KR20220046812A - Reward-oriented task offloading under limited edge server power for mobile edge computing - Google Patents

Abstract

Disclosed is an offloading task to maximize a reward under an edge server limit power in a mobile edge computing. An offloading task method performed by a computer-implemented offloading task system according to one embodiment may comprise: a step of determining an edge server to execute an offloaded task based on the power information of each edge server; and a step of offloading the task according to the allocation information set in the determined edge server.

Description

{REWARD-ORIENTED TASK OFFLOADING UNDER LIMITED EDGE SERVER POWER FOR MOBILE EDGE COMPUTING}

The description below relates to techniques for offloading tasks from mobile edge computing to edge servers.

Advances in network and system technologies have increased the feasibility of applications that depend on wireless networks. Augmented reality and smart traffic control systems based on connected Internet of Things (IoT) devices such as sensors, smart personal devices, and driverless vehicles are representative examples. Smart IoT devices are growing rapidly, and it is predicted that by 2030, 500 billion smart IoT devices will be needed in smart city applications. These applications will have to deal with the vast amount of data that will be generated by smart IoT devices.

Some IoT devices need to perform tasks that require fast computation. It also entails regular collection of sensor data that must be processed regularly. When computing capabilities are limited, these tasks may need to be offloaded to cloud servers. However, the latency imposed by the core network around a central server in the cloud means that a cloud-based approach may not be appropriate. For IoT-based systems that require rapid processing, we solve problems through edge computing services. The edge server (ES) located at the edge of the cloud produces the effect of reducing the load on the central server along with fast processing, so the demand for edge computing is expected to increase rapidly. Edge computing service can be implemented by placing an edge server at each base station (BS) or connecting the edge server to multiple base stations. This is called mobile edge computing (MEC). IoT devices located within range of these base stations can offload or access tasks from the base stations to edge servers.

Edge Infrastructure Providers (EIPs) have started offering edge computing services to their customers. EIP installs edge servers at the edge of the cloud and uses them to deliver computing power to customers. Improve edge server shared resource utilization among customers and get financial return on EIP. The priority of operation on the server is determined according to the contractual agreement between EIP and the customer. It is necessary to determine which tasks are offloaded within the context of maximizing the total reward for EIP.

The edge server is a remote micro data center, and the power consumption is expected to be 10kW. Thousands of such servers will consume megawatts, a major concern for EIP, so power limits are essential. The power drawn from the edge server remains below the predefined power limit. The power consumed by the server depends on the CPU utilization. This is determined by the workload of the server, which means it is determined by the number of offloaded operations that the edge server accepts.

Accordingly, there is a need for techniques to reduce the energy used by mobile devices using task offloading techniques.

It is possible to provide a method and system for maximizing the rewards associated with the benefits accrued while limiting the power consumed by an edge server. In detail, the maximum allowable CPU utilization rate of each edge server is based on the maximum allowable CPU utilization rate of each edge server set by examining the power characteristics of the edge server to formulate the optimization problem, and determining the edge server to execute each offloaded task It is possible to provide a method and system for estimating a reward in consideration of the power consumption associated with

A task offloading method performed by a computer-implemented task offloading system includes the steps of: determining an edge server to execute an offloaded task based on power information of each edge server; and offloading the job according to the determined assignment information set in the edge server.

The step of offloading the task comprises: determining allocation information including the maximum allowable utilization rate of each edge server and task allocation for maximizing the compensation of the edge server provider under the power limit of each edge server. may include

The step of offloading the job may include determining the maximum allowable utilization rate of each edge server while giving high priority to offloading jobs with a high compensation ratio of the edge server provider for the power demand. can

The step of offloading the task may include searching for values of a maximum allowable utilization rate while satisfying a power constraint of each edge server.

The step of offloading the task may include iteratively calculating the power increment and the reward increment from the task assignment, giving a higher reward to the power ratio until the power constraint is satisfied and the maximum allowable usage rate is determined at each edge server. and prioritizing the maximum allowable utilization value.

The step of offloading the job may include searching for a value for each job for allocating the job to each edge server.

The step of offloading the job includes the step of performing job assignment to the edge server using either a method that allows job partitioning using a minimum cost maximum flow graph or a method that does not allow job partitioning. may include

The method for allowing the work partitioning is a process in which a residual graph is generated from the minimum cost maximum flow graph and the value and cost of a binary function to represent the flow between each node is initialized, using a shortest path and fastest algorithm to minimize the cost. When a flow exists between vertices and a path augmentation process that creates a path in the residual graph as a flow is derived and updates the flow, an edge server is selected to process one of the sub-tasks constituting the job. The finishing process of updating the index value of the edge server to which the job is assigned can be performed.

The method that does not allow partitioning of work operates on the minimum cost maximum flow graph, a minimum cost flow process that computes a minimum cost flow from a source vertex to a sink vertex, removes the work vertex associated with the least cost flow and uses capacity usage The process of updating the index value of the edge server to which the task is assigned and removing the edges associated with the flow can be performed.

The computer-implemented job offloading system includes: a server determining unit configured to determine an edge server to execute an offloaded job based on power information of each edge server; and a job offloading unit for offloading a job according to the determined assignment information set in the edge server.

The present invention determines the maximum utilization rate of each server, while giving high priority to offloading tasks with a high compensation ratio of the edge server provider to the power demand, and modeling the task allocation problem using the minimum cost maximum flow graph. By retrieving the highest reward and offloading tasks that are affected by server capacity, higher rewards can be achieved than alternative plans under the same power constraints.

1 is a diagram for explaining a mobile edge computing structure according to an embodiment.
2 is a block diagram illustrating the configuration of a work offloading system according to an embodiment.
3 is a flowchart illustrating a task offloading method for maximizing compensation under an edge server limit power in mobile edge computing of a task offloading system according to an embodiment.
4 is a diagram illustrating a pseudo code for a maximum allowable utilization determination (MUD) algorithm in a work offloading system according to an embodiment.
5 is an example for explaining modeling an edge server allocation problem as a flow network in the task offloading system according to an embodiment.
6 is an example of a residual graph for representing a flow between tasks in a task offloading system according to an embodiment.
7 is a diagram illustrating a pseudo code for an edge server allocation algorithm when job division is permitted in the job offloading system according to an embodiment.
8 is a diagram illustrating a pseudo code for an edge assignment algorithm without job assignment in a job offloading system according to an embodiment.
9 to 14 are graphs for explaining the experimental results of the work offloading system according to an embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

In mobile edge computing, jobs are offloaded from mobile devices to servers at the edge of the network, speeding up job processing without incurring the latency required to reach a central server. However, the power used by the edge server is significant and needs to be cost-effective. The embodiment describes a method of offloading tasks to a server for the purpose of maximizing the compensation of an edge server provider within limited power consumption, server processing capacity, and wireless network coverage.

1 is a diagram for explaining a mobile edge computing structure according to an embodiment.

Edge servers may be assigned to at least one or more base stations to handle tasks offloaded from IoT devices within the coverage of the base stations. If a job is offloaded from an edge server to an EIP (referred to as an 'edge server provider'), that edge server provider may be rewarded based on the amount of computation performed by the edge server. Otherwise, offloaded tasks are handed over to a central server in the cloud.

Meanwhile, the SPEC Power Committee provides tools and standards for measuring the energy consumption of edge servers. Energy measurements for edge servers have been collected since 2006 and the relationship between energy consumption and CPU utilization has been analyzed, and it has been found that this is strongly influenced by server characteristics.

Many IoT applications such as surveillance, air quality monitoring and traffic monitoring require periodic collection of sensor data that must be processed periodically. For job offloading, each job may calculate the amount of computation required for a certain time and receive a certain percentage of the processor capacity provided by the edge server. Job usage can be expressed as the numerator of the amount of calculation required for the job, and the total amount of calculation provided by the edge server for a certain period of time as the denominator. Each task can then use the processor capacity allocated for each cycle.

(i=1, ..., NES) 를 I^{h igh} 의 지수를 가진 최고 처리용량을 갖는 엣지 서버와 비교하여 최대 처리용량을 나타내기 위해 사용할 수 있다. 그 후, 작업

의 정규화 처리된 U_i(i=1, ..., N^task)는 엣지 서버 I^{h igh} 에서 실행된다는 사실을 기반으로 결정될 수 있다. 여기서, N^task는 총 작업 수를 나타낸다. 쉽게 설명하기 위하여, U_i와

가 정수 값을 갖는다고 가정하자. 이를 위해, 그레인 크기를 최소 할당 단위로 도입하고

와 U_i를 그레인의 배수로 표현할 수 있다. 예를 들면, 만약, 그레인의 크기가 0.01이면,

=100이 된다. Because different edge servers may have different processing capacities, the usage of each job can be normalized to the edge server with the highest processing capacity. If NES is the number of edge servers,

(i=1, ..., NES) can be used to represent the maximum throughput by comparing it with the edge server with the highest throughput with an exponent of I ^{h igh} . After that, work

The normalized U _i (i=1, ..., N ^task ) of can be determined based on the fact that it is executed on the edge server I ^{h igh} . Here, N ^task represents the total number of tasks. For ease of explanation, U _i and

Assume that is an integer value. For this, we introduce the grain size as the smallest allocation unit and

and U _i can be expressed as multiples of grain. For example, if the grain size is 0.01,

= 100.

,

를 작업

( i=1,..., N^task)가 할당된 엣지 서버의 지수라고 하자. 어떠한 엣지 서버로도 할당되지 못하는 작업

을 처리하는 클라우드의 중앙 서버를 나타내기 위해 지수 0을 도입할 수 있다. 요약하자면, 만약

=0이면, 작업

는 클라우드 중앙 서버에 의해 처리될 수 있다. 반면에, 만약

=j이면, 작업

는 엣지 서버 j에 의해 처리될 수 있다. 모든 작업에 대하여 합쳐친 총 사용량은 처리용량을 초과할 수 없으므로

가 된다. 이에 따라, 엣지 서버에서의 총 CPU 사용률

은 다음과 같이 표현될 수 있다.

,

work

Let ( i=1,..., N ^task ) be the index of the assigned edge server. Tasks that cannot be assigned to any edge server

An exponent of 0 can be introduced to indicate a central server in the cloud that handles In summary, if

=0, work

can be processed by a cloud central server. On the other hand, if

If =j, work

can be processed by edge server j. The total usage for all jobs cannot exceed the processing capacity.

becomes Accordingly, the total CPU utilization on the edge server

can be expressed as

Equation 1:

사이의 정수이다. 이때, 각 엣지 서버 j에서 총 CPU 사용률 Y_j을 초과해서는 안된다. 만약, CPU 사용률 Y_j가 0이면, 각 엣지 서버 j는 어떠한 작업도 처리할 수 없다. 만약, Y_j=

라면, 각 엣지 서버 j의 처리용량을 완전히 사용할 수 있다.In order to limit the amount of power consumed by each edge server, CPU usage may be limited. To this end, a maximum allowable CPU utilization Y _j can be introduced for each edge server j. Here, CPU utilization Y _j is 0 and

is an integer between At this time, the total CPU usage rate Y _j in each edge server j must not be exceeded. If the CPU usage rate Y _j is 0, each edge server j cannot process any tasks. If Y _j =

If so, the processing capacity of each edge server j can be fully used.

일 때, 활성 전력 계수를 표현하기 위해

를 사용하고 각 엣지 서버 j의 유휴 전력 계수를 표현하기 위해

를 사용할 수 있다. 이에, 각 엣지 서버 j에 의해 도출된 전력

를 사용률 x에 따라 다음과 같이 정의될 수 있다.Active power is the power consumed while the edge server is performing computational tasks, and it varies non-linearly with CPU utilization. Even when computation is not in progress, the edge server uses fixed idle power. Accordingly, the utilization rate on each edge server j is x

When , to express the active power factor

and to express the idle power factor of each edge server j

can be used Accordingly, the power derived by each edge server j

can be defined as follows according to the usage rate x.

Equation 2:

2 is a block diagram for explaining the configuration of a task offloading system according to an embodiment, and FIG. 3 is a task off maximizing compensation under the edge server limit power in mobile edge computing of the task offloading system according to an embodiment. It is a flowchart for explaining the loading method.

The processor of the task offloading system 100 may include a server determining unit 210 and a task offloading unit 220 . These processor components may be representations of different functions performed by the processor according to control instructions provided by program code stored in the task offloading system. The processor and components of the processor may control the task offloading system to perform steps 310 to 320 included in the task offloading method for maximizing compensation under the edge server limit power in the mobile edge computing of FIG. 3 . In this case, the processor and the components of the processor may be implemented to execute instructions according to the code of the operating system included in the memory and the code of at least one program.

The processor may load into memory the program code stored in the program's file for a task offloading method that maximizes compensation under edge server limit power in mobile edge computing. For example, when a program is executed in the task offloading system 100 , the processor may control the task offloading system 100 to load a program code from a file of the program into the memory according to the control of the operating system. At this time, each of the processor and the server determining unit 210 and the task offloading unit 220 included in the processor executes the instruction of the corresponding part of the program code loaded in the memory to execute the subsequent steps 310 to 320 . may be different functional representations of the processor for

The server determining unit 210 may determine an edge server to execute an offloaded job based on power information of each edge server ( 310 ), and the job offloading unit 220 may be configured according to allocation information set in the determined edge server. The task may be offloaded ( 320 ).

In the embodiment, each task may be assigned to an appropriate edge server in order to maximize the overall reward under the three constraints described below.

Edge Server Processing Capacity: The total usage of each edge server must not exceed the maximum allowed usage.

Equation 3:

를 오프로드 하는 요청을 받아들일 수 있다. 만약, H _i,j =0이라면, 각 엣지 서버 j는 작업

를 오프로드 하는 요청을 받아들일 수 없다. 이에 따라, 작업

은 H _i,j =1일 경우에만, 엣지 서버 X로 할당될 수 있다.Edge Server Coverage: The edge server can only receive jobs from devices that exist within the edge server area. You can introduce a binary constant H _i,j that indicates whether the operation can be accepted by the edge server. If H _i,j = 1, each edge server j

can accept requests to offload If H _i,j = 0, each edge server j

Could not accept request to offload Accordingly, work

can be assigned to the edge server X only when H _i,j =1.

Power: P ^{l imit} is the maximum total power of all edge servers allowed to be used.

Equation 4:

을 처리할 때, 엣지 서버 제공자에 의해 받은 보상인 R _i,j 를 도입하기로 한다. 그리고 나서, 엣지 서버 할당 문제와 최대 허용 사용량 결정 문제를 공식화할 수 있다. 이는, 모든 엣지 서버의 보상들의 합계를 극대화하기 위한 목적으로 각각의 작업

에서의 값

, (

=0, ??, N^ES)를 찾는 것과 모든 엣지 서버 j에서의 최대 허용 가능한 사용률

, (

=0, ??,

)를 찾는 것으로 구성될 수 있다. Therefore, work on each edge server

When processing , we will introduce R _i,j , which is the reward received by the edge server provider. Then, we can formulate the problem of allocating edge servers and determining the maximum allowable usage. This is done for the purpose of maximizing the sum of all edge server rewards.

value at

, (

=0, ??, N ^ES ) and the maximum allowable utilization on all edge servers j

, (

=0, ??,

) can consist of finding

Referring to FIG. 4 , a diagram illustrating a pseudo code for a maximum allowable utilization determination (MUD) algorithm.

에 대한 값

과 각각의 엣지 서버 j에 대한 최대 허용 가능한 사용률

을 결정해야 한다. 각각의 엣지 서버 j는

처리용량을 갖고 있으며,

와

의 모든 조합들을 고려하는 것은 비실용적이다. 이에 따라 이를 2개의 하위 문제로 나눈다. 사용률 결정 문제(UDP: utilization determination problem,)는 에너지 제약조건을 충족하면서 최대 허용 가능한 사용률

의 값들을 찾는 것이다. 엣지 서버 할당 문제(EAP: edge server allocation problem,)는 각각의 작업

을 엣지 서버로 할당하는

값들을 찾는 것이다.The task offloading system may determine the maximum allowable utilization rate. To solve the optimization problem, each task

value for

and maximum allowable utilization for each edge server j

have to decide Each edge server j is

has a processing capacity,

Wow

It is impractical to consider all combinations of Accordingly, we divide it into two subproblems. The utilization determination problem (UDP) is the maximum allowable utilization rate while satisfying the energy constraint.

to find the values of The edge server allocation problem (EAP) is

to assign as an edge server

to find values.

The problem of determining the utilization rate will be described.

과 그에 해당하는 에너지 요구조건들이 제공된다면, 사용률 결정 문제는 다중 선택 배낭 문제(multiple-choice knapsack problem, MKCP)로 축소될 수 있다. 과부하 없이 아이템들을 배낭에 넣었을 때 얻는 총 이익을 극대화시키기 위해서, 다수의 아이템들을 포함하는 몇몇 클래스들의 각각에 대해 가중치와 이익이 있는 하나의 아이템을 선택해야 한다.Possible rewards on edge server j

Provided and the corresponding energy requirements are provided, the problem of determining the utilization rate can be reduced to the multiple-choice knapsack problem (MKCP). In order to maximize the total benefit of putting items in the backpack without overloading them, you should choose one item with a weight and a benefit for each of several classes that contain multiple items.

사용률 값을 각각 갖고 있는 엣지 서버이며, 각 사용률 값은 총합이

를 초과할 수 없는 전력과 극대화 되어야 하는 보상과 연관되어 있다. MCKP는 NP-난해 이다. 하지만 이익-가중치 비율에 기반을 둔 그리디 알고리즘(greedy algorithm)은 보통 잘 수행된다. 사용률 결정 문제는 작업 할당으로부터 전력과 보상 증가량을 반복적으로 계산하여 부등식 (4)에서의 전력 제약조건이 충족되고 각각의 엣지 서버 j에서 최대 허용 가능한 사용률

이 결정될 때까지 전력 비율에 더 높은 보상을 주는 사용률 값에 우선순위를 부여할 수 있다.In the utilization determination problem, the class is

Each edge server has a utilization value, and each utilization value is the sum of

It is related to the power that cannot be exceeded and the compensation that must be maximized. MCKP is NP-complex. However, greedy algorithms based on profit-weight ratios usually perform well. The utilization determination problem is to iteratively calculate the power and compensation increments from the task allocation so that the power constraint in inequality (4) is satisfied and the maximum allowable utilization rate at each edge server j

Priority may be given to utilization values that give higher rewards to power ratios until this is determined.

Next, the edge server allocation problem will be described.

는

개의 하위 작업

으로 분할될 수 있다. 작업 오프로딩 시스템은 각각의 하위 작업의 CPU 사용이 정수 값을 갖도록 조건화 하였고(작업의 사용률 값과 마찬가지로), 따라서 작업

가 분할될 수 있는 최대 하위 작업 수는

이다.

(

)이 하위 작업

이 할당된 엣지 서버의 지수라고 하자. 최소 비용 최대 흐름 그래프 그래프에서 최적의 정수 기반 흐름 집합을 찾을 수 있고, 이는 최대 가능한 흐름 전달 비용을 최소화한다. 이러한 흐름들에 부합하는

값들을 결정하기 위해 이 알고리즘을 사용할 수 있다.A job offloading system can model the assignment of jobs to an edge server using a minimum cost maximum flow graph (MCMF). In an embodiment, two possible edge allocation algorithms can be considered. Algorithms that allow task partitioning and those that do not. If task partitioning is allowed, the created subtasks can be assigned to different edge servers. work through modeling

Is

sub-tasks

can be divided into The task offloading system conditioned the CPU usage of each subtask to have an integer value (just like the task's utilization value), so the task

The maximum number of subtasks that can be partitioned is

am.

(

) this subtask

Let this be the index of allotted edge servers. Minimum Cost Maximum Flow Graph An optimal integer-based flow set can be found in the graph, which minimizes the maximum possible flow delivery cost. in line with these trends

You can use this algorithm to determine the values.

When task splitting is not allowed and each task has to be executed on a single edge server, EAP becomes the multiple knapsack problem (MKP), in which items with different weights and profits must be assigned to knapsacks with limited capacity in order to maximize profits. can be reduced In the multiple backpack problem, items are tasks, and the purpose of the edge server allocation problem is to maximize the total reward based on processing capacity on each edge server. Like the multiple-choice knapsack problem (MKCP), the multiple knapsack problem (MKP) is NP-hard. The job offloading system can reuse a greedy algorithm that allocates jobs to edge servers based on the maximum reward value until no processing capacity is left.

Next, an algorithm for determining the maximum allowable usage rate will be described.

값을 결정해야 한다. 먼저 알고리즘이 진행되는 동안

의 값들을 저장하기 위해 각 작업

에서 일시적인 변수

를 도입한다. 그리고

는 엣지 서버 j에 의해 도출된 전력 증가량을 나타내고, 새로운 작업

가 여기에 할당될 때 이는 다음과 같이 표현된다. requires 3 new variables

value must be determined. First, while the algorithm is running

Each operation to store the values of

temporary variable in

introduce And

represents the power increase derived by edge server j, and

When is assigned here, it is expressed as

Equation 5:

는 작업

이 각 엣지 서버 j에게 할당되기 전에, 엣지 서버 j에 의해 도출된 전력이고,

는 그 후에 엣지 서버 j에 의해 도출된 전력이다.From here

work

This is the power drawn by the edge server j before being allocated to each edge server j,

is then the power derived by edge server j.

는 작업

을 엣지 서버 j에게 할당함으로써 발생하는 보상의 증가량과 수학식 5의 증가량 간의 비율을 나타내는 특정 보상이다. 이 특정 보상은 다음과 같이 표현될 수 있다.

work

It is a specific reward representing the ratio between the increase amount of the reward generated by allocating to the edge server j and the increase amount of Equation 5. This particular reward can be expressed as

Equation 6:

과 특정 보상

을 계산한다. 그리고 나서 i=M이고j=H 일 때의 최대값

을 결정한다. 그리고 임시 변수

를 1로 설정하여

을 할당된 것으로 표시하고 총 전력 요구량

를 갱신한다. 이 과정은 최대 전력 조건이 충족되지 않을 때 혹은 작업이 더 이상 남지 않을 때까지 반복된다. The purpose of the maximum allowable utilization determination (MUD) algorithm of FIG. 4 is to obtain a large reward with a small amount of power consumption. The algorithm for determining the maximum allowable utilization rate is the amount of power required at each i and j.

and specific rewards

to calculate Then the maximum value when i=M and j=H

to decide and temporary variables

set to 1

mark as allocated and total power demand

update This process is repeated until either the maximum power condition is not met or there are no more jobs left.

5 is an example for explaining modeling an edge server allocation problem as a flow network in the task offloading system according to an embodiment.

Edge server allocation will be described.

The task offloading system can create a Minimum Cost Maximum Flow Graph (MCMF). The edge server allocation problem (EAP) can be effectively expressed as a minimum cost maximum flow problem. This can be explained by creating a max cost max flow graph, which is a flow network with source, sink and relay nodes directly connected to the edge. A source node has only an outward-facing edge, and a sink node has only an inward-facing edge. The edge from node a to node b has capacity and cost. The problem with the min-cost max flow graph is to find the set of flows from the source node to the sink node that maximizes the total flow at a low cost.

가 작업

에 상응하는 작업 정점

가 엣지 서버 j에 상응하는 엣지 서버의 정점

,

가 엣지 서버 j에 상응하는 엣지 서버의 정점

, 그리고 단일 싱크 정점 t이다. 이제, 엣지 서버 j에서 사용된 처리용량 단위당 보상을 다음과 같이 정의할 수 있다.As an example, the edge server allocation problem can be modeled as a flow network having the following four types of vertices as shown in FIG. 5 . single source vertex

autumn work

corresponding work vertex

is the vertex of the edge server corresponding to edge server j

,

is the vertex of the edge server corresponding to edge server j

, and a single sink vertex t. Now, the reward per unit of processing capacity used in edge server j can be defined as follows.

In this case, the rules for creating a flow network can be described as follows.

)는 용량

과 비용 0을 가진다.The source vertex s is connected to all working vertices. Each edge (

) is the capacity

and cost 0.

, (i=1,...,N^task)는

=1일 때, 정점

에 연결된다. 각 엣지는 용량

와 비용

을 가진다.each vertex

, (i=1,...,N ^task ) is

When =1, the vertex

is connected to Each edge has a capacity

and cost

have

, (j=1,...,N^ES) 는 싱크 정점 t에 연결된다. 각 엣지 (

)는 용량

와 비용 0을 가진다.each vertex

, (j=1,...,N ^ES ) is connected to the sink vertex t. each edge (

) is the capacity

and cost 0.

을 결정하고, 작업 분할이 허용되지 않을 때는

를 결정하여 해결할 수 있다. 총 보상을 극대화시키는 것은 총 비용

을 최소화시키는 것에 해당된다.The edge server allocation problem aims to maximize the total reward while satisfying the usage conditions on each edge server, and when division of work is allowed, the value

to decide, and when division of work is not allowed

can be resolved by determining Maximizing the total reward is the total cost

is to minimize the

에 대한 흐름 용량이

라면, 각 엣지 서버의 정점

에서 싱크 정점으로의 총 흐름은 엣지 서버 j의 용량 제약조건인

를 초과할 수 없다.

=1인 경우에만

와

사이의 엣지가 존재하기 때문에 엣지 서버 커버리지 제약조건은 최소 비용 최대 흐름 그래프에 표현될 수 있다.If the operation from the source node to the sink node is

flow capacity for

Ramen, the vertex of each edge server

The total flow from to the sink vertex is the capacity constraint of edge server j.

cannot exceed

only if =1

Wow

Because there is an edge between them, the edge server coverage constraint can be expressed in the minimum cost maximum flow graph.

7 is a diagram illustrating a pseudo code for an edge server allocation algorithm when job division is permitted in the job offloading system according to an embodiment.

The Edge Server Allocation Algorithm (EAA-TS) when task partitioning is allowed is presented in FIG. 7 . EAA-TS finds a path from source to synchro at a reduced cost, which also uses the shortest path faster algorithm (SPFA) to satisfy the capacity condition, which is an improved form of the well-known Bellman-Ford algorithm. .

In the embodiment, a binary function f(a, b) is introduced to express the flow between vertices (nodes) a and b. If the value of f(a, b) is 1, then a flow exists. The shortest path fast algorithm requires a reverse path, since the path chosen at the time of algorithm execution is not part of the final flow, which can be reversed. Therefore, it is necessary to build a residual graph where each edge represents an allowable additional flow between two vertices and a reverse path is allowed.

에 대한 용량 값은 1만큼 감소했지만, 역방향 경로의 엣지들의 용량은 1만큼 증가했다. 이 역방향 경로

에 대한 비용

은

로 설정된다. 역방향 경로는 알고리즘에 의해 선택될 수 있고, 이 경우 최초의 경로 선정은 취소된다. 최단 경로 빠른 알고리즘은 가능한 흐름이 없어질 때까지 잔차 그래프에서 작동한다.For example, suppose f(v1,w1) = 1 to flow from v1 to w1. A corresponding residual graph is shown in FIG. 6 . all the edges

The capacitance value for α was decreased by 1, but the capacitance of the edges of the reverse path increased by 1. this reverse path

cost for

silver

is set to The reverse path may be selected by an algorithm, in which case the initial path selection is cancelled. The shortest path fast algorithm operates on the residual graph until there are no possible flows.

EAA-TS includes initialization, path augmentation, and finalization processes that operate before and after the shortest path fast algorithm.

의 값과 비용

이 초기화 된다(라인5-9). Initialization: Residual graph is created from min-cost max flow graph (line 1) and flow

value and cost of

is initialized (lines 5-9).

에 생성되는 증강 경로 p가 없어질 때까지 반복된다(라인 12). Path Augmentation: Use the minimum cost maximum flow graph to find the minimum cost flow (line 11). A corresponding path is created in the residual graph (line 15). And the flow is updated (lines 17-18). This process

It is repeated until there is no augmentation path p being created in (line 12).

와

사이에 흐름이 있으면,

에 있는 하위 작업 중 하나를 처리하기 위해 엣지 서버 j가 선정되고,

값은 이에 따라 갱신된다(라인 21-29).finish:

Wow

If there is a flow between

Edge server j is chosen to process one of the subtasks in

The value is updated accordingly (lines 21-29).

8 is a diagram illustrating a pseudo code for an edge assignment algorithm without job assignment in a job offloading system according to an embodiment.

The Edge Allocation Algorithm without Work Partitioning (EAA-NTS) operates on the minimum cost maximum flow graph as follows.

Minimum cost flow: The minimum cost flow f from the source vertex to the sink vertex is computed (line 5).

과 지수

가 갱신되며 흐름과 연관된 엣지들이 제거된다(라인 6-16). 예를 들어, 최소 비용 흐름이 도5에서

라고 가정하자. 그렇다면

에 해당되는 정점은 들어오는 엣지 및 나가는 엣지와 함께 제거된다.

이 엣지 서버 2로 할당되었기 때문에

값은 2로 갱신된다. 엣지 서버 2에 의해 증가된 처리 사용량을 기록하기 위해서 임시 변수

는

만큼 증가되며 이는 Y_j를 초과해서는 안된다. 이러한 2가지 과정은 모든 작업들이 검토될 때까지 반복된다.Renew and Purge: Work vertices associated with the least cost flow are removed, capacity usage

and exponent

is updated and the edges associated with the flow are removed (lines 6-16). For example, the minimum cost flow is shown in FIG.

Let's assume if so

The corresponding vertices are removed along with the incoming and outgoing edges.

Because it was assigned to Edge Server 2

The value is updated to 2. Temporary variable to record increased processing usage by Edge Server 2

Is

, and it must not exceed Y _j . These two processes are repeated until all tasks have been reviewed.

9 to 14 are graphs for explaining the experimental results of the work offloading system according to an embodiment.

1. Simulation Settings

Simulations can be run to evaluate our scheme for power consumption and total compensation. To this end, it is possible to model the energy characteristics of four commercial servers released between 2019 and 2020. One of the characteristics of these commercial servers was randomly assigned to each edge server.

As an example, in Melbourne, Australia, a database of cell phone network location data was used to model the spatial relationship between edge servers and IoT devices. It contains the latitude, longitude, and IP addresses of 815 wireless access points (WAPs) and 125 base stations, which can be used to calculate the distance between WAPs and each base station.

=1 이다 (그렇지 않으면 0). 각 BS의 커버리지는 450m와 750m사이에 무작위로 선정된 거리로 표현된다.We will assume that each base station has a single edge server capable of performing tasks within edge server coverage. And the location information of the WAPs in the Melbourne database will be obtained to indicate the location of the IoT device. If the WAP connected to the equipment that wants to offload the work is within the coverage of the base station (BS) connected to the ES,

=1 (otherwise 0). The coverage of each BS is expressed as a randomly selected distance between 450m and 750m.

와

사이의 임의의 작업 수를 생성하는 IoT 장비들로 연결될 수 있다. 모든 엣지 서버 j에 대하여 동일하게, 작업

에 대한 기본 보상은

과

사이에 임의로 선정된 정수값이고, 작업

에 의해 요구되는 사용량

는

와

사이에 임의로 선정된 정수값이다. 다르게 서술되지 않는 경우에, 각 변수에 대한 기본 범위는 표1과 같다.To examine the effects of various variables, we varied work, reward, and work usage, which have different ranges, as shown in Table 1. Each WAP is

Wow

It can be connected to IoT devices that create any number of tasks in between. Same for all edge server j, work

The basic reward for

class

It is an integer value randomly selected between

Usage Required by

Is

Wow

An integer value randomly selected between Unless otherwise stated, the default ranges for each variable are shown in Table 1.

Table 1 Parameter Settings

Referring to FIG. 8 , two schemes are evaluated, namely, EAA-TS (algorithm ( FIG. 7 )) and EAA-NTS (algorithm ( FIG. 8 )). Both used a maximum allowable utilization determination algorithm to determine the maximum allowable usage on each edge server. To satisfy inequality (4), these schemes are compared in the following four ways of allocating tasks to edge servers while limiting power consumption.

RR + HR (Round-Robin Allocation + Maximum Reward) allocates tasks in a round-robin manner and successively selects the task that yields the highest reward until inequality (4) is not satisfied.

RA+HR (random task assignment + maximum reward) continuously selects the assignment with high reward for randomly selected tasks until inequality (4) is not satisfied.

MUD+HR (MUD+High Reward) is similar to the Maximum Allowable Utilization Determination Algorithm except that the value is set so that the assignment of work that yields a higher reward is performed first. The values are continuously updated until the inequality (4) is not satisfied.

MUD+HUR (MUD + high base reward) is similar to MUD+HR except that the value is set. The values are continuously updated until inequality (4) is not satisfied.

The first two schemes favor balance, MUD+HR aims for a high reward, and MUD+HUR aims for a high reward ratio based on usage.

2. Work Split Effect

Power limit values are expressed as a percentage of total power consumption when all edge servers are fully used. EAA-TS and EAA-NTS for all combinations of various range values for each variable in Table 1, looking at the three sum rewards of 40%, 60%, and 80% to examine the effect of dividing the work on the total reward. can be compared.

)과

의 다른 값들에 대한 EAA-TS와 EAA-NTS 의 결과 간의 평균 비율 차이를 나타낸다. EAA-TS와 EAA-NTS 결과 간의 비율 차이가 다른 매개변수들의 값에 상관없이 무시할 수 있음을 발견할 수 있다. 예를 들어

=12일 때 0.33%를 넘지 않았고,

=6일 때 0.48%를 넘지 않았다. 하지만,

이 증가함에 따라 차이가 커져가는 것을 확인할 수 있다. 그 이유를 다음과 같이 설명될 수 있다. EAA-NTS는 하나의 작업 전체를 처리하기 위해 남은 용량이 부족한 경우 엣지 서버에 작업을 할당할 수 없으므로, 사용률이 낮으며, 이는 주로 최소 작업 사용량 크기가 1을 초과할 때 발생한다.

이 증가함에 따라, 이 둘 사이의 차이는 줄어들고, 이는 가용 전력이 낮을 때 작업 분할이 효과적이라는 것을 의미한다.Table 1 (

)class

Shows the average ratio difference between the results of EAA-TS and EAA-NTS for different values of . It can be found that the ratio difference between the EAA-TS and EAA-NTS results is negligible regardless of the values of the other parameters. for example

=12 did not exceed 0.33%,

=6 did not exceed 0.48%. However,

It can be seen that the difference increases as this increases. The reason can be explained as follows. EAA-NTS cannot allocate a job to an edge server when the remaining capacity is insufficient to process one entire job, so the utilization rate is low, which mainly occurs when the minimum job usage size exceeds 1.

As α increases, the difference between the two decreases, which means that task splitting is effective when available power is low.

) 과

의 다른 값들에 대한 EAA-NTS와 EAA-TS 비교Table 2: (

) class

Comparison of EAA-NTS and EAA-TS for different values of

3. Power Limiting Effect on Total Rewards

The effect of power limiting on total compensation was examined. From now on, we will use EAA-NTS as the default scheme when comparing other schemes (RR+HR, RA+HR, MUD+HR and MUD+HUR) that do not allow work division. 9 shows the compensation for all five schemes when the default ranges for each parameter in Table 1 are used, and FIG. 10 shows the normalized results based on EAA-NTS.

Higher power limits provide greater rewards. Additionally, EAA-NTS yields high compensation with respect to all power limits. These rewards are 7% to 80% higher compared to other schemes. This difference is greater at lower power limits. The MUD+HUR scheme yields higher rewards than MUD+HR, which means that tasks with higher values should be assigned first. However, MUD+HUR operation does not consider edge server characteristics. This yields, on average, only 70% of EAA-NTS compensation and 40% of the minimum power limit.

4.Total compensation for the number of tasks

값의 평균 보상 결과인 40%, 60%, 80% 를 보여준다. 다시 EAA-NTS가 다른 스킴에 비해 좋은 성과를 보여주었고, 작업의 수가 증가할 수록 차이가 커지는 경향을 보였다. 이를 초기 선정 시에 높은 보상을 갖는 작업이 더 많이 가능하다는 것으로 해석할 수 있다. 예를 들어, 작업 수 범위가 (3,6)에서 (9,18)로 변화했을 때 EAA-NTS에 의한 보상이 5060만큼 증가한 반면에, MUD+HUR로 얻은 보상은 2684만큼 증가함을 확인할 수 있다. We can see how the total reward depends on the number of tasks generated within each WAP. 11 shows three

It shows the average compensation result of 40%, 60%, and 80% of the values. Again, EAA-NTS showed better performance than other schemes, and the difference tends to increase as the number of tasks increases. This can be interpreted as that more tasks with high rewards are possible at the time of initial selection. For example, when the number of tasks range is changed from (3,6) to (9,18), it can be seen that the reward obtained by EAA-NTS increases by 5060, while the reward obtained by MUD+HUR increases by 2684. there is.

5. Effect of Coverage

)의 효과를 살펴보았다. 도12는 40%, 60% 그리고 80%일 때 평균 보상 결과를 나타낸다. 다시 EAA-NTS가

과

간의 차이가 커지면서 마진이 증가하여 다른 스킴들보다 우수한 성능을 나타내는 것을 확인할 수 있다. 작업들에 대해 넓은 보상 범위를 갖는 것이 더 큰 최적화 범위를 부여하며, EAA-NTS이 이를 잘 활용하는 것으로 보여진다.Various compensation ranges (

) was examined. 12 shows the average compensation results at 40%, 60% and 80%. EAA-NTS again

class

As the difference between the two schemes increases, the margin increases, and it can be seen that the performance is superior to that of other schemes. Having a wide compensation range for tasks gives a larger optimization range, and EAA-NTS appears to make good use of this.

6. Task Usage Range Effect

)에 대한 총 보상과 3가지

값 40%, 60% 그리고 80%에 대한 평균값을 나타낸다. 더 적은 수의 작업이 엣지 서버로 할당될 수 있기 때문에, 총 보상은 범위가 (1,6)에서 (3,8)로 증가했을 때 약간 증가한다. 하지만 이 작업들은 더 높은 보상을 받는다. 하지만, 그 범위 밖에서

의 값은 보상 결과에 거의 영향을 주지 못한다. You can see how the total reward varies depending on the range of work usage values. 13 shows a number of different ranges (

) and 3

Average values for values 40%, 60% and 80% are shown. Because fewer jobs can be assigned to edge servers, the total reward increases slightly when the range is increased from (1,6) to (3,8). However, these tasks are highly rewarded. but out of range

The value of has little effect on the reward result.

7. Power consumption required by other schemes to get rewards like EAA-NTS

의 다양한 값들에 대해서 EAA-NTS를 사용했을 때 얻어지는 총 보상을 결정할 수 있다. 동일한 보상을 얻기 위해서 다른 스킴들이 필요로 하는 전력을 찾을 수 있다. 도 14는 각각의 경우를 EAA-NTS에 의해 도출된 전력으로 정규화한 결과를 나타낸다. 동일한 보상을 얻기 위해서 다른 스킴들은 EAA-NTS보다 13%과 75% 더 높은 전력을 필요로 하는 것을 확인할 수 있다. 예를 들어, MUD+HUR은 EAA-NTS보다 13%에서 29% 많은 전력을 필요로 했고, MUD+HR은 19%에서 30% 더 많은 전력을 필요로 했다.

=60%, 70% 및 80%일 때, RR+HR과 RA+HR은 모든 엣지 서버들이 완전히 사용되었음에도 불구하고 EAA-NTS와 같은 보상을 산출하지 못함을 확인할 수 있다.

It is possible to determine the total reward obtained when using EAA-NTS for various values of . You can find the power needed by different schemes to get the same reward. 14 shows the results of normalizing each case to the power derived by EAA-NTS. It can be seen that other schemes require 13% and 75% higher power than EAA-NTS to achieve the same reward. For example, MUD+HUR required 13% to 29% more power than EAA-NTS, and MUD+HR required 19% to 30% more power.

When =60%, 70% and 80%, it can be seen that RR+HR and RA+HR do not yield the same compensation as EAA-NTS, even though all edge servers are fully used.

According to an embodiment, it is possible to propose a scheme for maximizing the total compensation obtained by the edge server while limiting the power consumed by the edge server. In detail, first, the power generated by the edge server based on the processor capacity and usage of each edge server is investigated, and the maximum allowable utilization rate and task allocation of each edge server are determined with the purpose of maximizing the total compensation. Optimization problems can be formulated. To determine the maximum allowable usage rate, you can use a greedy algorithm that tries to get the highest reward with the lowest total power consumption on the edge server. Two algorithms were proposed to formulate the task allocation problem as a minimum-cost-maximum flow graph and to explore the highest reward path in the graph. One algorithm allows task partitioning, and the other algorithm does not allow task partitioning.

We can also run simulations to evaluate our scheme for total compensation, varying the number of jobs, power limit values, ranges of compensation and job usage, and compensation ratios for power consumption. Experimental results show that the scheme proposed in the example utilizes the available power well to obtain a large reward. The reward of the proposed scheme was 7% to 80% higher than that obtained using other schemes, and the other schemes required 13% to 75% more power than the proposed scheme to achieve the same reward. The scheme proposed in the embodiment is particularly effective when the available power is reduced, when the reward assigned to each task is highly variable, and when there are many tasks. Task splitting does not make a big difference when the minimum task usage size is 1, but it is effective when the average task requires high CPU processing capacity.

To support the growing number of IoT tasks, the demand for edge computing is rapidly increasing, resulting in increased power consumption by edge servers. Experimental results presented in the examples indicate that the proposed scheme can contribute to effective power management of edge servers.

The device described above may be implemented as a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA). , a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose or special purpose computers. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that can include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

Software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or apparatus, to be interpreted by or to provide instructions or data to the processing device. may be embodied in The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible from the above description by those skilled in the art. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (10)

A work offloading method performed by a computer-implemented task offloading system, the method comprising:
determining an edge server to execute an offloaded task based on power information of each edge server; and
Offloading the job according to the assignment information set in the determined edge server
A method of offloading work that includes According to claim 1,
The step of offloading the task is:
determining allocation information including a maximum allowable utilization rate of each edge server and a job allocation for maximizing an edge server provider's compensation under the power limit of each edge server;
A method of offloading work that includes According to claim 1,
The step of offloading the task is:
Determining the maximum allowable utilization rate of each of the edge servers while giving high priority to offloading jobs with a high compensation ratio of the edge server provider for the power demand
A method of offloading work that includes According to claim 1,
The step of offloading the task is:
Searching for values of the maximum allowable usage rate while satisfying the power constraint of each edge server
A method of offloading work that includes 5. The method of claim 4,
The step of offloading the task is:
By iteratively calculating power increments and reward increments from task assignments, priority is given to the maximum allowable utilization value that gives a higher reward to the power ratio until the power constraint is met and the maximum allowable utilization is determined on each edge server. step to grant
A method of offloading work that includes According to claim 1,
The step of offloading the task is:
Searching for a value for each job for allocating a job to each edge server
A method of offloading work that includes 7. The method of claim 6,
The step of offloading the task is:
Performing task assignment to the edge server using either a method that allows work splitting or a method that does not allow work splitting using the minimum cost maximum flow graph.
A method of offloading work that includes 8. The method of claim 7,
The method to allow the splitting of work is:
A residual graph is generated from the minimum cost and maximum flow graph, and the process of initializing the value and cost of a binary function to express the flow between each node, and the process of deriving the minimum cost flow using the shortest path fast algorithm to the residual graph The index value of the edge server to which the sub-task is assigned by selecting an edge server to process one of the sub-tasks constituting the task when there is a flow between the vertex and the path augmentation process of creating a route and updating the flow to perform the finishing process to update
How to offload work. 8. The method of claim 7,
The method that does not allow the above work division is,
A minimum cost flow process that operates on the minimum cost maximum flow graph and calculates a minimum cost flow from a source vertex to a sink vertex, removes the task vertex associated with the minimum cost flow, and the capacity usage and the index value of the edge server to which the task is assigned to update and remove the edges associated with the flow.
How to offload work. In the computer-implemented task offloading system,
a server determining unit that determines an edge server to execute an offloaded task based on power information of each edge server; and
A job offloading unit for offloading a job according to the assignment information set in the determined edge server
A work offloading system comprising a.

Images