JPH0713817B2 - Dynamic load balancing method for loosely coupled parallel computers - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of use] The present invention dynamically loads evenly on a loosely coupled parallel computer in which a plurality of processors are connected by a network and the communication speed is slower than the processing speed of the processors. The present invention relates to a dynamic load equalization method for a loosely coupled parallel computer.

[Prior Art] FIG. 5 shows, for example, the journal “IEE Transactions on Comput
ers, peges 1,098-1,109 vol 38, No.8, August 1989 ''.
The paper “GAMMON: A Load Balancing Strategy for Local Com” published by M. Baumgartner and BW Wah.
This is an example of the dynamic load balancing method on the distributed computer systems connected by the local area network, which is described in the 10th line from the top right of page 1099 of "puter Systems with Multiaccess Networks". In Fig. 5, 1-4 are "CPU1" and "CPU".
A plurality of processors with symbols “2”, “CPU3”, and “CPU4”, and these plurality of processors 1 to 4 are divided into a master processor that performs load distribution and a slave processor that receives load distribution. Specifically, "CPU
"1" processor as master processor 1 and "CPU
2 ”,“ CPU3 ”, and“ CPU4 ”are slave processors 2 to 4. Reference numerals 5, 6, 7, and 8 denote queues (queues) for storing executable jobs in the master processor 1 and slave processors 2 to 4, and 9
Reference numerals 16 to 17 indicate jobs that can be executed by the master processor 1 and the slave processors 2 to 4, and 17 to 10 are networks that mutually connect the master processor 1 and the slave processors 2 to 4, respectively.

Next, the operation of the conventional dynamic load balancing method on the distributed computer system will be described. In FIG. 5, the master processor 1 divides the given work to generate executable jobs 9 to 16 and allocates them to the slave processors 2 to 4. For example, in the above jobs 6 to 16, job 9
~ 13 are stored in the queue 5 of the master processor 1, jobs 14 and 15 are stored in the queue 6 of the slave processor 2, job 16 is stored in the queue 7 of the slave processor 3, and executed in the queue 8 of the slave processor 4. It is assumed that possible jobs are not stored. In this state, the master processor 1 allocates the jobs 9 to 13 to the slave processors 2 to 4 and equalizes the loads.

In order to equalize the load, the master processor 1 periodically checks the load weights of all the slave processors 2 to 4 and allocates jobs to the dynamically detected lightly loaded slave processors. The weight of the above load is the jobs of queues 2-4 in slave processors 2-4.
You can tell by the length of 12 to 16 stored. This means that a slave processor with a long queue 6-8 has a heavy load and is busy, and a short slave processor has a light load and is free. That is, the master processor 1 periodically polls the lengths of the queues 6 to 8 of the slave processors 2 to 4 and detects the slave processor having the shortest length of the queues 6 to 8 as the lightest loaded slave processor. .
To simplify the explanation here, the queue length is "0".
The slave processor of is the one with the lightest load. Therefore, the master processor 1 detects the slave processor 4 as the slave processor with the lightest load and distributes the load to the slave processor 4.

FIG. 6 shows how the master processor 1 polls the lengths of the queues 6 to 8 of all the slave processors 2 to 4 at regular intervals to find a slave processor having a light load. First, the master processor 1 sends a communication message (ReqL: Request Queue Length) 18,19,2 to the slave processors 2 to 4 to check the queue length.
Send 0. The slave processors 2 to 4, which have received these communication messages 18 to 20, communicate the communication messages (RQL: Report Queue Lengt
h) Send 21, 22 and 23 to the master processor 1 to report. In this example, the slave processor 2 reports “2” to the master processor 1, the slave processor 3 reports “1” to the master processor 1, and the slave processor 4 reports “0” to the master processor 1.
When the communication messages 21 to 23 from all the slave processors 2 to 4 arrive at the master processor 1, the master processor 1 finds the communication message 21 to 23 having the smallest queue length. In this example, since the length of the queue 8 according to the communication message 23 from the slave processor 4 is “0”, the master processor 1 determines that the slave processor 4 has the lightest load,
The master processor 1 newly allocates the job 9 to the slave processor 4.

FIG. 7 shows how the master processor 1 allocates the job 9 to the slave processor 4 having the lightest load. In FIG. 7, all slave processors 2 to 4 are in a busy state, that is, a heavy load. Then, after a lapse of a certain time, the master processor 1 detects again the slave processor with a light load among the slave processors 2 to 4, and the master processor 1 detects the slave processor with a light load among the jobs 10 to 13 among the slave processors with a light load. At least one of these is newly allocated. As described above, the master processor 1 dynamically equalizes the loads on the slave processors 2 to 4.

[Problems to be Solved by the Invention] In the conventional dynamic load equalization method, as described above, the master processor 1 slaves the slave processors 2 to 2 at regular intervals.
4, the queue length must be polled, and when allocating one job, when the number of processors including the master processor 1 and the slave processor 2 is N, (N-1) × 2 Need communication messages 18-23. Therefore, the number of message communications increases in proportion to the number of processors. Therefore, in a loosely coupled parallel computer in which the transfer speed of the network is slow compared to the processing speed of the processor, if the number of processors increases, network communication becomes a bottleneck, and it becomes difficult to perform load equalization completely. There is a problem that the expandability of the number of processors is reduced.

The present invention has been made to solve the above problems, and improves the expandability of the number of processors by distributing at least one job with one communication message from a slave processor to a master processor. In addition, a dynamic load balancing method for a loosely coupled parallel computer that can improve efficiency and speed of dynamic load balancing without requiring special hardware or operating system functions therefor is provided. With the goal.

[Means for Solving the Problems] A dynamic load balancing method for a loosely coupled parallel computer according to the present invention is a master processor of a plurality of processors, which assigns a job to a high priority job and a plurality of processors. Of the low-priority jobs that are equal to or greater than the number of slave processors, and this master processor assigns the above-mentioned high-priority jobs and low-priority jobs to slave processors, and the slave processors assign high-priority jobs. Slave processor that sends a communication message requesting a new job to the master processor when low priority jobs are executed by executing the communication messages in descending order, and the master processor sends the communication message by the arrival of this communication message At least one of the above low priority jobs It is characterized by allocation.

[Operation] According to a dynamic load equalization method in a loosely coupled parallel computer of the present invention, a master processor divides a given work into a job having a high priority and a job having a low priority equal to or more than the number of slave processors. When a slave processor executes a low-priority job while it is spawning and the master processor has assigned at least one low-priority job to the slave processor, the slave processor has its queue empty and the load lightened. Notify the master processor that By this notice,
The master processor assigns at least one low priority job to the lightly loaded slave processor.

[Embodiment] An embodiment of a method for dynamically balancing loads in a loosely coupled parallel computer according to the present invention will be described with reference to FIG. 1 to FIG. To do.

FIG. 1 is a configuration diagram showing a dynamic load equalization method in a loosely coupled parallel computer as an embodiment of the present invention. First
In the figure, "CPU1", "CPU2", "CPU3", "CPU4"
Four processors 1-4 marked with a symbol are network
They are connected at 17, and these processors 1 to 4 are divided into one master processor for load balancing and three slave processors for load balancing. That is, the processor "CPU1" is the master processor 1, and the processors "CPU2", "CPU3", and "CPU4" are the slave processors 2 to 4. Each of the master processor 1 and the slave processors 2 to 4 has two queues inside. The master processor 1 has a queue 25 for storing high priority jobs and a queue 35 for storing low priority jobs, and the slave processor 2 has a queue 26 for storing high priority jobs. And a queue 36 for storing low priority jobs, and the slave processor 3 has a queue 27 for storing high priority jobs and a queue 37 for storing low priority jobs. However, the slave processor 4 has a queue 28 for storing high priority jobs and a queue 38 for storing low priority jobs.

Further, the master processor 1 divides a given job into a plurality of jobs 9 to 16 and 30 to 34 that can be executed, and these jobs 9 to 16 and 30 to 34 are divided into high priority jobs 9 to 16 and slave processors. 30 or more low-priority jobs
Generate as 34.

And the master processor 1 is the slave processor 2
The jobs of high priority and at least one job of low priority are allocated to 4 to. For example, in jobs 9 to 16 with high priority, jobs 9 to 13 are stored in queue 25 of master processor 1, jobs 14 and 15 are stored in queue 26 of slave processor 2, and job 16 is queue of slave processor 3. Stored in 27. Among the jobs 30 to 34 having low priorities, the jobs 33 and 34 are stored in the queue 35 of the master processor 1, the job 30 is stored in the queue 36 of the slave processor 2, and the job 31 is stored in the queue 37 of the slave processor 3. , The job 32 is stored in the queue 88 of the slave processor 4.

In this allocated state, the master processor 1 and the slave processors 2 to 4 execute jobs 9 to 9 respectively.
Execute 16,30 to 34 in descending order of priority. That is, the master processor 1 executes the high priority job 13, the slave processor 2 executes the high priority job 15, the slave processor 3 executes the high priority job 16, and the slave processor 4 executes the priority job. Run job 32 with a low

FIG. 2 shows how the slave processor 4 in FIG. 1 executes a low priority job 32 and at the same time sends a communication message (Reqjob: Request Job) 18 for requesting a new job to the master processor 1. It is a representation. In FIG. 2, when the communication message 18 transmitted from the slave processor 4 arrives at the master processor 1, the master processor 1 can detect which slave processor among the slave processors 2 to 4 is currently lightly loaded. .

FIG. 3 shows that the master processor 1 that has detected the lightly loaded slave processor 4 has a high priority job 9 in its own queue 25 and at least 1 in its own queue 35.
This shows a state in which the low-priority jobs 33 are assigned to the slave processor 4 having a light load. In FIG. 3, a high priority job 9 is assigned from the queue 25 of the master processor 1 to the queue 28 of the slave processor 4, and a low priority job 33 is assigned from the queue 35 of the master processor 1 to the queue 38 of the slave processor 4. Assigned. The job 33 of low priority is executed by the slave processor 4 when the load of the slave processor 4 is lightened again by executing the job 33 after the slave processor 4 executes the job 9. This is a job for transmitting the communication message 18 (see FIG. 2) for requesting a job again.

In FIG. 4, the slave processor 3 notifies the master processor 1 that the load is lightened by executing the low priority job 31 in FIG. 3, and the master processor 1 notifies the high priority job 10 and the priority The low job 34 and the slave job 3 are newly allocated to the slave processor 3.

As described above, the slave processors 2 to 4 each request the master processor 1 to allocate a job with the communication message 18 each time the load is lightened by executing a job having a low priority. By dynamically allocating high-priority jobs and low-priority jobs to the respective slave processors 2 to 4 whose loads are lightened, the loads are equalized.
That is, each time a single communication message 18 is sent from the slave processors 2 to 4 to the master processor 1, at least one low-priority job is dynamically assigned to the lightly loaded slave processors 2-4. Such a communication bottleneck does not occur, jobs are distributed, and the expandability of the number of processors is improved.

As described above, in this embodiment, the work to which the master processor 1 responsible for load balancing is given is divided into high priority jobs 9 to 16 and low priority jobs 30 to 34 equal to or more than the number of slave processors. While slave processor 2
4, the slave processors 2 to 4 execute the low-priority jobs 30 to 34 thus assigned, and at the same time request a new job from the master processor 1 by a communication message 18 to receive the communication message 18. The master processor 1 that has done the above communication message 18
Jobs with high priority and jobs with low priority are assigned to the slave processors 2 to 4 having the lightest load that communicated with each other.

Therefore, in the above-mentioned conventional method, in order to detect the idle slave processors 2 to 4, the master processor 1 polls the lengths of the executable jobs in the queues 5 to 8 of the slave processors at regular intervals, and the To distribute the jobs, when the number of processors including the master processor 1 and the slave processors 2 to 4 is N,
According to the present embodiment, (N-1) × 2 message communication is required, but according to this embodiment, the special priority of the slave processors 2 to 4 is assigned without requiring any special hardware or operating system function. The communication message 18 is transmitted to the master processor 1 every time the jobs 30 to 34 having a low communication rate are executed.
By receiving 18 from the master processor 1, the master processor 1 was able to detect the slave processor with the lighter load among the slave processors 2 to 4, and the master processor 1 had the lighter load. Jobs of high priority and jobs of low priority can be distributed by one communication message 18 to the slave processors. As a result, in this example, the minimum number of communication messages 18
The load can be dynamically equalized and the expandability of the number of processors is improved.

In the above embodiment, the queues storing executable jobs are queues 25 to 28 storing high priority jobs.
Although a loosely coupled parallel computer having two types of queues, namely, queues 35 to 38 for storing jobs with low priority, and a queue for storing jobs with multi-level priority, the above implementation is performed. Jobs with low priority in the example 30 ~
By replacing 34 with the job with the lowest priority, the same application as in the above embodiment is possible.

As described above, according to the present invention, a master processor divides a given work into a high-priority job and a low-priority job equal to or more than the number of slave processors among a plurality of processors. , The master processor assigns at least one low-priority job to each slave processor, and the slave processor executes the low-priority job and notifies the master processor that the load is lightened at the same time. , The master processor assigns at least one low-priority job to the lightly loaded slave processor, so that the slave processor has a low priority without requiring any special hardware or operating system functionality. By running the job The master processor can detect the lightly loaded slave processor and one communication message sent from the slave processor to the master processor causes the master processor to distribute at least one job to the lightly loaded slave processor. As a result, even if the number of processors increases, the number of communication messages required for dynamic load balancing can be minimized, and the scalability of the number of processors can be improved. The effect is that dynamic load equalization can be performed efficiently and at high speed.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing a dynamic equalization method in a loosely coupled parallel computer according to an embodiment of the present invention, and FIG. 2 is a job in which a slave processor has a low priority in the above embodiment. Is executed and at the same time a communication message requesting a new job is transmitted to the master processor, and FIG. 3 is an illustration showing a job newly allocated to the slave processor having a light load in the above embodiment. FIGS. 4 and 5 are explanatory views showing how jobs are newly allocated to another slave processor having a light load in the above embodiment, and FIG. 5 is dynamic load distribution in a distributed computer as a conventional loosely coupled parallel computer. FIG. 6 is a block diagram showing the equalization method, and FIG. 6 shows that the master processor finds a slave processor with a light load in the above conventional example. FIG. 7 is an explanatory view showing the child, and FIG. 7 is an explanatory view showing how the master processor allocates a job to the slave processor having the lightest load in the conventional example. 1 ... Master processor (CPU1), 2,3,4 ... Slave processor (CPU2,3,4), 9,10,11,12,13,14,15,16 ...
… High priority jobs, 17 …… Network, 18 …… Communication messages, 25,26,27,28 …… Queues for storing high priority jobs, 30,31,32,33,34… ... low priority jobs, 35, 36, 37, 38 ... queues for storing low priority jobs.

Claims (1)

[Claims] 1. A master processor that performs load balancing and a slave processor that receives load balancing by this master processor are connected by a network, and load is loaded on a loosely coupled parallel computer whose communication speed is slower than the processing speed of the processor. In the dynamic load balancing method of dynamically balancing, the master processor divides a given work into a job with a high priority and a plurality of jobs with a low priority equal to or more than the number of slave processors. The master processor allocates the high-priority job to some of the slave processors, and at least allocates one low-priority job to all the slave processors. Jobs from highest priority to lowest priority And sending a communication message requesting a new job to the master processor when the low priority job is executed, and the arrival of this communication message causes the master processor to send the communication message with the above priority. A method of allocating at least one low priority job to a slave processor which does not have a low job, and a dynamic load balancing method in a loosely coupled parallel computer including: