KR-20260064521-A - METHOD AND SYSTEM FOR EFFICIENT POWER SCHEDULING IN A COMPUTING SYSTEM

KR20260064521AKR 20260064521 AKR20260064521 AKR 20260064521AKR-20260064521-A

Abstract

The present disclosure relates to an efficient power scheduling method in a computational system performed by at least one processor, comprising: a step of assigning a task to a specific computational unit among a plurality of computational units included in the computational system upon a task request from a user; a step of calculating an expected power consumption resulting from performing the assigned task through the specific computational unit; and a step of transferring the assigned task to another computational unit if, based on the calculated expected power consumption, it is determined that the power consumption of the specific computational unit or an upper computational layer including the specific computational unit will exceed a reference power consumption.

Inventors

조강원
박정호
정우근

Assignees

주식회사 모레

Dates

Publication Date: 20260507
Application Date: 20251013
Priority Date: 20241030

Claims (12)

In an efficient power scheduling method in a computational system performed by at least one processor, A step of assigning a task to a specific operation unit among a plurality of operation units included in the above-described operation system as a task is requested by a user; A step of calculating the expected power consumption resulting from performing the assigned task through the specific operation unit; and Based on the estimated power consumption calculated above, if it is determined that the power consumption of the specific computational unit or the upper computational layer including the specific computational unit will exceed the reference power consumption, the step of transferring the assigned task to another computational unit. An efficient power scheduling method in a computing system including
In paragraph 1, The step of calculating the above-mentioned estimated power consumption is, A step of generating input data based on the attributes of the above-mentioned assigned task; and A step of deriving the estimated power consumption for the assigned task as result data by inputting the generated input data into a previously trained artificial intelligence model. Includes, The attributes of the above operation are, At least one of the above-mentioned type of operation, operation graph, input data size, and batch size An efficient power scheduling method in a computing system including
In paragraph 1, The step of calculating the above-mentioned estimated power consumption is, A step of calculating the expected power consumption for the assigned task through a statistical analysis of the average power consumption for tasks identical to the assigned task over a past predetermined period and the power consumption for computational units among the plurality of computational units that are performing tasks identical to the assigned task. An efficient power scheduling method in a computing system including
In paragraph 1, The step of calculating the above-mentioned estimated power consumption is, A step of calculating the estimated power consumption corresponding to the above-mentioned assigned task based on a power consumption calculation model. Includes, The above power consumption calculation model is, An efficient power scheduling method in a computing system, defined based on information regarding the attributes of multiple tasks executed during a predetermined past period and the power consumption corresponding to each of the multiple tasks, stored in a power consumption database, and a mathematical formula model that mathematically expresses the correlation between the attributes of the tasks and the power consumption.
In paragraph 1, The step of transferring the above-mentioned assigned task to another operation unit is, Based on the estimated power consumption calculated above, if it is determined that the power consumption of the specific computational unit will exceed a first reference power consumption set corresponding to the specific computational unit, the step of transferring the assigned task to another computational unit placed in the same rack as the specific computational unit. An efficient power scheduling method in a computing system including
In paragraph 1, The step of transferring the above-mentioned assigned task to another operation unit is, Based on the estimated power consumption calculated above, if it is determined that the power consumption of a specific rack containing the specific computing unit exceeds a second reference power consumption set corresponding to the specific rack, the assigned task is transferred to another computing unit included in another rack placed on the same line as the specific rack. An efficient power scheduling method in a computing system including
In paragraph 1, The step of transferring the above-mentioned assigned task to another operation unit is, Based on the estimated power consumption calculated above, if it is determined that the power consumption of a specific line in which a rack including the specific computing unit is placed will exceed a third reference power consumption set corresponding to the specific line, the assigned task is transferred to another computing unit included in a line adjacent to the specific line. including, Efficient power scheduling method in computing systems.
In paragraph 1, The step of transferring the above-mentioned assigned task to another operation unit is, If it is determined that the power consumption of the specific operation unit or the upper operation layer including the specific operation unit will exceed the reference power consumption, the assigned task is transferred to another operation unit, If there is no computational unit to transfer the above-mentioned assigned task, the step of reducing the computational load for the above-mentioned assigned task or delaying the processing of the above-mentioned assigned task. including, Efficient power scheduling method in computing systems.
In paragraph 1, The step of transferring the above-mentioned assigned task to another operation unit is, A step of transferring the assigned task to another computational unit when, based on the calculated estimated power consumption, the number of times the power consumption of the specific computational unit or the upper computational layer including the specific computational unit exceeds the reference power consumption exceeds a preset number, or when, based on the calculated estimated power consumption, the time during which the power consumption of the specific computational unit or the upper computational layer including the specific computational unit exceeds the reference power consumption exceeds a preset time. An efficient power scheduling method in a computing system including
In paragraph 1, The above assigned task is, It is represented as a computation graph composed of multiple computation nodes and data dependency relationships between said computation nodes, and The step of calculating the above-mentioned estimated power consumption is, A step of individually calculating the expected power consumption for each of the above-mentioned plurality of operation nodes or one or more sub-operation graphs including two or more operation nodes. Includes, The step of transferring the above-mentioned assigned task to another operation unit is, If it is determined that the power consumption of the specific operation unit or the upper operation layer including the specific operation unit will exceed the reference power consumption, the step of transferring at least one operation node among the plurality of operation nodes or the one or more sub-operation graphs to another operation unit, reducing the amount of operation, or delaying processing. An efficient power scheduling method in a computing system including
A computer program stored on a computer-readable recording medium for executing a method according to any one of paragraphs 1 through 10 on a computer.
As an information processing system, Memory; and At least one processor connected to the memory and configured to execute at least one computer-readable program contained in the memory. Includes, The above at least one program is, As a task is requested by a user, the task is assigned to a specific computational unit among multiple computational units included in the computational system, and Calculate the expected power consumption resulting from performing the assigned task through the aforementioned specific operation unit, and An information processing system comprising instructions for transferring the assigned task to another computational unit when, based on the estimated power consumption calculated above, it is determined that the power consumption of the specific computational unit or the upper computational layer including the specific computational unit will exceed a reference power consumption.

Description

Method and System for Efficient Power Scheduling in a Computing System The present disclosure relates to an efficient power scheduling method and system in a computing system, and more specifically, to an efficient power scheduling method and system in a computing system that enables the efficient use of power, a limited resource, in a computing system operating multiple computing units. In data center or cluster environments that support high-performance computing, power consumption acts as one of the biggest constraints. Generally, when configuring a cluster by deploying multiple computing devices such as servers or GPUs, the total power capacity is calculated based on the maximum power consumption of each device, and the number of devices that can be installed is limited accordingly. For example, if the maximum power consumption of a single GPU server is 4kW and the total power or cooling capacity of the data center room is 40kW, only up to 10 units can be installed. While this approach has the advantage of guaranteeing facility safety and stable power supply, it presents a problem in that there is a significant discrepancy with the power consumption patterns that occur during actual operation. In real-world environments, it is rare for all devices to operate at maximum power at all times. Most computing units experience load fluctuations depending on the situation; some devices consume significant power under high loads, while others consume less power under relatively low loads. Therefore, calculating power capacity based solely on the maximum values of all devices results in an overly conservative design relative to actual power consumption. This reduces the number of devices that can be installed within the same infrastructure, lowering the efficiency of power resource utilization and ultimately leading to increased data center operating costs and resource waste. To mitigate these problems, various methods have been proposed in conventional technologies, such as power capacity management techniques, resource monitoring systems, and setting power caps per device. For instance, methods have been utilized to assign power upper limits to each computing unit or to control loads to prevent them from exceeding a certain level. Additionally, attempts have been made to monitor the power consumption of the entire cluster and delay or limit new tasks when power thresholds are approached. However, these techniques have limitations in adequately reflecting dynamic changes in power usage or resolving issues of unbalanced load distribution among devices. In particular, existing methods often limited themselves to simple restrictions at the individual device level or blanket constraints at the entire cluster level, failing to adequately consider interactions between devices or actual operating patterns. As a result, inefficiencies occurred where some devices were overloaded while others remained idle. Furthermore, the design practice of calculating power capacity based solely on maximum values affected peripheral infrastructure resources, such as cooling capacity, making it difficult to optimize resource utilization at the overall system level. As such, power management issues in cluster and data center environments are recognized as critical challenges that must be addressed to ensure the efficient utilization of power resources, reduce operating costs, and provide stable services, and various studies to improve this are continuously being conducted. The aforementioned background technology is one that the inventor possessed or acquired in the process of deriving the contents of the disclosure of the present application, and it cannot be considered as prior art disclosed to the general public prior to the filing of this application. Embodiments of the present disclosure will be described with reference to the accompanying drawings described below, wherein similar reference numerals indicate similar elements, but are not limited thereto. FIG. 1 is a diagram showing examples of a main process and a sub-process for efficient power scheduling in a computational system according to one embodiment of the present disclosure. FIG. 2 is a block diagram showing the internal configuration of an information processing system according to one embodiment of the present disclosure. FIG. 3 is a flowchart of an efficient power scheduling method in a computational system according to one embodiment of the present disclosure. FIG. 4 is a flowchart of a method for dynamically transferring tasks by considering power consumption by computational hierarchy, such as computational units, racks, and lines, according to one embodiment of the present disclosure. FIG. 5 is a flowchart of a method for transferring work depending on the presence or absence of a computational unit capable of transferring work according to one embodiment of the present disclosure. Hereinafter, specific details for implementing the present disclosure will be descri