For example, most of the CPU power is not used when the user is reading a sentence displayed on the display or performing character input. Therefore, when looking at the entire LAN, a lot of CPU power is left.

It is an object of the present invention to collect surplus CPU power not effectively used in a LAN and use it for large-scale calculations and the like. In the present invention, a large-scale calculation can be performed with little cost because the CPU power that has not been used conventionally is effectively used. For example,
A solution based on GA can be executed at high speed using LAN.

[0013]

A program for performing a large-scale calculation is divided into a number of processes that can be executed in parallel, and each process is processed in parallel on each computer on a LAN. Here, the minimum unit for parallel processing is called a task. A task includes a process that can be calculated by itself and a process that performs an interaction with another task. The processing content needs to be executable on various computers.

For example, a task is described by a binary code of a virtual machine (or an actual machine), and an interpreter for interpreting and executing the binary code is mounted on each computer. Alternatively, instead of the interpreter, a compiler that converts the code into a code that can be processed by each computer immediately before executing the task may be implemented.

Here, a program that accepts and executes a task is called a task server. Task server in advance
Make it resident on each computer on the LAN. In other words, the task server is run in the background separately from the normal business processing. However, the task server runs at a low priority so as not to affect the operational feeling of the GUI in normal business.

The task server is a CPU of the computer.
Monitor the computer load to determine if there is excess power. When the load is light, an acceptance permission is returned to the task acceptance request. When the load is heavy,
Reduce or stop accepting tasks. Further, a task acceptance request is sent to another computer arbitrarily selected, and the task currently being processed is moved to the computer for which permission has been returned, thereby reducing the number of tasks to be processed.

In the case of a multi-user multi-task OS,
The load fluctuates slowly because many users use it at the same time. Therefore, even if a task is transmitted and received after detecting a change in load, load distribution can be achieved. However, in the case of a single-user multitasking OS mainly used in personal computers, the load changes abruptly when a user starts an application. Therefore, if a task is transmitted / received after detecting a change in load, a decrease in performance cannot be avoided. That is, the user feels a decrease in usability such as a slow start of the application.

Therefore, a plurality of computers whose load changes rapidly are collected and handled as a single task server. This is called a task server cluster. A master task server (hereinafter abbreviated as “master”) resides on one computer of the task server cluster, and a slave task server (hereinafter abbreviated as “slave”) resides on the other computers.

The slave sends the load detected every moment to the master, and the master calculates the average value of the load of each slave and the load detected by the master itself. When the average value of the load is light, the master returns acceptance permission in response to the task acceptance request. If the average load is heavy, reduce or stop accepting tasks.

The tasks accepted by the master are copied to other slaves. However, the task can be executed only by the computer with the lightest load in the cluster, and the other computers are in a state of waiting for execution. When the load on the computer that is executing the task increases, the execution is immediately stopped, and only the minimum data necessary to continue the execution of the task is transferred to the computer with the lightest load in the cluster at that time. To take over the task execution.

When a large amount of data is required when executing a task, a disk may be shared between the computers in the cluster, and the amount of data to be transferred may be reduced by placing data on the disk. Or during task execution,
The data may be sequentially transferred to another computer in the cluster, and the amount of data transferred when the task is taken over may be reduced.

When a user starts an application or starts a calculation that takes a long time to process, the performance of the requested process reduces the amount of processing. When the load suddenly increases, the processing amount of each task being processed may be reduced or temporarily interrupted so that the performance of the processing requested by the user does not deteriorate. Alternatively, instead of detecting a change in load, a change in load may be predicted by monitoring a system call of the OS, and measures such as task movement may be taken before the load actually becomes heavy.

To start large-scale calculation, first, a plurality of tasks are generated. A program that generates a task is called a client. The client has a unique number through the LAN and gives the created task that number.
Next, the client sends a task acceptance request to the task server, and sends the task to the task server that has accepted the task.

An estimate of the amount of memory required for executing the task is sent together with the task acceptance request, and the task server issues an acceptance permission if the amount of memory required by the task exceeds the available memory at that time. It may not be necessary. Tasks with the same client number interact via LAN. Also, the task notifies the client of the processing status and the calculation result at that time, such as when the condition specified by the client is satisfied in advance, for example, when the calculation of the task is completed.

The client sends a message together with the number to each task server to notify the client of the processing status of the task having the same client number, the calculation result, and the like. Alternatively, by sending a calculation end request message together with the client number, the calculation of the task having the same client number is ended, and the task is deleted.

[0026]

BEST MODE FOR CARRYING OUT THE INVENTION A LAN of a branch office of a bus company will be described below.
The present invention will be described in detail with reference to an example.
FIG. 1 shows a configuration diagram. In business offices, several to dozens of computers are used for the purpose of creating bus schedules, managing bus operations, managing employees, and the like, and are connected via a LAN to share bus schedule databases and the like.

The calculation that requires the CPU power at the sales office is the preparation of a bus schedule, that is, the scheduling of assigning crew members to each bus operation. The method of solving the bus schedule diagram based on GA is described in the above-mentioned conventional example. In this solution, in order to generate a working schedule that satisfies the constraints in a medium-sized office,
It takes about 10 minutes with a normal personal computer.

A method for reducing the time required for creating a bus schedule by distributing and executing the GA on a LAN will be described. A task server 120 is resident on each computer 110 on the LAN. The task server is a load monitoring unit 12
1. It comprises a task management unit 122 and a task execution unit 123.

The load monitor 121 monitors the load on the computer 110. Generally, multitask OS is “run queue”
It has an execution queue of executable threads called.
The time average of this matrix length is called the load average. The higher the load average value, the heavier the computer load. That is, the execution speed of the program is reduced. Therefore, the average value of the queue length of the run queue for 30 seconds is referred to as the load 1 (ell) of the computer 110.

If the load 1 calculated by the load monitoring unit 121 is less than 3, the task management unit 122 waits 1W seconds for a task acceptance request sent via the LAN, and then returns acceptance. If the load 1 is 3 or more, the task acceptance request is ignored. W is determined based on the performance of the computer. That is, a high-performance computer sets a small W value, and a low-performance computer sets a large W value. If you receive a task after returning permission,
It is accepted and passed to the task execution unit 123.

When the load 1 is 4 or more, the task execution unit 123
Arbitrarily selects one of the running tasks, stops its execution, and moves the task to another computer. The destination first sends an acceptance request to all task servers,
The task server to which permission is returned first is assumed. After the movement is completed, the client 101 is notified of the destination.

If the load 1 is less than 4, the task execution unit 123
Notifies the current load to each of the running tasks. Since tasks are executed with a low priority, when the load becomes heavy, the CPU time spent for executing the tasks decreases.

The client 101 sends a task acceptance request to all task servers, and sends tasks to the N task servers in order from the task server to which acceptance acceptance is returned earlier. The client stores a list 102 of the task servers that have sent the task, and when the task is moved, receives a notification of the destination from the task server and updates the list.

A task is composed of the following variables and methods. I: Client number H: Transfer destination task server T: Transfer interval Mmax, Mmin: Upper and lower limit of number of individuals R: Number of transfers S: Individual set GA method Transfer method Load adjustment method Each individual included in individual set S is one bus Corresponds to the work schedule. The individual set is randomly generated when a client generates a task. Therefore, the bus schedule corresponding to the initial individual is of low quality that does not satisfy the constraints, but by executing the GA method the bus schedule is `` evolved '', that is, improved, and by sufficiently changing generations, You can get the best work schedule. The GA method is started by the task execution unit 123 and is continuously executed until the task ends.

FIG. 2 shows a flowchart of the GA method. Steps 202 to 205 are genetic manipulations of the GA, which are the same as those in the above-described conventional technology. This genetic manipulation
After repeating T times, the transfer operations 208 to 210 are performed.

The transfer operation is an operation of sending an R individual randomly selected from the individual set to the transfer destination task server H. First, in step 208, the task server H is inquired about whether a task having the client number I exists in the task server H.

If the task having the client number I has moved to another task server and does not exist in H,
Inquire the client 101 of another task server where a task having the client number I exists (step 209),
Update the value of the transfer destination task server H. Client 101
Returns a task server randomly selected from the server list 102 in response to this inquiry.

If the task exists in H, step 210
Sends R individuals to the transfer-destination task server H. The task server to which the individual has been sent adds the individual to the individual set of the task by calling the transfer method of the task having the corresponding client number. Since the GA method and the transfer method are executed asynchronously, a semaphore is used to exclusively change the set of individuals.

The load adjustment method is called by the task manager when the load 1 is less than 3. The load adjustment method sets the upper limit Mmax of the number of individuals close to the lower limit Mmin if l is large. For example, Mmax ← Mmin + (Mmax−Mmin) / 10
And If l is small, the upper limit of the number of individuals is increased. For example, Mmax ← Mmin × (4-1).

The number of individuals fluctuates due to in / out. Therefore, adjustment is performed in the cross-proliferation 202 and the natural selection 204 so that the number of individuals is not less than Mmin and not more than Mmax. That is, when the number of individuals is below Mmin, the number of individuals generated by crossing is increased, and when the number of individuals exceeds Mmax, the number of individuals to be removed by natural selection is increased.

The client 101 periodically inquires each task server in the server list 102 about the evaluation value of the best individual at that time. The bus schedule corresponding to the individual with the highest evaluation value among the task servers is sent to the diamond composition system 103 and presented to the user. When the user determines that the displayed bus schedule is at a satisfactory level, the user presses the stop button. Then, the client ends the calculation by sending an end message to each task server.

The level to be satisfied by the bus schedule may be set in advance in the GA method of the task, and when the evaluation value of the best individual exceeds the level, the client may be notified of that and the calculation may be terminated.

[0043]

As described above, an easy-to-use GUI requires a high-speed CPU. However, when the user is reading or typing characters on the screen, most of the CPU power is not used. The surplus CPU power is enormous when viewed over the entire LAN. According to the present invention, it is possible to utilize the surplus CPU power. Because an existing LAN is used, large-scale parallel computation can be performed at extremely low cost.

The computation time can be reduced by executing the GA-based solution on a large number of computers on a LAN in a distributed manner. Alternatively, more individuals can be evaluated in the same calculation time, so that a high-quality solution can be obtained.
For example, in the example of creating a bus schedule, it can be solved within one minute by dispersing and processing the calculations that conventionally took about ten and several minutes by several tens of computers. Therefore, it is possible to sufficiently repeat the trial and error of repeating the generation of the work schedule while changing various conditions (for example, the departure and arrival times of the bus).

[Brief description of the drawings]

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

FIG. 2 is a flowchart of a GA method.

Claims (3)

[Claims] 1. A program for processing a task sent from another computer via a network is resident on each of a personal computer and a server computer connected to a local area network, and the local area network is loosely coupled to one. A method for utilizing as a parallel computer, wherein the program monitors a load of a computer in which the program resides, and accepts many tasks when the load is light, and reduces tasks to accept when the load is heavy. And how to use the network. 2. When the load is heavy, a group of a plurality of computers accepts the same task, and the computer with the lightest load processes the task at the time of acceptance, and the other computers wait for the execution of the task. 2. The network using method according to claim 1, wherein the network is used. 3. When the load of a computer processing the task becomes heavy, transferring data necessary for the continuous execution of the task to another machine of the computers belonging to the group and taking over the task processing. The method according to claim 1, wherein: