JP2526399B2 - Load balancing method for parallel computers - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a plurality of processing
The present invention relates to a load balancing method in a parallel computer composed of elements.

[0002]

2. Description of the Related Art In a conventional load balancing method for a parallel computer, processing executed as described in Information Processing Society of Japan, Research Report 89-ARC-77; "Slit check mechanism for speeding up software interrupt processing between processors" There is a load balancing method in which a processor that has become empty requests all other processors to distribute the processing and receives the processing (hereinafter, referred to as a first conventional technique). Also,
Ohmsha's Computer Architecture Series, The Institute of Electronics, Information and Communication Engineers, "Parallel Processing Machine" As described in the threshold method on page 154, when the load of the user is less than the threshold, a new process is assigned to the operator and the threshold If it exceeds the value, it is assigned to another processing element that does not exceed the threshold value (hereinafter referred to as the second conventional technique).

[0003]

In order to improve the performance of a parallel computer, it is necessary to improve the operating time of each processing element and reduce the overhead of the synchronization processing between each processing element. A load balancing method that combines the first and second prior arts described above, that is, the processing engine with no load.
When the threshold value is set to a large value, processing load does not occur when the threshold value is set to a large value, so the load on each processing element varies. Occurs and the operating rate decreases. Further, when the threshold value is made small, distribution occurs not only from the processing element having a large amount of processing but also from the processing element having a small processing, and this distribution increases unnecessary overhead.

It is an object of the present invention to provide a load balancing system in a parallel computer which improves the operating rate of processing elements by distributing a small amount of processing, thereby improving the overall performance.

[0005]

In order to achieve the above object , a plurality of professionals connected to each other by a network are used.
By operating the processing and elemet in parallel, each processing
Each processing element is executed and executed.
From a processing element that has no processing to do
To other processing elements via the network
Requesting distribution of processing and other processing elements
Exceeds a certain threshold amount of work to be performed
, The processing requester of the requesting distributor of the processing
In a parallel computer that distributes processing to
Set an initial value to the threshold value of the sessing element,
One process for any processing element
Distribution requests are not made by other processing elements.
New processing needs to be distributed within a specified time after
If requested, leave it in that processing element
Decrease the threshold value, and within that time new processing
If there is no distribution request, change the threshold value to the initial value
I do.

[0006]

When a plurality of processing distribution requests arrive within a certain period of time, it is determined that many processing elements do not have processing and the processing distribution frequency is lowered in order to lower the processing distribution threshold. The operation rate is improved. If multiple processing distribution requests do not arrive within a certain time, it is determined that many processing elements already have processing, and the processing distribution threshold value is set to a certain value (initial value).
To change to a higher value, that is, to change to a higher value
In, it is possible to improve the operating rate in distribution of less processing for unnecessary processing of the distribution from the processing elements that do not have processing much does not occur.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows a flow chart in the case where a load balancing system which is an embodiment of the present invention is realized in a processing element of a parallel computer. The configuration of the parallel computer is shown in FIG. First, FIG. 1 will be described. Step 1
Then, it is determined whether or not there is a process to be executed by the processing element. Step 2 if there is processing to be executed
If not, go to step 3. In step 2, the process is executed. Here, the processing is, for example, a partial program of a program in which one program is made up of a plurality of partial programs, which can be executed in parallel. After this, the process proceeds to step 4. In step 3, the other processing elements are requested to distribute the processing. Here, the processing distribution request is sent as a network message to another processing element. This message may be sent to all the other processing elements, or may be sent to any one of the processing elements. In step 4, it is determined whether there is a message that has arrived from the network. If there is an arrival message, proceed to step 5, otherwise proceed to step 1. In step 5, it is determined whether the arrival message is a processing distribution request. If it is a processing distribution request, the process proceeds to step 7, and if not, the process proceeds to step 6. In step 6, the arrival message is processed as a normal message. The processing in step 6 is processing related to step 2 (for example, data communication between two partial programs) or processing for receiving distribution of processing (for example, registration of a partial program in an execution waiting queue). Not shown).

Step 7 and thereafter are cases where there is a processing distribution request. First, in step 7, the time NEWT when the processing distribution request was previously received is set to OLDT. Where N
The EWT and OLDT may be registers or memories. Next, the value of TOD is set in NEWT. TOD is a clock unique to the system. After that, TT is compared with the value obtained by subtracting OLDT from NEWT, and if TT or less, step 10
Otherwise, go to step 11. Here, TT is a preset time. In step 10, one half of the threshold value TH is newly set to TH. The value to be newly set is truncated after the decimal point. In step 11, a threshold value TH is set to a predetermined initial value. In step 12, if the load amount is larger than the threshold value, the process proceeds to step 13, and if not, the process proceeds to step 4. In step 13, one of the processes to be executed is converted into a process distribution message and sent to the request source to distribute the process, and then the process returns to step 4.

FIG. 2 shows the configuration of the parallel computer as described above. The parallel computer includes a plurality of processing elements PE1 to PEn and a network NET interconnecting the processing elements PE1 to PEn. This processing element PE1
.. to PEn, the execution of processing is controlled based on the flowchart shown in FIG. The network NET only needs to be able to realize the transfer of messages between arbitrary processing elements, and its internal configuration does not matter.

Next, an operation example of this parallel computer will be described. Here, the case where the number of PEs is 4 is shown. An example is shown in which the threshold is reduced and the processing is distributed, the threshold is changed to the initial value, and the disposal is suppressed. First, in FIG. 3A, it is assumed that all PEs initially have four processes (load amount = 4). At this time, the threshold value TH is set to the initial value, which is 16 here. Next, in FIG. 3A, it is assumed that time elapses from the state of FIG. 3A and the processes other than PE1 are lost one after another. Further, it is assumed that the number of processes increases as a result of execution in PE1 and eight processes remain (load amount = 8). At this time, PE2, PE
3, PE4 sequentially sends processing distribution request messages to PE1 via the network. Then, if the preset time TT in FIG. 1 is an appropriate value,
The arrival time interval of the process distribution request message becomes smaller than TT, which causes PE1 to perform the step 10 of FIG.
Is executed, the threshold TH is changed from 16 to 8, and further, 8
It is changed from 4 to 4 and becomes smaller than the load amount of PE 1, which is 8. At this point, the distribution of the processing in step 13 of FIG. 1 is executed and PE2, PE3, PE are processed as shown in FIG.
Processing is distributed to four. If the execution time of the distributed process is long, or if it is a process that creates a new process, PE2, PE3, and PE4 send a process distribution request message.
Is no longer sent to PE1. After that, a new load
When the request message arrives, as shown in FIG.
The threshold value TH of PE1 is set to 1 by step 11 in FIG.
6, the distribution of processing is suppressed. By controlling the above processing, it is possible to suppress the wasteful distribution of processing while improving the operating rate of each PE.

[0011]

According to the present invention, it is possible to reduce the overhead while improving the operating rate of the processing element, and thereby improve the performance of the parallel computer.

[Brief description of drawings]

FIG. 1 is a flowchart illustrating a load distribution method according to an embodiment of the present invention.

FIG. 2 is a diagram showing the configuration of a parallel computer that implements the present invention.

FIG. 3 is a diagram showing a part of an operation example of a parallel computer embodying the present invention.

FIG. 4 is a diagram showing another portion of the operation example of the parallel computer embodying the present invention.

[Explanation of symbols]

PE1, PE2, PE3, PE4, PEn ... Processing element, NET ... Network.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Machiko Asaya 1-280, Higashi Koigokubo, Kokubunji City, Tokyo Metropolitan Research Laboratory, Hitachi, Ltd. (56) Reference JP-A-3-88060 (JP, A)

Claims (2)

(57) [Claims] Claim: What is claimed is: 1. A plurality of processing elements connected to each other by a network are operated in parallel so that each processing element executes a process, and a processing element from which a process to be executed is no longer present is the network. Requesting the distribution of processing to other processing elements via the
If the amount of processing to be performed by an element exceeds a certain threshold value, the threshold of each processing element is determined by the parallel computer that distributes the processing to the processing element that requested the distribution of the processing. Set the initial value
If a distribution request for one process is made to one of the processing elements from another processing element and a distribution request for a new process is made within a predetermined time, the processing Load balancing in a parallel computer, characterized in that the threshold value in the element is reduced, and if there is no request for distribution of new processing within that time, the threshold value is changed to the initial value. Method. 2. The parallel computer according to claim 1, wherein the process of reducing the threshold value comprises a process of changing the value to one half of the current threshold value (rounding down after the decimal point). Load balancing method.