KR102116858B1 - Scheduling apparatus, and control method thereof - Google Patents

Abstract

The present invention relates to a scheduling device capable of improving the execution performance of an application program as well as optimizing the performance of a high performance computing system by scheduling concurrent executions between applications that have less interference with each other based on execution characteristics according to the use of resources of the application program, and an operation method thereof. The scheduling device comprises: a classification part; and a determination part.

Description

Scheduling device and its operation method {SCHEDULING APPARATUS, AND CONTROL METHOD THEREOF}

The present invention relates to a technique for scheduling concurrent execution between application programs with low resource interference based on execution characteristics of an application program in a high performance computing (HPC) environment.

With the development of process technology and the increase in performance required from application programs, as the performance required from applications increases, modern processors are able to integrate multiple cores on a single chip.

Advances in this technology have led to the emergence of a many-core type processor that contains dozens to hundreds of cores on a single chip, and high-performance computing that can execute multiple applications simultaneously on the processor. HPC, High Performance Computing (HPC) system has been established.

However, in order to optimize system performance in a high performance computing (HPC) environment supporting concurrent execution between application programs, it is necessary to consider the effect of interference between application programs.

Accordingly, the present invention intends to propose a new method for scheduling concurrent execution between application programs in consideration of the effect of interference between resources used when executing the application program.

The present invention was created in view of the above circumstances, and the object to be reached in the present invention is that, in a high performance computing (HPC) environment, mutual interference is based on execution characteristics according to resource usage of an application program. A small number is in scheduling concurrent execution between applications.

Scheduling apparatus according to an embodiment of the present invention for achieving the above object, a collection unit for collecting a performance event that is information about the resource used by each application program for execution for each application; A classification unit for clustering application programs for each resource according to resource usage of each application program identified from the performance event; And it characterized in that it comprises a determination unit for determining the execution characteristics of each application by collecting the clustering results in each resource for each application.

Specifically, the scheduling apparatus may further include a scheduling unit that determines the similarity between execution characteristics determined for each application program and schedules simultaneous execution between application programs having similarity or less than a threshold.

Specifically, the classification unit may set two or more clusters for distinguishing relative resource usage between application programs for each resource, and classify one application program into any one of the two or more clusters for each resource.

Specifically, the determining unit, if the group according to the resource usage according to the length from the center of the regular polygon to each vertex on the basis of the regular polygon having a category corresponding to each resource as a vertex, the application within the regular polygon The execution characteristics of each application can be determined according to the distribution of relative resource usage diagrammed in.

Specifically, the scheduling unit may determine that the similarity between execution characteristics of an application program increases as the area where the distribution of relative resource usage overlaps within the regular polygon is larger.

Specifically, the scheduling unit clusters the application programs for each resource when the area where the distribution of the relative resource usage diagrammed in the regular polygon overlaps or exceeds the maximum area or the ratio of applications that are less than the minimum area is greater than or equal to the threshold ratio. You can adjust the number of clusters to do.

Specifically, the scheduling unit, if the ratio of the application where the area where the distribution of relative resource usage schematically within the regular polygon overlaps the maximum area is greater than or equal to the threshold ratio, the number of clusters for clustering the application for each resource Can be adjusted upward.

Specifically, the scheduling unit, when the proportion of applications with an area in which the distribution of relative resource usage diagrammatically overlapped within the regular polygon is less than a minimum area is greater than or equal to a threshold ratio, decreases the number of clusters for clustering applications for each resource. Can be adjusted.

An operation method of a scheduling apparatus according to an embodiment of the present invention for achieving the above object includes: a collection step of collecting performance events, which are information on resources used for execution by each application program for each application program; A classification step of clustering application programs for each resource according to resource usage of each application program identified from the performance event; And a determination step of collecting clustering results from each resource for each application program to determine execution characteristics of each application program.

Specifically, the method may further include a scheduling step of determining similarity between execution characteristics determined for each application program and scheduling concurrent executions between application programs having similarity or less than a threshold.

Specifically, in the classification step, two or more clusters for distinguishing relative resource usage between application programs for each resource may be set, and one application program may be classified into any one of the two or more clusters for each resource.

Specifically, in the determining step, when the clusters according to resource usage are classified according to the length from the center of the regular polygon to each vertex based on a regular polygon having a category corresponding to each resource as a vertex, the regular polygon is applied for each application. The execution characteristics of each application can be determined according to the distribution of the relative resource usage diagrammatically within.

Specifically, in the scheduling step, it can be determined that the larger the area where the distribution of relative resource usage overlaps within the regular polygon, the higher the similarity between execution characteristics of the application program.

Specifically, in the scheduling step, if the area where the distribution of the relative resource usage diagrammed in the regular polygon overlaps each other exceeds the maximum area or the ratio of applications that are less than the minimum area is greater than or equal to the threshold ratio, the application program is applied to each resource. It is possible to adjust the number of clusters for clustering.

Specifically, in the scheduling step, when the proportion of applications in which the distribution of the relative resource usage diagrammed in the regular polygon overlaps the maximum area is greater than or equal to the critical ratio, the clustering for clustering applications for each resource The number can be adjusted upward.

Specifically, in the scheduling step, when the proportion of applications in which the distribution of the relative resource usage diagrammed in the regular polygon overlaps each other is less than the minimum area, the threshold ratio is greater than or equal to the threshold ratio, the number of clusters for clustering applications for each resource is determined. It can be adjusted downward.

Accordingly, in the scheduling apparatus and the operation method of the present invention, in a high performance computing (HPC) environment, by scheduling concurrent executions between applications with little interference between each other based on execution characteristics according to resource usage of the application program In addition, it is possible to achieve an effect of not only optimizing performance of a high-performance computing system, but also improving performance of an application program.

1 is a schematic configuration diagram for explaining a high performance computing environment according to an embodiment of the present invention.
2 is a configuration diagram for explaining the configuration of a scheduling apparatus according to an embodiment of the present invention.
3 is an exemplary view for explaining a performance event collection process according to an embodiment of the present invention.
4 is an exemplary diagram illustrating execution characteristics of each application according to an embodiment of the present invention.
5 and 6 are exemplary views for explaining the similarity of execution characteristics between application programs according to an embodiment of the present invention.
7 is a schematic flowchart for explaining a method of operating a scheduling apparatus according to an embodiment of the present invention.

1 schematically shows a high performance computing environment according to an embodiment of the present invention.

As shown in FIG. 1, a high performance computing environment according to an embodiment of the present invention may include a scheduling device 100 that schedules concurrent execution between application applications.

Scheduling device 100 refers to a device for scheduling concurrent execution between application programs, for example, implemented in the form of a server accessible through a wired or wireless communication network, or in the form of a program in a computer system (eg, computer, mobile phone) Can also be implemented.

When the scheduling device 100 is implemented in the form of a server, for example, it may be implemented in the form of a web server, a database server, a proxy server, etc., a network load balancing mechanism, or a service device operating on the Internet or other network One or more of a variety of software that can be installed can be installed, which can also be implemented as a computerized system.

As described above, in order to optimize system performance in a high performance computing (HPC) environment supporting concurrent execution between application programs, it is necessary to minimize the influence of interference between application programs.

In this regard, in the related art, there is a method of calculating interference between application programs based on a distance between clusters that classifies a single characteristic of an application program by simply using a single cluster analysis result.

However, in the case of the existing technology, since interference between applications is considered based on a single characteristic of the application, there is a limit to sufficiently reflect the effect of interference between individual use resources that cause substantial interference when executing each application. It is reality to follow.

Accordingly, according to an embodiment of the present invention, a new method for scheduling concurrent execution between application programs having little interference with each other is proposed based on execution characteristics according to resource usage of each application program. The specific configuration of the device 100 will be described.

2 shows a configuration of the scheduling device 100 according to an embodiment of the present invention.

As shown in FIG. 2, the scheduling apparatus 100 according to an embodiment of the present invention includes a collection unit 10 for collecting performance events for each application, a classification unit 20 for clustering application programs, and an application program. It may have a configuration including a determining unit 30 for determining the execution characteristics.

In addition, the scheduling apparatus 100 according to an embodiment of the present invention may further include a configuration of the scheduling unit 40 for scheduling concurrent execution between application programs in addition to the above-described configuration.

The entire configuration or at least a portion of the scheduling device 100 including the above-described collection unit 10, classification unit 20, determination unit 30, and scheduling unit 40 is implemented in the form of a hardware module or a software module, or , It can also be implemented in the form of a combination of hardware modules and software modules.

Here, the software module may be understood as, for example, instructions executed by a processor that controls operations within the scheduling apparatus 100, and these instructions may have a form mounted in a memory in the scheduling apparatus 100. .

Meanwhile, the scheduling apparatus 100 according to an embodiment of the present invention may further include a configuration of the communication unit 50, which is an RF module in charge of a communication function for supporting a wired / wireless communication network connection, in addition to the above-described configuration.

Here, the communication unit 50 includes, for example, an antenna system, an RF transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a codec (CODEC) chipset, and memory, but is not limited thereto. Any known circuit to be performed may be included.

As described above, the scheduling apparatus 100 according to an embodiment of the present invention can schedule concurrent executions between applications having little interference with each other based on execution characteristics according to resource use by application through the above-described configuration. Each configuration in the scheduling apparatus 100 for realizing this will be described in more detail.

The collection unit 11 performs a function of collecting performance events for each application.

More specifically, the collection unit 11 collects performance events, which are information on resources used for execution by each application for each application.

At this time, the collection unit 11 may collect performance events for resources used while each application program is running in the system using a hardware performance counter built into the system processor.

For example, referring to FIG. 3, performance events collected for each application may include, for example, 'Compute', a resource category for processor use, 'Memory', a resource category for memory use, and a resource category for disk use. Phosphorus 'I / O' may be included.

The classification unit 20 performs a function of clustering application programs.

More specifically, when the collection of performance events for each application is completed, the classification unit 20 classifies the applications according to resource usage of each application identified from the collected performance events to cluster the applications for each resource. do.

At this time, the classification unit 20 may set two or more clusters for distinguishing the relative resource usage between application programs for each resource, so that one application can be classified into any one of two or more clusters according to the usage amount for each resource. have.

Here, the cluster is, for example, 1 (low cluster), 2 (normal cluster) for each of the resource categories 'Compute', 'Memory', and 'I / O' included in the performance event according to an embodiment of the present invention. , And 3 (high cluster).

The determining unit 30 performs a function of determining an execution characteristic of the application program.

More specifically, when clustering in each resource is completed for each application, the determination unit 30 collects the clustering results of the application for each resource and determines execution characteristics for each application.

At this time, the determining unit 30 is based on a regular polygon having a category corresponding to each resource as a vertex, when classifying clusters according to resource usage according to the length from the center of the regular polygon to each vertex, within the regular polygon for each application program The execution characteristics of each application can be determined according to the distribution of relative resource usage diagrammed in.

That is, the distribution and relative area of relative resource usage within a regular polygon for each application can be regarded as execution characteristics of each application.

In this regard, FIG. 4 exemplarily shows a state in which execution characteristics for each application program are schematically illustrated.

In FIG. 4, each of the equilateral triangles represents each application program in which execution characteristics are determined, and the vertices of each triangle represent a resource category (Compute, Memory, and I / O) according to an embodiment of the present invention. Each of the boundaries from the smallest equilateral triangle to the outermost equilateral triangle sequentially refers to the lower cluster 1, the normal cluster 2, and the higher cluster 3.

The scheduling unit 40 performs a function of scheduling concurrent execution between application programs.

More specifically, the scheduling unit 40 determines the similarity between execution characteristics determined for each application program, and schedules concurrent executions between application programs having similarity or less than a threshold.

In this case, when the regular polygons expressed for each application program overlap each other, the scheduling unit 40 determines the similarity between applications according to the area where the distribution of relative resource use overlaps, and the area where distribution of relative resource use overlaps each other. It can be determined that the wider the degree of similarity between application programs, the greater.

Here, the high degree of similarity between applications means that the resource interference rate is high because the execution characteristics due to resource use are similar to each other, whereas the similarity between applications is low, the execution characteristics due to resource usage are different. Therefore, it means that the mutual resource interference rate is low.

Accordingly, the scheduling unit 40 identifies applications with low resource interference rate (low similarity) according to the similarity determination result between application programs, and schedules concurrent execution in the system for the identified application programs.

For reference, FIG. 5 shows a state in which distributions according to execution characteristics for each application are superimposed on each other, and the table in FIG. 6 shows a resource interference rate between applications calculated when application-specific execution characteristics are overlapped. Giving.

On the other hand, in one embodiment of the present invention, if the area where the distribution of the relative resource usage diagrammed in the regular polygon overlaps each other exceeds the maximum area, or the proportion of applications that are less than the minimum area is greater than or equal to the critical ratio, application programs for each resource A feature configuration that adjusts the number of clusters for clustering can be applied.

According to the result of checking the resource interference rate between applications, the application for each resource is considered in consideration of the case where the ratio of applications that can be concurrently executed is higher than the set value or the ratio of applications that can be concurrently executed is lower than the set value. It can be understood that the number of clusters for clustering is adjusted to adjust the number of concurrently executable applications.

That is, the scheduling unit 40 determines the number of clusters for clustering applications for each resource when the proportion of applications in which the distribution of the relative resource usage diagrammed in the regular polygon overlaps the maximum area is greater than or equal to the threshold ratio. By adjusting upward, the number of applications that can be concurrently executed can be increased.

On the contrary, the scheduling unit 40 lowers the number of clusters for clustering applications for each resource when the ratio of applications in which the distribution of relative resource usage schematically within a regular polygon overlaps with a minimum area is greater than or equal to a threshold ratio. By adjusting, it is possible to reduce the number of applications that can be concurrently executed.

As described above, according to the configuration of the scheduling device 100 according to an embodiment of the present invention, in a high-performance computing (HPC, High Performance Computing) environment, based on the execution characteristics of the application using the resources of each other It can be seen that by scheduling concurrent executions between applications with low interference, it is possible to achieve not only optimization of performance of a high performance computing system, but also improvement of execution performances of applications.

Hereinafter, a description of the operation method of the scheduling apparatus 100 will be continued with reference to FIG. 7.

First, the collection unit 11 collects performance events, which are information on resources used for execution by each application for each application (S10).

At this time, the collection unit 11 may collect performance events for resources used while each application program is running in the system using a hardware performance counter built into the system processor.

3, the performance events collected for each application include, for example, 'Compute', a resource category for processor use, 'Memory', a resource category for memory use, and a resource category for disk use. 'I / O' may be included.

Subsequently, when the collection of performance events for each application is completed, the classification unit 20 classifies the applications according to resource usage of each application identified from the collected performance event and clusters the applications for each resource (S20). ).

At this time, the classification unit 20 may set two or more clusters for distinguishing the relative resource usage between application programs for each resource, so that one application can be classified into any one of two or more clusters according to the usage amount for each resource. have.

Here, the cluster is, for example, 1 (low cluster), 2 (normal cluster) for each of the resource categories 'Compute', 'Memory', and 'I / O' included in the performance event according to an embodiment of the present invention. , And 3 (high cluster).

Furthermore, when clustering in each resource is completed for each application, the determination unit 30 collects the clustering result of the application for each resource and determines execution characteristics for each application (S30-S40).

At this time, the determining unit 30 is based on a regular polygon having a category corresponding to each resource as a vertex, when classifying clusters according to resource usage according to the length from the center of the regular polygon to each vertex, within the regular polygon for each application program The execution characteristics of each application can be determined according to the distribution of relative resource usage diagrammed in.

That is, the distribution and relative area of relative resource usage within a regular polygon for each application can be regarded as execution characteristics of each application.

As illustrated above, FIG. 4 exemplarily shows a state in which execution characteristics for each application program are schematically illustrated.

In FIG. 4, each of the equilateral triangles represents each application program in which execution characteristics are determined, and the vertices of each triangle represent a resource category (Compute, Memory, and I / O) according to an embodiment of the present invention. Each of the boundaries from the smallest equilateral triangle to the outermost equilateral triangle sequentially refers to the lower cluster 1, the normal cluster 2, and the higher cluster 3.

Subsequently, the scheduling unit 40 determines the similarity between execution characteristics determined for each application program, and schedules concurrent executions between application programs having similarity or less than a threshold (S50-S60).

In this case, when the regular polygons expressed for each application program overlap each other, the scheduling unit 40 determines the similarity between applications according to the area where the distribution of relative resource use overlaps, and the area where distribution of relative resource use overlaps each other. It can be determined that the wider the degree of similarity between application programs, the greater.

Here, the high degree of similarity between applications means that the resource interference rate is high because the execution characteristics due to resource use are similar to each other, whereas the similarity between applications is low, the execution characteristics due to resource usage are different. Therefore, it means that the mutual resource interference rate is low.

Accordingly, the scheduling unit 40 identifies applications with low resource interference rate (low similarity) according to the similarity determination result between application programs, and schedules concurrent execution in the system for the identified application programs.

FIG. 5 exemplified above shows a state in which distributions according to execution characteristics for each application are superimposed on each other. In addition, in the table of FIG. 6, resource interference rates between applications are calculated when application-specific execution characteristics are superimposed on each other. Is showing.

On the other hand, in one embodiment of the present invention, if the area where the distribution of the relative resource usage diagrammed in the regular polygon overlaps each other exceeds the maximum area, or the proportion of applications that are less than the minimum area is greater than or equal to the critical ratio, application programs for each resource A feature configuration that adjusts the number of clusters for clustering can be applied.

According to the result of checking the resource interference rate between applications, the application for each resource is considered in consideration of the case where the ratio of applications that can be concurrently executed is higher than the set value or the ratio of applications that can be concurrently executed is lower than the set value. It can be understood that the number of clusters for clustering is adjusted to adjust the number of concurrently executable applications.

That is, the scheduling unit 40 determines the number of clusters for clustering applications for each resource when the proportion of applications in which the distribution of the relative resource usage schematically within the regular polygon overlaps the maximum area is greater than or equal to the threshold ratio. By adjusting upward, the number of applications that can be concurrently executed can be increased.

On the contrary, the scheduling unit 40 lowers the number of clusters for clustering applications for each resource when the ratio of applications in which the distribution of relative resource usage schematically within a regular polygon overlaps with a minimum area is greater than or equal to a threshold ratio. By adjusting, it is possible to reduce the number of applications that can be concurrently executed.

As described above, according to the operating method of the scheduling apparatus 100 according to an embodiment of the present invention, in a high performance computing (HPC, High Performance Computing) environment, based on the execution characteristics of the application using the resource interaction It can be seen that by scheduling concurrent executions between application programs with less inter-interference, it is possible to achieve not only optimization of performance of a high performance computing system, but also improvement of execution performances of application programs.

On the other hand, steps of the method or algorithm described in connection with the embodiments presented herein may be implemented directly in hardware or in the form of program instructions that can be executed through various computer means to be recorded on a computer readable medium. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floptical disks. -Hardware devices specifically configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

The present invention has been described in detail with reference to preferred embodiments, but the present invention is not limited to the above-described embodiments, and the technical field to which the present invention pertains without departing from the gist of the present invention claimed in the claims below. Anyone with ordinary knowledge in the technical idea of the present invention extends to the extent that various modifications or modifications are possible.

According to the scheduling apparatus and its operating method according to the present invention, in a high performance computing (HPC, High Performance Computing) environment, in the point that it is possible to schedule concurrent execution between applications with low resource interference based on the execution characteristics of the application, As it surpasses the limitations of the existing technology, it is not only the use of the related technology, but also the possibility of commercially available or sales of the applied device, as well as the degree to which it can be practiced clearly and in practice.

100: scheduling device
10: collection unit 20: classification unit
30: decision unit 40: scheduling unit

Claims (16)

A classification unit for clustering application programs for each resource according to resource usage of each application; And
It includes a decision unit that collects clustering results from each resource for each application and determines execution characteristics according to resource usage of each application.
The determining unit,
In the case of classifying clusters according to resource usage according to the length from the center of the regular polygon to each vertex based on a regular polygon having a category corresponding to each resource as a vertex, the relative resource usage plotted within the regular polygon for each application program Scheduling device characterized in that to determine the execution characteristics of each application according to the distribution. According to claim 1,
The scheduling device,
And a scheduling unit determining the similarity between execution characteristics determined for each application program and scheduling concurrent executions between application programs having similarity or less than a threshold. According to claim 1,
The classification unit,
Scheduling apparatus characterized in that by setting two or more clusters for distinguishing relative resource usage between application programs for each resource, one application is classified into any one of the two or more clusters for each resource. delete According to claim 2,
The scheduling unit,
Scheduling device characterized in that the larger the area where the distribution of relative resource usage overlaps within the regular polygon, the higher the similarity between execution characteristics of the application program. According to claim 2,
The scheduling unit,
Adjust the number of clusters for clustering applications for each resource when the area where the distribution of relative resource usage schematically within the regular polygon exceeds the maximum area or the proportion of applications that are less than the minimum area is greater than or equal to a threshold ratio. Scheduling device characterized in that to do. The method of claim 6,
The scheduling unit,
When the proportion of applications in which the distribution of the relative resource usage schematically within the regular polygon overlaps the maximum area is greater than or equal to a critical ratio, the number of clusters for clustering applications for each resource is adjusted upward. Scheduling device. The method of claim 6,
The scheduling unit,
When the ratio of applications in which the distribution of relative resource usage schematically plotted within the regular polygon overlaps with a minimum area is greater than or equal to a threshold ratio, the number of clusters for clustering applications for each resource is adjusted downward. Scheduling device. A classification step of clustering application programs for each resource according to resource usage of each application program; And
It includes a decision step of collecting clustering results from each resource for each application and determining execution characteristics according to resource use of each application.
The determining step,
In the case of classifying clusters according to resource usage according to the length from the center of the regular polygon to each vertex based on a regular polygon having a category corresponding to each resource as a vertex, the relative resource usage plotted within the regular polygon for each application program The operating method of the scheduling device, characterized in that to determine the execution characteristics of each application according to the distribution. The method of claim 9,
The above method,
And determining a degree of similarity between execution characteristics determined for each application program and scheduling concurrent executions between applications having a similarity level below a threshold value. The method of claim 9,
The classification step,
A method of operating a scheduling apparatus, characterized in that two or more clusters are set to distinguish relative resource usage between application programs for each resource, and one application is classified into one of the two or more clusters for each resource. delete The method of claim 10,
The scheduling step,
The larger the area where the distribution of relative resource usage overlaps within the regular polygon, the higher the similarity between execution characteristics of the application program is. The method of claim 10,
The scheduling step,
Adjust the number of clusters for clustering applications for each resource when the area where the distribution of relative resource usage schematically within the regular polygon exceeds the maximum area or the proportion of applications that are less than the minimum area is greater than or equal to a threshold ratio. The operation method of the scheduling device, characterized in that to be. The method of claim 14,
The scheduling step,
When the proportion of applications in which the distribution of the relative resource usage schematically within the regular polygon overlaps the maximum area is greater than or equal to a critical ratio, the number of clusters for clustering applications for each resource is adjusted upward. How to operate the scheduling device. The method of claim 14,
The scheduling step,
When the ratio of applications in which the distribution of relative resource usage schematically plotted within the regular polygon overlaps with a minimum area is greater than or equal to a threshold ratio, the number of clusters for clustering applications for each resource is adjusted downward. How the scheduling device works.

Images