KR102068676B1 - The method for scheduling tasks in real time using pattern-identification in multitier edge computing and the system thereof - Google Patents

Abstract

The present invention relates to a method for scheduling a task in real time by using pattern recognition in multi-layer edge computing and a system thereof. The method for scheduling a task in real time by using pattern recognition in multi-layer edge computing according to an embodiment of the present invention can comprise the steps of: receiving task information from an IoT apparatus and training a task pattern with respect to a timeslot offline; and optimally allocating a new task based on the trained task pattern.

Description

TECHNICAL METHODS AND SYSTEMS FOR SCHEDULE WORKS IN REAL TIME USING PATTERN IDENTIFICATION IN MULTIPLE EDGE EDGE COMPUTING.THE METHOD FOR SCHEDULING TASKS IN REAL TIME USING PATTERN

The present invention relates to a method for scheduling a task in real time using pattern identification in multi-layer edge computing. More specifically, the present invention relates to a method for scheduling a task by allocating an optimal task online using a self-learning algorithm.

Smartization of manufacturing is one of the major trends to realize the Fourth Industrial Revolution. In smart factories, production machines and auxiliary devices that implement the Industrial Internet of Things (IIoT) are connected to integrated networking. In this environment, all connected devices interact with each other and request work orders from the central management.

If the correct production process encounters an unexpected problem, it coordinates the reaction of the machine and robot by requesting trigger feedback from the Control Entities.

For example, if smoke is detected in a warehouse, an immediate shutdown command must be sent to the electrical work chain and a real-time activation command must activate the fire extinguishing system.

However, since such Internet of Things (IIoT) devices are assumed to be lightweight embedded platforms with limited computing resources, it is not appropriate to run real-time services using their own capabilities.

However, advances in 5G Mobile Edge Computing (MEC) technology can provide autonomous and reliable real-time offload services.

Mobile Edge Computing (MEC), defined by the European Telecommunications Standards Institute (ETSI), provides cloud computing capabilities and IT service environments at the edge of the network. More specifically, multi-layer MEC (mMEC) is a virtualized hierarchical MEC framework in which edge performance implementation entities (ECEs) are layered according to computing power.

The multi-layer mobile edge computing (mMEC) framework is supervised by a central orchestrator. The central orchestrator schedules offloaded tasks on IIoT devices for the purpose of maximizing computing performance and reducing energy consumption while maintaining the requirements of IIoT applications.

Therefore, there is a need for a mechanism for scheduling offloaded IIoT tasks in real time to optimize task assignment.

(Non-Patent Document 0001) Amit Agarwal, Saloni Jain (2014). Efficient Optimal Algorithm of Task Scheduling in Cloud Computing Environment. UCTT, Vol. 9, 344-349.

(Non-Patent Document 0002) Haoyu Wang, Jiaqi Gong, Yan Zhuang, Haiying Shen, John Lach (2017). Healthedge: Task Scheduling for Edge Computing with Health Emergency and Human Behavior Consideration in Smart Homes. IEEE.

The technical problem to be solved by the present invention is to provide a method for scheduling offloaded IIoT tasks in real time for optimization of task assignment.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

According to an embodiment of the present invention, a method of scheduling a job in real time using pattern identification in multi-layer edge computing includes: receiving task information from an IoT apparatus and learning a task pattern on a timeslot offline; And optimally allocating a new task based on the learned task pattern.

Optimally assigning a new task according to an embodiment of the present invention includes: identifying a set of self-organization maps for the task pattern; Calculating, before each time slot, the optimal task assignment expected for the self-organizing map; And allocating the task by finding a neuron that best matches a task newly arriving at the mobile edge computing framework based on the calculated task assignment.

In a method of scheduling a task in real time according to an embodiment of the present invention, identifying a set of self-organizing maps and calculating the optimal task assignment are performed offline, and assigning the task is performed online. It is characterized in that the implementation.

According to an embodiment of the present invention, calculating the optimal task assignment may be to find a neuron whose equation F is the minimum.

[Equation 1]

는 뉴런 j 가 ECEs s에 할당되면 1, 그 이외는 0이며,

는 뉴런j가 ECEs s에 할당될 때 총 대기 시간을 의미하며 하기 수학식 2에 의해 결정된다.Where indicators

Is 1 if neuron j is assigned to ECEs s, 0 otherwise

Denotes the total waiting time when neuron j is assigned to ECEs s and is determined by Equation 2 below.

[Equation 2]

Here, w _j1, w _j2, w _j3, w _j4 are weight vectors, and mean values of relative job size, relative average processing complexity, relative response size, and relative execution deadline, respectively, and b _s is a job waiting to be processed in ECE. The current buffer size, ra _js is The access data rate (bps) from the device with task j to the network, rb _js is the internal forwarding rate for ECEs in bps, C _s is the CPU cycle of the ECEs, the number of k neurons, and m is the number of neurons.

In the multi-layer edge computing according to an embodiment of the present invention, the system for scheduling a job in real time using pattern identification includes: a task pattern learning unit learning task patterns with respect to timeslots by receiving task information from an IoT apparatus; An optimal task allocator configured to optimally allocate a new task based on the learned task pattern, wherein the optimal task allocator includes: a self-organized map set identification unit identifying a self-configured map set for the task pattern; An optimal task assignment calculator that calculates the optimal task assignment expected for the self-organization map before each time slot and a neuron that best matches the newly arrived task in the mobile edge computing framework based on the calculated task assignment. It may include an online neuron matching unit for finding and assigning the task.

According to a method and system for scheduling a task in real time using pattern identification in multi-layer edge computing according to an embodiment of the present invention, task allocation delay may be greatly reduced.

According to an embodiment of the present invention, since task assignment calculation time is shortened and computation complexity and calculation wait time are reduced, task allocation efficiency can be improved in an environment using the Industrial Internet of Things (IIoT).

In addition, an embodiment of the present invention may continuously perform real-time task assignment without waiting for task assignment calculation to be completed.

Effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a conceptual diagram of a method and system for scheduling a job in real time using pattern identification in multi-layer edge computing according to an embodiment of the present invention.
2 is a flowchart of a real-time job scheduling method according to an embodiment of the present invention.
3 is a diagram illustrating a self-learning algorithm used according to an embodiment of the present invention.
4 is a flowchart for optimally allocating a new task according to an embodiment of the present invention.
5 shows a timeline of a method for calculating task assignment in accordance with an embodiment of the present invention.
6 is a graph comparing the results of the real-time job scheduling method according to an embodiment of the present invention with an optimal solution.
7 is a graph comparing job processing waiting time of a real-time job scheduling method according to an embodiment of the present invention with another algorithm.
8 illustrates execution error rate of a real-time job scheduling method according to an embodiment of the present invention.
9 illustrates an average buffering wait time of a real time job scheduling method according to an embodiment of the present invention.

As the invention allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the drawings, similar reference numerals are used for similar elements.

Terms such as first, second, A, and B may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term and / or includes a combination of a plurality of related items or any item of a plurality of related items.

When a component is said to be "connected" or "connected" to another component, it may be directly connected to or connected to that other component, but it may be understood that another component may be present in the middle. Should be. On the other hand, when a component is said to be "directly connected" or "directly connected" to another component, it should be understood that there is no other component in between.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art, and shall not be construed in ideal or excessively formal meanings unless expressly defined in this application. Do not.

Throughout the specification, Mobile Edge Computing (MEC) refers to a technology for creating a new local network by applying distributed cloud computing technology to a wireless base station and deploying various services and caching contents close to a user terminal. Is a virtualized multi-layered MEC framework, which refers to Edge Computing-enabled entities layered according to computing capability.

Throughout the specification, a task is referred to as a kind of work to be performed by an industrial IoT apparatus, and an edge computing entity (hereinafter, referred to as ECEs) is referred to as a configuration for performing a specific task.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a conceptual diagram of a method and system for scheduling a job in real time using pattern identification in multi-layer edge computing according to an embodiment of the present invention.

Referring to FIG. 1, a multi-layer edge computing (mMEC) framework 101 (hereinafter referred to as framework in mMEC) is managed and supervised by a central orchestrator 104, and the mMEC framework 101 is an industrial object. Not only serves to connect the Internet (IIoT) device 102 to the Internet, but also process data received by the IIoT device 102.

The mMEC framework 101 consists of a router 103, a central orchestrator 104, a wireless access point / switch 105, and a local server / storage 106.

Data generated at the IIoT device 102 is collected via the access point 105 and processed at the central orchestrator 104 and stored in the storage 106.

That is, the real-time job scheduling method according to an embodiment of the present invention operates in the mMEC framework 101 and is implemented in the central orchestrator 104.

2 is a flowchart of a real-time job scheduling method according to an embodiment of the present invention.

In a method of scheduling a job in real time using pattern identification in multi-layer edge computing according to an embodiment of the present invention, receiving task information from an IoT apparatus and learning a task pattern with respect to timeslots offline. Step S201; And optimally allocating a new task based on the learned task pattern (S202).

The learning of the task pattern is performed by the central orchestrator receiving task information from the IoT apparatus and applying a self-organizing map (SOM) to construct a map representing the characteristics of the task pattern.

3 is a diagram illustrating a self-learning algorithm used according to an embodiment of the present invention.

Referring to FIG. 3, the self-learning algorithm (hereinafter referred to as SOM) first finds a neuron that best matches task i, and all neurons in the vicinity of the neuron proceed in order to update the weight vector close to task i. .

는, j개의 뉴런에 대해 뉴런과 태스크 i사이에 유클리디언(Euclidean) 거리를 계산하여 태스크 i에 가장 작은 유클리디언 거리를 가지는 유닛으로 결정된다.Neurons that best match task i

Is calculated as the unit having the smallest Euclidean distance to task i by calculating the Euclidean distance between the neurons and task i for j neurons.

[Equation 1],

의 주변의 모든 뉴런들은 태스크 i에 가깝게 가중치 벡터를 업데이트하는데, 하기 수학식 2를 기초로 업데이트 된다.And the neurons

All neurons in the vicinity of update the weight vector close to task i, which is updated based on Equation 2 below.

[Equation 2]

를 업데이트함에 있어, 이웃함수

, 학습요소

를 이용한다. 여기서, 이웃함수

는 이웃에 있는 것을 정의하기 위해 뉴런

로부터 거리를 결정하며, 학습요소

는 입력 태스크i에서 뉴런 j까지의 거리에 따라 입력 태스크 i에서 뉴런 j에 대한 영향의 양을 계산하는 학습인자를 나타낸다. weight

In updating, the neighbor function

, Learning factor

Use Where neighbor function

Neurons to define what is in the neighborhood

Distance from the learning element

Denotes a learning factor that calculates the amount of influence on neuron j in input task i according to the distance from input task i to neuron j.

는 태스크 세트로 입력값을 의미한다.here

Means a set of inputs to a task set.

와 학습요소

의 관계는 수학식 3과 수학식 4와 같이 나타낼 수 있다.Neighbor function

And learning elements

Can be represented as in Equation 3 and Equation 4.

[Equation 3]

[Equation 4]

는 이웃 반지름인데, 수학식 5와 같이 정의 될 수 있다.here

Is the neighbor radius, which can be defined as Equation 5.

[Equation 5]

Where R is the maximum radius from the center of the self-study map, and X is the number of tasks.

4 is a flowchart for optimally allocating a new task according to an embodiment of the present invention.

More specifically, FIG. 4 illustrates a more detailed step S202 of optimally allocating a new task based on the learned task pattern of FIG. 2.

Therefore, optimally allocating a new task includes identifying a set of self-organization maps for the task pattern (S301); Calculating (S302) an optimal task assignment expected for the self-organizing map before each time slot; And (S303) finding a neuron that best matches a task newly arriving at the mobile edge computing framework based on the calculated task assignment.

Referring to step S301 of identifying the self-organization map set for the task pattern in more detail, the IIoT task i may be represented by a fourth order feature vector (Equation 6).

[Equation 6]

Each fourth order vector represents, in order, task size, average processing complexity, relative response size, and relative execution deadline.

Since the production process runs continuously for a given period, the average number of task sets X (t) of IIoT devices offloaded to the mMEC framework within each time slot of this period N neurons in the SOM if there are N _x tasks k is selected as in Equation 7 below.

[Equation 7]

Here, <·> is the nearest integer function, and λ is the scale level of reducing the SOM size.

Based on these settings, the k-neuron SOM is trained using all IIoT tasks that are offloaded for a given period. During each characteristic period, the IIOT task is typically represented by obtaining a trained SOM. A unique period is defined as the offloaded IIoT task arrival cycle, after which the workload is repeated again in terms of volume and characteristics.

The step S302 of calculating the optimal task assignment expected for the self-organizing map SOM before each time slot is described with reference to FIG.

5 shows a timeline of a method for calculating task assignment in accordance with an embodiment of the present invention.

5, prior to time slot t, in order to determine the optimal task assignment expected during timeslot t, the task assignment calculation mechanism according to one embodiment of the present invention performs an H (t) function at t-Δt. To perform.

The H (t) function uses the Hungarian Method, which solves the problem of expected optimal task allocation between a set of neurons in a typical SOM and a set of current ECEs in time slot t.

The result of the H (t) function can be applied to online task assignment for a period from t to t + 1, even if the H (t + 1) function starts at (t + 1)-Δt.

The expected optimal task allocation problem can define the current performance state (ie, job buffer and CPU frequency) and transfer state (ie, transfer data rate and forwarding data rate between the IIoT device and the mMEC framework) of the ECEs. Task allocation aims to minimize the total latency of task execution when allocating a set of neurons from a typical SOM to ECEs.

The expected optimal task allocation problem can be formulated as: W and M represent a set of k neurons in the SOM and a set of m ECEs in the mMEC framework.

Since SOM neurons were trained using recorded tasks, the weight vector of the neurons reflects the mean value of the task features. The weight vectors w _j can be classified into w _j1 , w _j2 , w _j3 , and w _j4 , which represent the average values of relative job size, relative average processing complexity, relative response size, and relative execution deadline, respectively. When the current buffer size of a task waiting to be processed in ECEs is b _s CPU cycle, when the neuron j is allocated to the ECEs, the total waiting time L _js is expressed by Equation 8.

[Equation 8]

Where ra _js is The access data rate (bps) from the device with task j to the network, rb _js is the internal forwarding rate for ECEs in bps, C _s is the CPU cycle of the ECEs, the number of k neurons, and m is the number of neurons.

In Equation 8, (a) represents an upload waiting time, (b) represents a queue waiting time, (c) represents a job processing waiting time, and (d) represents a response waiting time.

The access data rate ra _js may be expressed as in Equation 9.

[Equation 9]

Here, BW _js is the bandwidth allocated to the IIoT device, and SINR _j is the channel quality between the IIoT device and the network.

The optimal task assignment problem F expected based on this may be expressed as in Equations 10 to 14 below.

[Equation 10]

[Equation 11]

[Equation 12]

[Equation 13]

[Equation 14]

Equation 13 can assign neurons to only one ECEs and set the maximum match for all m neurons in the SOM.

That is, by finding a neuron of which F is the minimum in Equation 10, the found neuron may correspond to a specific task.

The optimal task assignment problem F finds the optimal solution by applying two sets W and M to the graph using the Hungarian method (hereinafter referred to as H (·)). The function H (·) is performed as follows.

(I) Supplement the addition of constants (e.g. 0) to reinforce the square matrix (referred to as V) of the dimension maximum (| W |, | M |) to a queue of all possible work assignments in the neuron set and the ECE set. do.

(II) Subtract the smallest row item from each row item of V, and subtract the smallest column item from each column item of V.

(III) Find the minimum number of rows and columns that cover all zero entries.

(IV) If ν is equal to the magnitude of V, an optimal solution is found. Select an item with two or fewer items in the same row or column, and if υ is less than the size of V, determine the minimum item c found. Then, subtract c from all uncovered rows and add c to all covered columns. Go back to the previous step and find the minimum number υ again.

For each time slot, the function H (·) is calculated before Δt before the start of the time slot to find the optimal task assignment solution.

On the basis of the calculated task assignment, which is the last step of the step of optimally allocating a new task (S202), the step of finding a neuron that best matches the newly arrived task in the mobile edge computing framework and assigning the task (S303) is online. Is performed.

When the IIoT task arrives at the mMEC framework in online mode, the task matches the SOM to apply Equation 1 to the computational complexity function to find the best matching neurons. According to the optimal assignment for neurons determined by function H (·) in offline mode, this assignment is performed immediately for the IIoT task that arrives.

6 to 9 are experimental results showing the superiority of the real-time job scheduling method and the system according to an embodiment of the present invention.

6 is a graph comparing the results of the real-time job scheduling method according to an embodiment of the present invention with an optimal solution.

More specifically, we show the result of the optimal task assignment F upon various IIoT job arrivals for 300 timeslots. The numerical results indicate the approximate performance of real-time task scheduling (hereinafter, referred to as pattern-identified online task scheduling (PIOTS)) according to an embodiment of the present invention in comparison with an optimal solution. Almost identical to the Optimal value).

Meanwhile, the optimal value is calculated based on the set of IIoT tasks collected in each time slot.

The difference in average job processing latency is 0.159 ms, and the real-time job scheduling method according to an embodiment of the present invention performs the expected task assignment based on a typical task set derived from the SOM map, and the actual task assignment is online. It is noteworthy that it does.

7 is a graph comparing job processing waiting time of a real-time job scheduling method according to an embodiment of the present invention with another algorithm.

Referring to FIG. 7, the PIOTS method according to an embodiment of the present invention has a task processing time longer than other types of algorithms, OHTA (Online Greedy Task Assignment) and OGTA (regardless of the number of IIoT devices). You can see a shorter one.

The PIOTS method according to an embodiment of the present invention achieves effective performance because a general set of tasks is used to predetermine the expected task assignment for an incoming task. The OHTA, on the other hand, performs task assignment based on input tasks collected during each time slot. OHTA better adapts to various job arrivals, but has the disadvantage of waiting until the end of each timeslot to collect jobs and determine the optimal allocation. On the other hand, the OGTA plan can confirm that the performance against latency is the lowest.

8 illustrates execution error rate of a real-time job scheduling method according to an embodiment of the present invention.

Execution error rate% is defined as the ratio of deadline-violated IIoT devices among all connected devices in the network. 8 shows simulation results corresponding to three thresholds of execution deadlines 0.5 seconds, 1.0 seconds and 1.5 seconds. During 300 simulated time slots, the execution error rate of the 0.5 second deadline is approximated in the PIOTS, OHTA and OGTA schemes (0.970%, 0.976% and 0.977%). On the other hand, with a 1.5 second deadline, the PIOTS approach reduced the execution error rate to 0.826%. This reduction is about 5.7% and 1.2% reduction compared to the OGTA and OHTA systems. Analysis according to FIG. 8 shows that the PIOTS scheme provides balanced work distribution among ECEs to meet task execution deadlines.

9 illustrates an average buffering wait time of a real time job scheduling method according to an embodiment of the present invention.

9, the average buffering wait time of IIoT tasks arriving at ECEs during 100, 200, and 300 time slots is shown by comparing PIOTS, which is an embodiment of the present invention, with other schemes, OHTA and OGTA.

Arrived task volumes are configured to accommodate ECE performance leading to saturation in the network, resulting in increased latency due to time delay. It is observed that the OGTA scheme has significant buffering latency in all time slots. On the other hand, the PIOTS method according to an embodiment of the present invention achieves a low buffering latency of 44.33% and 4.32% in time slot 300, respectively, compared to the OGTA method and the OHTA method.

Although the embodiments of the present invention have been described above with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing the technical spirit or essential features thereof. I can understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

101: mMEC framework
102: Industrial Internet of Things (IIoT) devices
103: router
104: central orchestrator
105: wireless access point / switch
106: local server / storage

Claims (6)

Receiving task information from the IoT apparatus and learning a task pattern with respect to the timeslot offline;
Determining a total waiting time including upload waiting time, queue waiting time, job processing waiting time, and response waiting time corresponding to each neuron based on the learned task pattern; And
Optimally allocating a new task based on the neurons with the lowest total latency;
How to schedule jobs in real time using pattern identification in multi-layer edge computing. The method of claim 1,
Identifying a set of self-organization maps for the task pattern;
Calculating, before each time slot, the optimal task assignment expected for the self-organizing map; And
Based on the calculated task assignment, finding a neuron that best matches the newly arrived task in the mobile edge computing framework, and assigning the task;
How to schedule jobs in real time using pattern identification in multi-layer edge computing. The method of claim 2,
Identifying the self-organizing map set and calculating the optimal task assignment are performed offline, and assigning the task is performed online.
How to schedule jobs in real time using pattern identification in multi-layer edge computing. The method of claim 1,
The step of calculating the optimal task assignment is to find a neuron that the following equation (F) (Equation 1) is the minimum in scheduling a task in real time using pattern identification in multi-layer edge computing.
[Equation 1]

Where indicators

Is 1 if neuron j is assigned to ECEs s, 0 otherwise, and L _js is the total waiting time when neuron j is assigned to ECEs s and is determined by Equation 2 below.
[Equation 2]

Here, w _j1, w _j2, w _j3, w _j4 are weight vectors, and mean values of relative job size, relative average processing complexity, relative response size, and relative execution deadline, respectively, and b _s is a job waiting to be processed in ECE. The current buffer size, ra _js is The access data rate (bps) from the device with task j to the network, rb _js is the internal forwarding rate for ECEs in bps, C _s is the CPU cycle of the ECEs, the number of k neurons, and m is the number of neurons. A task pattern learning unit that receives task information from the IoT apparatus and learns a task pattern with respect to a timeslot offline; and
Based on the learned task pattern, a total waiting time is determined including upload waiting time, queue waiting time, job processing waiting time, and response waiting time corresponding to each neuron, and based on the neurons having the minimum waiting time. An optimal task allocator that optimally allocates new tasks is included.
A system for scheduling jobs in real time using pattern identification in multi-layer edge computing. The method of claim 5,
The optimal task allocator is a system for scheduling a task in real time using pattern identification in multi-layer edge computing characterized in that the following equation (F) to find the minimum neuron.
[Equation 3]

Where indicators

Is 1 if neuron j is assigned to ECEs s, 0 otherwise, and L _js is the total waiting time when neuron j is assigned to ECEs s and is determined by Equation 4 below.
[Equation 4]

Here, w _j1, w _j2, w _j3, w _j4 are weight vectors, and mean values of relative job size, relative average processing complexity, relative response size, and relative execution deadline, respectively, and b _s is a job waiting to be processed in ECE. The current buffer size, ra _js is The access data rate (bps) from the device with task j to the network, rb _js is the internal forwarding rate for ECEs in bps, C _s is the CPU cycle of the ECEs, the number of k neurons, and m is the number of neurons.

Images