CN110069341A - What binding function configured on demand has the dispatching method of dependence task in edge calculations - Google Patents

Abstract

The invention discloses a scheduling method for tasks with dependencies that are configured on demand in combination with functions in edge computing. The parameters are greedy initial configuration of the edge server to obtain server configuration information; step 3, use a directed acyclic graph to represent the tasks with dependencies in step 1, and topologically sort the tasks in the directed acyclic graph into a topological sequence; step 4. Use the server configuration information of step 2 to iterate the topology sequence of step 3, calculate the earliest completion time of each task running on each edge server, and obtain the task allocation and scheduling plan; step 5, under the actual capacity constraints of the edge server , and assign and schedule each task according to the task assignment and scheduling scheme in step 4. This method can minimize the completion time of an application composed of multiple dependent tasks in an edge computing environment.

Description

Scheduling method for dependent tasks with on-demand configuration combined with functions in edge computing

technical field

The invention relates to the field of edge computing, in particular to a scheduling method for tasks with dependencies in edge computing that combine functions with on-demand configuration.

Background technique

In recent years, with the rapid development of cellular networks and the Internet of Things (IOT), high-speed and high-reliability air interfaces have enabled applications with high complexity and high energy consumption to be offloaded to remote cloud data centers for operation, thereby making up for the computing power of mobile terminals. and reduce its energy consumption. However, long-distance propagation inevitably leads to severe communication delays, which cannot meet the real-time response requirements of applications such as augmented reality (VR), cognitive assistance, and connected vehicles. To alleviate this problem, there has been an important paradigm shift in the field of mobile computing, from centralized cloud computing to edge computing (Edge Computing, also known as fog computing, micro-cloud computing). The idea of edge computing is to deploy small servers, called edge servers, at the edge of the Internet (such as Wi-Fi access points or cellular base stations), which have more powerful computing and storage capabilities than mobile devices and are geographically located. Close to mobile users, usually mobile users can directly connect to edge servers through wireless networks, which greatly reduces communication latency and enables mobile users to seamlessly access cloud services with low latency.

However, as the performance requirements and resource demands of mobile applications are increasing dramatically, edge computing faces many challenges in practical applications, such as:

(1) Capacity limitation and on-demand configuration: Compared with remote cloud computing, edge servers have relatively limited computing and storage capabilities, and only a limited number of functions can be configured on edge servers. In order to run a certain task, the edge server needs to perform operations such as database caching, image download, installation and startup, and additional environment configuration. These series of operations can be called function configuration. Therefore, a task can only be performed on the edge with the required function. run on the server. If the current edge server does not have enough capacity to configure the function corresponding to the currently scheduled task, you need to decide to remove the configured functions on some edge servers. The on-demand configuration of functions will significantly affect the performance of mobile applications and the utilization of edge servers, so how to provide intelligent function configuration strategies is crucial.

(2) Task dependencies and parallel execution: A mobile application consists of multiple tasks with dependencies, usually represented by a directed acyclic graph (DAG). The points in the graph represent different types of tasks, and the value on the directed edge represents that a certain amount of data needs to be transmitted after the end of the task as the input of the arrow pointing to the task, so the edge set in the graph also defines the sequence or parallel relationship of task execution. . In addition, different tasks may have different preferences for edge servers. For example, in a Facebook video processing application, encoding operations are computationally intensive tasks and are more suitable for edge servers with more powerful computing performance. In order to minimize the application completion time as much as possible, how to design a reasonable scheduling strategy is a problem that needs to be solved, including deciding which edge server to place each task in the DAG and the order in which the tasks are executed on each edge server.

At present, in the field of mobile edge computing, there is a lot of work on task scheduling and function configuration, but the existing algorithms do not consider the dependencies of tasks in the application, but assume that the application is an independent and inseparable whole. With the increasing complexity of mobile applications, distributing parallel tasks to different edge servers can effectively optimize the performance of mobile applications. However, in a resource-constrained edge computing environment, how to configure functions and schedule tasks with dependencies is an urgent problem to be solved.

SUMMARY OF THE INVENTION

Based on the problems existing in the prior art, the purpose of the present invention is to provide a scheduling method for tasks with dependencies in edge computing combined with on-demand configuration of functions, which can solve the problem that task scheduling in existing edge computing does not consider tasks in application programs. Dependency, which leads to the problem of low application operation efficiency in edge computing.

The purpose of this invention is to realize through the following technical solutions:

Embodiments of the present invention provide a scheduling method for tasks with dependencies configured on demand in combination with functions in edge computing, including:

Step 1, obtain the relevant parameters of the edge computing network and the application including the task with dependency, and select an edge server from the edge computing network as the initialization server for processing the input and output of the application;

Step 2, using the relevant parameters of the application obtained in the step 1 to perform a greedy initial configuration on each edge server in the edge computing network to obtain server configuration information;

Step 3, use a directed acyclic graph to represent the tasks with dependencies of the application in the step 1, and perform topological sorting on the tasks in the directed acyclic graph to obtain the topological sequence of the tasks;

Step 4: Use the server configuration information obtained in step 2 to iteratively calculate the topology sequence of the tasks obtained in step 3, and calculate that each task in the topology sequence is placed on each edge server of the edge computing network and runs the earliest to end. According to the completion time of the last task and store the corresponding allocation process, reversely search the stored allocation processes according to the completion time of the last task to reversely reconstruct the allocation and scheduling scheme of all tasks;

Step 5: All tasks are allocated and scheduled according to the task allocation and scheduling scheme finally determined in step 4 under the constraints of the actual capacity of the edge server.

It can be seen from the technical solutions provided by the present invention that the embodiments of the present invention provide a scheduling method for tasks with dependencies that are configured on demand in combination with functions in edge computing, and the beneficial effects are as follows:

The method of the invention realizes the on-demand configuration of the decision-making function, as well as which edge server each task is placed on and the execution sequence of the tasks on each edge server. In edge computing, under the premise of considering the limited number of functions on the edge server, the efficient scheduling of tasks with dependencies and the on-demand configuration of functions can improve the utilization of edge servers, and minimize one in the edge computing environment. Completion time of an application composed of multiple dependent tasks (completion time refers to the time it takes for the application to obtain the final running result and return it to the mobile user after the application is uninstalled from the edge server or remote cloud server), reduce the application running time, and use it with other modifications. Compared with the method in this scenario (such as the HEFT algorithm proposed by topcuoglu), it can reduce the application completion time by 1.54 to 2.8 times and improve the user experience.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a flowchart of a method for scheduling tasks with dependencies that are configured on demand in combination with functions in edge computing provided by an embodiment of the present invention;

2 is a model diagram of a configuration method provided by an embodiment of the present invention;

3 is a schematic structural diagram of a directed acyclic graph of three applications provided by the implementation of the present invention;

FIG. 4 is a comparison diagram of the implementation performance of the method provided by the implementation of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the specific content of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention. Contents that are not described in detail in the embodiments of the present invention belong to the prior art known to those skilled in the art.

As shown in FIG. 1 , an embodiment of the present invention provides a method for scheduling tasks with dependencies in edge computing combined with on-demand configuration of functions, which can minimize the completion of an application composed of multiple dependent tasks in an edge computing environment time, thereby improving the efficiency of application operation in edge computing and improving user experience, including:

Step 1, obtain the relevant parameters of the edge computing network and the application including the tasks with dependencies, and select an edge server from the edge computing network as the initialization server of the application;

Step 2, using the relevant parameters obtained in step 1 to perform greedy initial configuration on each edge server in the edge computing network to obtain server configuration information;

Step 3, use a directed acyclic graph to represent the tasks with dependencies in the step 1, and perform topological sorting on the tasks in the directed acyclic graph to obtain a topological sequence of tasks;

Step 4: Use the server configuration information obtained in step 2 to iteratively calculate the topology sequence of the tasks obtained in step 3, and calculate that each task in the topology sequence is placed on each edge server of the edge computing network and runs the earliest to end. and store the corresponding process, and reversely search the stored processes according to the completion time of the last task to reversely reconstruct the allocation and scheduling scheme of all tasks;

Step 5: All tasks are allocated and scheduled according to the task allocation and scheduling scheme finally determined in step 4 under the constraints of the actual capacity of the edge server.

In step 1 of the above method, the edge computing network includes:

A remote cloud server and multiple heterogeneous edge servers, each edge server has a limited capacity, wherein the bidirectional data transmission rate between any two edge servers is equal.

In step 1 of the above method, obtaining the relevant parameters of the edge computing network and the task with dependencies includes:

The running time of each task on each edge server of the edge computing network and the time it takes for each edge server to configure different functions.

In step 1 of the above method, the server where the greedy initial configuration starts is determined by the minimum running time of all tasks on all servers. This initialization server actually simulates a mobile device to offload one of its computing tasks (such as a face recognition application) to an edge server (that is, the initialization server here), hoping to use the resources of the initialization server to help it faster. After completing the computing task, the mobile device will provide input data required by the computing task, and the initialization server will return the final computing result.

In step 2 of the above method, the relevant parameters obtained in step 1 are used to perform greedy initial configuration on the edge servers in the edge computing network to obtain the server configuration information, and a corresponding array record configuration function is set on each edge server. The information stored in the array is the obtained server configuration information. The configuration process includes the following steps:

Step 21: Under the premise of ignoring the actual capacity of the edge server, greedily ensure that each task is configured with the corresponding function on the edge server with the least running time, record the function number in the array of the edge server, and calculate Under the current configuration, the maximum capacity cost value in the edge server is used as the virtual capacity;

Step 22, set the capacity of all edge servers as virtual capacity, and continue to perform the following configuration on the edge servers that are not fully configured in step 21: sort the running time of all tasks on all edge servers from small to large, and determine the running time in turn Whether the corresponding edge server is fully configured, if it is fully configured, skip the next steps; if not, determine whether the function corresponding to the task has been configured on the under-configured edge server. If so, skip the next steps. Otherwise, configure and save it to the array, and then skip to the next runtime for judgment until all edge servers are full.

The processing of the above-mentioned step 21 is to count how much capacity each edge server needs to spend under the greedy configuration scheme. maximum capacity consumption).

In the above step 22, under-provisioning means that the number of configured functions is less than the virtual capacity of the edge server;

Full configuration means that the number of functions configured is equal to the virtual capacity of the edge server.

The processing of the above step 22 is specifically to continue to perform function configuration on the edge servers that are not fully configured in step 21, so that each edge server consumes its virtual capacity, and in practice, a two-dimensional array can be used to record the functions configured on each edge server. .

In step 3 of the above method, using a directed acyclic graph to represent the tasks with dependencies applied in step 1 includes:

Among the tasks with dependencies, each task represents a computing module in the application, and the tasks with dependencies are set as a task set Among them, the running time of task v _j on the edge server _sk is t _jk ;

with a directed acyclic graph Representing an application, set the directed edge e:=(vi,v _j )∈ε to indicate that the execution of task v _j requires the result of task v _{i ,} as input, where the amount of transmitted data is w _ij .

In the above step 4, the assignment and scheduling scheme of tasks is obtained in the dynamic programming process. In the dynamic programming method, the results of the sub-problems will be memorized and stored. When the iteration reaches the last step (that is, the last task), the minimum value is obtained. To calculate the completion time, the task allocation and scheduling scheme can be reconstructed in reverse from the result, and the assignment of each task under the completion time and the task scheduling order on each server can be recorded with an array, and then the finalized task allocation and scheduling scheme can be obtained.

The method of the invention realizes the on-demand configuration of the decision-making function, as well as which edge server each task is placed on and the execution sequence of the tasks on each edge server. In edge computing, under the premise of considering the limited number of functions on the edge server, the efficient scheduling of tasks with dependencies and the on-demand configuration of functions can improve the utilization of edge servers, and minimize one in the edge computing environment. Completion time of an application consisting of multiple dependent tasks (completion time refers to the time it takes for an application to be uninstalled from an edge server or a remote cloud server to obtain the final running result and return it to the mobile user), reduce application running time and improve user experience.

The embodiments of the present invention will be described in further detail below.

The method for scheduling tasks with dependencies that are configured on demand in combination with functions in an edge computing embodiment of the present invention is a method for scheduling tasks with dependencies that are configured in combination with functions, including: model definition and processing steps;

Among them, (1) the network environment and each model used in the scheduling method are defined as follows:

(11) Edge computing network: the edge computing network to which the method of the present invention is applied is an edge cloud system, in which there are K heterogeneous edge servers, which are where, each edge server _sk has a limited capacity C _k ; the data transmission rate from edge servers _si to s _j is d _ij , set d _ij =d _ji ; the edge cloud system includes a remote cloud server s _K .

(12) Task dependency graph: One computing module in each application is a task, and multiple tasks of each application have dependencies, and set multiple task sets of one application Among them, the running time of task v _j on the edge server _sk is denoted by t _jk ; a directed acyclic graph (DAG) To represent an application, set the directed edge e:=(vi, v _j )∈ε to indicate that the execution of task v _j requires the result of task v _{i ,} as input, where the amount of data transmitted is w _ij .

(13) Server configuration: task vi can only be run on edge servers configured with corresponding functions _, and it takes time r _ij to configure task _vi on edge server s _j ; set the capacity of each function to account for one unit, The edge server _si can be configured with at most C _i functions at a time. The default server will open an instance (also called a thread) for the configured function to process the corresponding task; if there is not enough capacity on the edge server to configure the new function, Then a decision needs to be made to throw away some configured functions, and at the same time terminate the corresponding instance of the thrown function.

(14) Directed acyclic graph simplification: In order to simplify the application representation, add an empty node, and generate directed edges pointing to all entry nodes (in-degree 0) in the directed acyclic graph, and the amount of data transmission on these edges Initialize the amount of data for the application, and another empty node is used to collect the results of all end nodes (out-degree 0); empty nodes do not occupy any capacity and have no execution time, and two empty nodes can only be placed on the same edge server. , indicating that the initialization server of the task request is also the edge server that receives the final execution result. This step is added because the user's mobile device generally unloads the computing application to the edge server closest to it (referred to as the initialization server in this application), and at the same time sends the initial input data of the application to the initialization server. When scheduling , if the initialization server decides to place the entry node (that is, the task that requires initial input data as input) on another edge server, it is necessary to consider the communication delay caused by the transmission of the initial input data from the initialization server to other edge servers. Conversely, the final result is also returned to the user's mobile device by the initialization server, which simplifies and characterizes the process of the directed acyclic graph.

The specific processing of the method includes the following steps:

Step 1: Enter the information of each parameter in the model definition, including: network, directed acyclic graph, the running time of each task on each server, and the time it takes for each edge server to configure different functions, and specify the entry node of the application the server on which it resides.

Step 2, using the information in the above step 1, specifically the running time information of each task on different servers, perform greedy initial configuration on the edge servers in the network:

Step 21, without considering the actual capacity constraints of edge servers, greedily ensure that each task is configured on the server with the least running time, and calculate the maximum capacity cost among all servers (that is, the maximum cost on the server). number of functions configured), this value is called virtual capacity;

Step 22: Set the capacity of all edge servers as virtual capacity, and continue to configure the edge servers that have not reached full configuration (that is, the number of configured functions is less than the virtual capacity) in step 21 above: run all tasks on all edge servers. The time is sorted from small to large, and it is judged in turn whether the edge server _sk corresponding to the running time _{t jk} is fully configured. If the function corresponding to _i has been configured, skip the next steps, and then skip to the next running time to judge until all edge servers are fully configured. This configuration also ensures that each task has an executable edge server (this is a necessary condition for the operation of step 4) when on-demand configuration is not allowed (that is, when the function of the server configuration does not change);

Step 3: Topological sorting is performed on the directed acyclic graph input in the above step 1 to obtain the topological sequence of the tasks; the parameter f _ij is defined to represent the earliest end time of the task vi placed on the server _{s j .} Initialized to f _ij :=∞, for all 1≤i≤J and 1≤j≤K;

Step 4, using the server configuration information of the above step 2 and the definition of step 3, iterate according to the topology sequence, and use the method of dynamic programming to obtain the value of f _ij step by step, Until the completion time of the last task f _Ja is calculated, and the assignment and scheduling scheme of the task is obtained;

Step 5: Under the constraint of the actual capacity of the server, the tasks of the application are allocated and scheduled according to the result of step 4, and the real completion time of the application is calculated. Considering that the actual capacity is often smaller than the virtual capacity, for some tasks that can be run directly after satisfying the directed acyclic graph dependency constraint under the virtual capacity, additional queuing waiting time and function configuration time are required under the actual capacity limit. The task completion time is not f _Ja obtained in step 4, but f _Ja plus some tasks due to queue waiting time and function configuration time. Among them, the queue waiting time is the time when a task can run under the time constraint of the directed acyclic graph, and it needs to wait for an idle instance (not processing any task) on the server where the task is located. The function configuration time refers to the time required to configure the function corresponding to the task and replace the function corresponding to the idle instance.

Example

The method for scheduling tasks with dependencies configured on demand in combination with functions in edge computing provided by the embodiment of the present invention specifically includes the following steps:

Step 1, input the edge computing network set in Figure 2 and the directed acyclic graph of the three applications in Figure 3 give the edge server s _a initialized by the application;

Step 2: Use the information in Step 1 above to perform a greedy initial configuration of the edge servers in the network, ignoring the capacity of the edge servers, and greedily ensure that each task is configured on the server with the least running time, and set the virtual capacity value as All servers spend the maximum value of capacity and set the capacity of the edge server to be the virtual capacity value. Under this virtual capacity setting, edge servers with remaining capacity that can be configured greedily and non-repeatedly select the task of configuring the minimum running time until all servers are fully configured. This configuration also ensures that each task has an executable server without allowing on-demand configuration;

Step 3: Perform topological sorting on the directed acyclic graph in Step 1 above to obtain a topological sequence of tasks. The parameter f _ij is defined to represent the earliest finish time of the task vi placed on the server _{s j .} Initialized to f _ij :=∞, for all 1≤i≤J and 1≤j≤K;

Step 4, use the server configuration information of 2 and the definition of 3, iterate according to the topology sequence, and use the method of dynamic programming to obtain the value of f _ij step by step, Until the completion time of the last task f _Ja is calculated, and the assignment and scheduling scheme of the task is obtained;

Step 5: Using the task allocation and scheduling sequence scheme obtained in the above step 4, recalculate the application completion time in the actual edge computing network given in the above step 1. If a certain function is missing when the task is running, configure it as needed, and add it. The actual completion time can be calculated after the above configuration and the waiting time in the queue.

In the edge computing network shown in FIG. 2 above, an edge cloud system composed of three edge servers and a remote cloud is described, in which the edge servers are allocated and configured with some functions, an application is initialized in the edge server s ₁ , and the DAG is determined by the present invention. In the figure, tasks 1 and 3 are placed to run on s ₁ , while task 2 is placed on s ₂ and task 4 is placed on s ₃ . Since the function corresponding to task 2 is not configured on s ₂ , when task 2 runs, server s ₂ needs to download the function from the cloud and replace the function corresponding to task 1.

Fig. 4 shows a performance comparison diagram of the method of the present invention; in Fig. 4, the abscissa "chain query" corresponds to a query chain application, "Video Processing" corresponds to a video processing application, and "CDA" corresponds to a complex data analysis application. The algorithm ALG-ODM proposed in the present invention reduces the completion time by at least 2.8, 2.28, 1.54 times under these three applications.

The edge computing network configuration applied in this embodiment is as follows:

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the program can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. The storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM) or the like.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. Substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (6)

1. what binding function configured on demand has the dispatching method of dependence task in a kind of edge calculations, which is characterized in that packet It includes:

Step 1, edge calculations network and the relevant parameter comprising the application with dependence task are obtained, from edge calculations net The initialization server that an Edge Server is output and input as the processing application is chosen in network；

Step 2, the relevant parameter of the application obtained using the step 1 is to each Edge Server in the edge calculations network It carries out greedy initial configuration and obtains server configuration information；

Step 3, the task with dependence of the application in the step 1 is indicated with directed acyclic graph, and to described oriented Task in acyclic figure carries out topological sorting, obtains the topological sequences of task；

Step 4, the server configuration information obtained using the step 2 changes to the topological sequences for the task that the step 3 obtains In generation, calculates, and each task in topological sequences that calculates is placed on operation on each Edge Server of edge calculations network to be terminated earliest Deadline and store corresponding assigning process, respectively distributed according to what the last one task completion time reverse search was stored Journey reversely rebuilds distribution and the scheduling scheme of all tasks；

Step 5, under the constraint of Edge Server actual capacity, according to the distribution and scheduling of the task that the step 4 finally determines Scheme is allocated and dispatches to each task.

2. what binding function configured on demand has the dispatching party of dependence task in edge calculations according to claim 1 Method, which is characterized in that in the step 1 of the method, edge calculations network includes:

The Edge Server of one long-distance cloud and multiple isomeries, each Edge Server have limited capacity, wherein any two Bidirectional data transfers rate between Edge Server is equal.

3. what binding function configured on demand has the scheduling of dependence task in edge calculations according to claim 1 or 2 Method, which is characterized in that in the step 1 of the method, obtain edge calculations network and include answering with dependence task Relevant parameter includes:

Runing time and each Edge Server of each task on each Edge Server of edge calculations network configure different function The time that can need to be spent.

4. what binding function configured on demand has the scheduling of dependence task in edge calculations according to claim 1 or 2 Method, which is characterized in that the relevant parameter for the application of the step 2 of the method obtained using the step 1 is to the edge The greedy initial configuration of Edge Server progress calculated in network obtains in server configuration information, sets on each Edge Server Set the number of a corresponding array record configuration feature, configuration process the following steps are included:

Step 21, under the premise of ignoring the actual capacity of Edge Server, greedy ensures each task in its runing time Good corresponding function is configured on least Edge Server, and the number of function is recorded in the array of the Edge Server, and It calculates under current-configuration, maximum capacity cost value is as virtual capacity in Edge Server；

Step 22, the capacity of all Edge Servers is set as virtual capacity, to the edge service of non-full configuration in the step 21 Device continues to configure as follows: the runing time by all tasks in all Edge Servers sorts from small to large, successively judges Whether full configuration if full configuration skips subsequent step on Edge Server corresponding to runing time, if non-full configuration, judgement Whether the corresponding function of configuration task if configured skips subsequent step to the Edge Server of non-full configuration, otherwise, carries out Configure and be saved in after array and jump to next runing time again and judged, until all Edge Servers match it is full.

5. what binding function configured on demand has the dispatching party of dependence task in edge calculations according to claim 4 Method, which is characterized in that in the step 22, non-full configuration refers to that the function number of configuration is less than the virtual capacity of Edge Server；

Full configuration refers to that the function number of configuration is equal to the virtual capacity of Edge Server.

6. what binding function configured on demand has the scheduling of dependence task in edge calculations according to claim 1 or 2 Method, which is characterized in that in the step 3 of the method, indicate that having for the application in the step 1 relies on directed acyclic graph The task of relationship includes:

In task with dependence, a computing module in the application of each task presentation, by appointing with dependence Business is set as a set of tasks v={ v₁, v₂..., v_J, wherein task v_jIn Edge Server s_kRuning time be t_jk；

With a directed acyclic graphIt indicates an application, sets directed edge e:=(v_i, v_j) ∈ ε expression task v_j's Execution needs task v_i, result as input, wherein the data volume transmitted be w_ij。