JP3103129B2 - Load balancing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[Object of the Invention]

[0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a load distribution method used in a computer system having one or more processors.

[0003]

2. Description of the Related Art In a conventional load distribution system, when a large number of tasks are arranged on a single processor, a cache memory overflows, data in the cache memory is frequently replaced, and cache thrashing occurs. There was something. Further, when a related task group distributed to a plurality of processors reads and writes common data, when a certain processor updates data in the cache memory and invalidates data in the cache memory of another processor, Since the processor reads the invalidated data from the main storage device again, the exchange of data between the cache memory and the main storage device becomes frequent, and cache thrashing may similarly occur.

[0004]

As described above, in the conventional load distribution method, when a large number of tasks are arranged on one processor and a cache memory overflows, a related task group distributed to a plurality of processors is used. However, there is a problem that cache thrashing occurs when common data is accessed.

The present invention has been made to solve such a problem, and an object of the present invention is to provide a load distribution method for realizing efficient use of a system by preventing cache thrashing.

[Configuration of the Invention]

[0007]

For load balancing scheme SUMMARY OF THE INVENTION The first invention to achieve the above object, in a computer system having multiple processors, each processor,
It has a function to communicate its current load status with each other.
When a task group to be processed occurs, this task group
Is processed by itself and a group of processors with a lower load
Predict the average load for the case, and this average load
The other processor higher than the load and the next highest
Processor that has determined that the current load of
The processor having a load equal to or less than the average load
Each of the above processors is assigned a task group
It is characterized in that information for task distribution is broadcasted to the server .

[0008] For load balancing method of the second invention to achieve the above object, in a computer system having multiple processors, each processor of its own current
It has a function to communicate the load status with each other, and
When a task group occurs, the task group
Average negative load when processing is performed by processors with lower load
Predict the load, and if this average load is higher than
Below the other processor's load, and
A processor that determines that the value is less than a predetermined value
But to a processor having a load less than the average load,
Each of the above tasks is assigned to assign a newly generated task group.
It is characterized in that information for task distribution is broadcast-communicated to the processor .

[0009]

According to the first invention, a newly generated task group is set to a negative
By assigning only to processors with low load, the number of processors sharing the required task group is reduced,
This makes it possible to prevent cache thrashing.

In the second invention, the average load of each processor
Task group assignment only when
By doing so, the number of tasks assigned to each processor can be reduced, thereby preventing cache thrashing.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of a computer system using a load distribution method according to an embodiment of the present invention.

As shown in FIG. 1, the computer system comprises a plurality of processors 1, 2, 3,..., N connected through a communication medium 10. Each of the processors 1 to n includes process management devices 21, 22, 23,..., 2n and processing devices 31, 32, 33,. The process management devices 21 to 2n manage the execution request task group while communicating with other process management devices. Processing device 31
3 to 3n are mainly composed of a central processing unit, an execution control unit, etc., and actually execute tasks.

FIG. 2 shows an example of a method of storing data stored in the process management devices 21 to 2n. As shown in the figure, the process management devices 21 to 2n store a newly generated task group and a task that has started execution in separate queues 41 and 42, respectively. The following values are stored in the process management devices 21 to 2n as data for performing the load distribution processing.

(1) The total load Ni on the own processor.

(2) The number Li of itself and a group of processors with a lower load than itself.

(3) The total load Si of the processor itself and the processor group having a lower load than the processor itself.

(4) The minimum load Ui (the next highest load) among the loads of the processor group having a higher load than itself.
However, if the processor itself has the maximum load, a sufficiently large number is stored as Ui.

Here, the processors having the same load are ordered according to the order predetermined by the processor name.

Next, the load distribution method of this embodiment will be described with reference to the flowcharts of FIGS. First, at the time of system startup, the above data is initialized as follows. That is, since there is no task on each processor when the system is started, Nl = 0, Li = i, Si = 0, Ui = 0 or a sufficiently large integer.

When N new tasks are generated on a certain processor, a request for starting the tasks and information including the number of tasks are broadcast to all processors as task start request communication.

Each processor that has received the task activation request communication (step 301) first receives the average load Avei of each processor group when the processor itself and a processor group with a lower load process the newly generated task group. Is calculated (step 302). Here, the average load Avei
Is greater than its own load Ni (step 303) and does not exceed its next highest load Ui (step 30).
4) The processor i that has determined the above performs the following processing.

That is, when the maximum integer s equal to or less than the average load Avei is equal to or less than a predetermined constant s ^* (step 305), s' = s + a (a ≧ 0 is a constant)
(Step 306). If s is larger than the constant s ^* , s' = s is set (step 307). At this time, information consisting of s and the number R of surplus tasks when a newly generated task group is distributed so that the load of each processor is equal to or less than s ′ in order from the processor with the smallest load is broadcast communication as task distribution determination communication. (Step 308,
309).

On the other hand, the processor which has received the task distribution determination communication (step 401) performs the following processing.

That is, when the load of the self is equal to or less than s (step 402), it is determined that the self is the processor that handles the task. And when s ≦ s ^* (step 40
3), replacing s with s + a (step 404),
Based on the algorithm, “from the beginning of the task group, tasks are received from the beginning of the task group so that the load on each processor becomes s until there are no unassigned tasks, starting from the processor with the lowest priority. Calculates the number of tasks to be assigned to the processor group having a lower rank for (a), selects the task to be assigned to the processor itself (step 405), activates these tasks, and stores the task based on the number of accepted tasks. Update data. When the load of the own processor exceeds s, the data stored in the own processor is updated based on the number of tasks assigned to the processor whose load order is lower than that of the own processor.

When s> s ^* , the above processing except for step 404 is executed.

When one task is completed in the processor i, the load on the own processor is updated and the own processor i
The information consisting of the load Ni and the own processor name is broadcasted as task end communication.

The processor that has received the task end message,
The data of the load of the other processor stored in the own processor is updated, these values are sorted, the order of the processors is determined, Li, Si, and Ui are calculated and updated.

As described above, in a situation where the load on the processor is somewhat low, that is, when s ≦ s ^* , the upper limit s of the load on the processor as a result of assigning the task group is increased to some extent, that is, s is given by s + a. In this case, the number of processors that handle the task group can be reduced without breaking the load balance of the entire system, so that cache thrashing can be prevented.

Next, another embodiment of the present invention will be described. FIG.
FIG. 6 is a flowchart showing the flow of the load distribution method of this embodiment. In this embodiment, step 5
The operation up to the operation of calculating the average load Avei of each processor group up to 02 is the same as in the previous embodiment.

In this embodiment, the average load Avei is larger than the own load Ni (step 503), does not exceed the next highest load Ui (step 504), and is equal to or less than a predetermined constant A. Only in a certain case (step 505), it is determined that the processor itself and the processor group having a lower load than the processor itself process the newly generated task group. Then, information consisting of the maximum integer s equal to or less than the average load Avei and the number R of surplus tasks when the load of each processor is set to s by distributing a newly generated task group is broadcast as task distribution determination communication. Communication is performed (steps 506 and 507).

On the other hand, the processor that has received the task allocation decision communication (step 601) determines that it is the processor that is responsible for the task if its own load is less than or equal to s (step 602). Then, the number of tasks to be handled by a processor having a lower rank with respect to the load is calculated to select the task to be handled by itself (step 60).
3) Activate these tasks and update the data stored in the own processor based on the number of accepted tasks. When the load of the own processor exceeds s, the data stored in the own processor is updated based on the number of tasks assigned to the processor whose rank is lower than that of the own processor.

When one task is completed in the processor i, the load of the own processor is updated, and information including the load Ni of the own processor and the name of the own processor is broadcasted as task end communication.

The processor which has received the task end message,
The data on the load of the other processor stored in the own processor is updated, these values are sorted, the order of the processors is determined, and Li, Si, and Ui are calculated and updated.

As described above, by allocating the task group only when the load of the processor becomes lower than a certain value, the number of tasks assigned to each processor can be reduced. Thrashing can be prevented.

[0036]

As described above, according to the load distribution method of the present invention, the number of processors sharing a newly generated task group can be reduced, and the number of tasks assigned to each processor can be reduced. This makes it possible to prevent cache thrashing.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a computer system using a load distribution method according to an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a data storage method of the process management device in FIG.

FIG. 3 is a flowchart for explaining the operation of the task activation request communication receiving side in the load distribution method according to one embodiment of the present invention.

FIG. 4 is a flowchart for explaining the operation of the task distribution determination communication receiving side in the load distribution method according to one embodiment of the present invention.

FIG. 5 is a flowchart illustrating an operation of a task activation request communication receiving side in a load distribution method according to another embodiment of the present invention.

FIG. 6 is a flowchart illustrating an operation of a task distribution determination communication receiving side in a load distribution system according to another embodiment of the present invention.

[Explanation of symbols]

 1, 2, 3, n processor 10 communication medium 21, 22, 23, 2n process management device 31, 32, 33, 3n processing device

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06F 9/46 G06F 15/16

Claims (2)

(57) [Claims] 1. A computer system having multiple processors, each processor its own current load shape
Has the function of communicating statuses with each other, and
When a task group occurs, this task group
Predict the average load when processing is performed by low-load processors.
The average load is higher than your current load, and
Exceed the current load of the other processor above itself
The processor that has determined that there is no
A task group newly generated in a processor having a load
Task allocation for each processor so that
A load balancing method characterized by broadcasting information for minutes . 2. A computer system having multiple processors, each processor its own current load shape
Has the function of communicating statuses with each other, and
When a task group occurs, this task group
Predict the average load when processing is performed by low-load processors.
This average load is higher than your own load,
Below the high load of the other processor, and
The processor that has determined that the value is
Processor with a load less than the average load
Each of the processors so as to allocate a generated task group
A load distribution method characterized in that information for task distribution is broadcasted to a user.