Task-aware service placement method for distributed learning in wireless edge network

Abstract

The invention discloses a task awareness service placement method for distributed learning in a wireless edge network. According to a service directed acyclic graph with resource and communication demand information obtained by merging original task directed acyclic graphs, services of distributed learning applications are placed, and the step comprises the following two stages: a first stage, a stage of searching for placed server nodes for each merged independent service directed acyclic graph; the second stage is an overall delay stage for evaluating a given service placement scheme; and by iterating the two steps, an optimal distributed learning application service placement scheme is found. Aiming at the problems of unreliable communication, random user movement and service queuing, the invention provides a task-aware service placement method considering the dependency relationship between placement services, the expected delay of all learning tasks is reduced, and the task completion rate of an edge network is improved. The method is suitable for the edge computing field.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things and edge computing, and in particular relates to a task-aware service placement method for distributed learning in a wireless edge network. Background Art

[0002] With the great progress of machine learning technology, more and more large-scale computing services and applications have been developed, such as intelligent network management and face recognition. Due to the large amount of data and intensive computing involved, machine learning algorithms are difficult to use on resource-constrained devices. Although edge computing can provide additional computing and storage resources for resource-constrained front-end devices, it requires users to transfer training data to centralized edge servers, which leads to privacy issues. To solve this problem, Distributed Learning in Wireless Edge Networks (DLWENs) came into being and received extensive research attention.

[0003] In the framework of DLWENs, edge servers act as "worker" nodes, run learning services, and collaborate to train models to handle offloaded tasks. Learning tasks are divided and assigned to multiple "worker" nodes according to data partitioning or function partitioning methods. When implemented in edge computing, subtasks located on different edge servers communicate with each other and collaborate to complete learning tasks. Obviously, to process specific subtasks, edge servers need to be placed in the corresponding services. However, since each edge server can only instantiate a certain number of services, the placement of services in wireless edge networks becomes a key issue for distributed learning performance.

[0004] Unlike service placement in traditional edge networks, service placement in distributed learning has the following obvious challenges. First, there are dependencies between the placed services to be considered. For example, an edge server running a training service depends on the output of a server with a data cleaning service. Therefore, placement not only needs to consider the resources of the service, but also the overhead of inter-service communication. Second, due to the unreliability of wireless communications and the randomness of front-end user mobility, learning tasks exhibit high uncertainty, which further affects service placement decisions. Third, at the system level, given the fixed number of service instances, multiple learning tasks may share the same type of functionality handled by the same service (e.g., data cleaning), which cannot be processed immediately, which will cause queuing delays, which is also ignored by existing work.

[0005] Based on this, a task-aware service placement method for distributed learning applications in wireless edge network scenarios that can consider the dependencies between placed services, is robust to communication unreliability and user mobility randomness, and reduces queuing delays is urgently needed in the field. Summary of the invention

[0006] The present invention provides a task-aware service placement architecture and method for distributed learning applications in wireless edge networks. By considering factors such as the dependencies between placed services, communication unreliability, user mobility randomness, and multiple task competition queuing, the architecture and method can solve the placement problem of distributed learning task services under complex subtask topology and unreliable wireless edge networks, thereby reducing the overall task delay.

[0007] The present invention is achieved through the following technical solutions:

[0008] A method for placing task-aware services for distributed learning in a wireless edge network comprises the following steps: step S1: acquiring global service directed acyclic graph information by merging task directed acyclic graphs according to all task directed acyclic graphs; step S2: searching for a placement scheme for the merged directed acyclic graph, i.e. searching for a suitable location for each service in the wireless edge network, wherein step S2 comprises the following two stages: the first stage is a stage of searching for a server node for placement for each independent service directed acyclic graph after merging; the second stage is a stage of evaluating the overall delay of a given service placement scheme; step S3: completing the placement of distributed learning task services in a target wireless edge network according to the placement scheme obtained by the placement method in the aforementioned stages.

[0009] In a certain embodiment, the task directed acyclic graph is merged, and the ownership values ​​of the same vertices and edges in the directed acyclic graph are accumulated into a merged service directed acyclic graph.

[0010] In one embodiment, in the first stage, a server that hosts a root service is first selected, and then a server that hosts subsequent services in the directed acyclic graph is selected.

[0011] In one embodiment, in the first stage, a server is randomly selected from the root server set for the root service; after the root node is determined, a greedy scheme is used to find a server that can host subsequent nodes in the directed acyclic graph. That is, for each subsequent service, an edge server with the maximum transmission rate with the server hosting the root service is found.

[0012] In one embodiment, in the first stage, after all subsequent services are placed, if there are multiple independent service directed acyclic graphs after merging, it is possible to move to the next root service and repeat the above iteration.

[0013] In one embodiment, in the second stage, by estimating the overall task delay of a specific service placement solution, the service placement in the iteration is updated to obtain a placement solution with the minimum overall task delay, which specifically includes the following steps:

[0014] Step 3.1) Evaluate the overall expected delay of a given placement solution. Specifically, calculate the task queuing delay, processing time, and communication delay of each path in the target edge network to obtain the longest path delay in the directed acyclic graph, which is the overall delay of the placement solution.

[0015] In a certain embodiment, the queuing delay of each service and its task queue is calculated, specifically: the queuing delay problem is modeled as an M / M / 1 queuing problem; for the queuing problem, the queuing delay is solved based on queuing theory and Poisson process to obtain the queuing delay.

[0016] Step 3.2), service placement plan update.

[0017] In a certain embodiment, when the service expected delay is reduced compared to the placement in the previous iteration, in order to avoid direct update causing local optimality, the entire placement solution has a chance to be updated to the placement setting in the current iteration with a certain probability.

[0018] In an embodiment, the probability that the placement solution is updated to the most recently generated location is related to the difference in expected latency between the current iteration and the previous iteration.

[0019] In one embodiment, the above iterations will continue until a preset timer is triggered.

[0020] The present invention completes the placement of distributed learning task services in the target wireless edge network according to the placement scheme obtained by the placement method in the aforementioned stages.

[0021] The present invention has the following advantages and beneficial effects:

[0022] 1. Aiming at the task queuing problem caused by communication unreliability, user mobility randomness and task uncertainty in the placement of distributed learning services in wireless edge networks, the present invention proposes to use Poisson process and queuing theory (M / M / 1 model) to accurately estimate the expected task delay of a given service placement scheme, thereby reducing the overall delay of distributed learning tasks.

[0023] 2. The present invention proposes a method that follows Gibbs sampling and iteratively approximates the optimal solution to the service placement search problem. By extracting the dependencies between functions and communication overhead in the distributed learning task structure, fine-grained task requests and dependencies between tasks are incorporated into the placement strategy, thereby realizing parallel computing of learning services.

[0024] Instruction Manual

[0025] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention.

[0026] Figure 1 Examples of service placement for distributed learning applications in wireless edge networks;

[0027] Figure 2 An example of the steps of merging the directed acyclic graph of tasks proposed by the present invention;

[0028] Figure 3 Schematic diagram of the model for overall service placement delay evaluation proposed in the present invention. DETAILED DESCRIPTION

[0029] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0030] First, the task-aware service placement architecture for distributed learning applications in wireless edge networks in the present invention is introduced. The architecture of this embodiment explicitly considers the dependencies between distributed subtasks, the task uncertainty in the wireless edge network, and the queuing process when multiple subtasks are queued on a server, thereby coping with complex subtask topologies and unreliable wireless edge networks, reducing the expected end-to-end delay of distributed learning tasks and improving the task completion rate of the edge network.

[0031] In an embodiment of the present invention, distributed learning in a wireless edge network scenario has the following specific features:

[0032] 1. The divided data and models / functions are processed by services running on edge servers. Each service corresponds to a specific type of subtask. To improve the efficiency and resource utilization of the edge network, a set of service images are placed on each edge node.

[0033] 2. Directed acyclic graphs are often used to represent task structures and subtask resource requirements. In order to process all nodes in the directed acyclic graph representing computing tasks, the same number of services need to be placed on the edge network.

[0034] Figure 1 A service placement diagram for distributed learning applications in wireless edge computing is shown. The task directed acyclic graph and the service topology on the edge network infrastructure can be very different. One edge server can host multiple services, and each service can perform multiple subtasks from different learning tasks. Therefore, another complete service chain that performs the same task with different data blocks can be placed in the same edge network.

[0035] The task-aware service placement method iteratively updates the placement of distributed learning services in the server cluster, so that the server can achieve the goal of minimizing the expected delay of the overall learning task while covering all service tasks. This method specifically includes the following steps:

[0036] Step S1: According to all task directed acyclic graphs, global service directed acyclic graph information is obtained by merging task directed acyclic graphs.

[0037] Merge Process Reference Figure 2 , where the arrival rates of tasks T1 and T2 are characterized by Poisson parameters λ1 and λ2. Since the vertex weights in the directed acyclic graph represent the service workload and the edge weights represent the communication requirements, for each vertex (i.e., subtask), its incoming and outgoing edges from all original task directed acyclic graphs to the merged directed acyclic graph are stored; then all weights of the same vertices and edges in the directed acyclic graph are accumulated into the merged service directed acyclic graph. The service directed acyclic graph thus obtained contains all types of services, the resource requirements of each service, and the communication overhead.

[0038] like Figure 2 As shown, the resource requirements of the merged public services B, D and E, and the link communication requirements of the subtasks B to D and D to E increase accordingly. Then, step S2 is executed.

[0039] Step S2: Search for a service placement solution for the merged directed acyclic graph, that is, search for a suitable placement location for each service in the wireless edge network. This step uses a Gibbs sampling scheme based on service dependencies to efficiently search for results, and is divided into the following two stages:

[0040] The first stage is to iteratively search for the server nodes to be placed for each independent service directed acyclic graph after merging. Specifically, it includes:

[0041] Step 2.1) Select the server that hosts the root service.

[0042] It is necessary to find a server that can host the root service. To this end, a server is randomly selected from the root server set to serve the root node in the directed acyclic graph, where each server saves the request for the root node.

[0043] Step 2.2) Select a server that hosts subsequent services in the directed acyclic graph.

[0044] After determining the root node, we need to find servers that can host subsequent nodes in the directed acyclic graph. Here, we use a greedy solution to place them. Specifically, for each subsequent service, we find the edge server that has the maximum transmission rate with the server that hosts the root service.

[0045] When the above steps are completed and all subsequent services are placed, if there are multiple independent service directed acyclic graphs after merging, it will move to the next root service and repeat the above iteration.

[0046] The second stage is the overall latency evaluation stage of the service placement solution based on the performance model. That is, the performance evaluation is performed on the distributed learning service placement to estimate the overall latency under a specific placement solution; if the latency of the service placement is reduced compared to the previous iteration, the entire placement solution has the opportunity to be updated to the placement settings in the current iteration. Specifically, it includes:

[0047] Step 3.1) Estimate the overall expected latency for a given placement solution.

[0048] refer to Figure 3 ,To evaluate the placement solution delay, it is necessary to calculate the longest path delay in the directed acyclic graph, as follows:

[0049] Since each task is further divided into multiple subtasks characterized by a directed acyclic graph, the task delay essentially represents the delay of completing all subtasks in the directed acyclic graph of the task, that is, the longest path delay in the directed acyclic graph. Here, it is necessary to check all paths and find the delay of the critical path. In order to calculate the path delay, the queuing delay, processing time and communication delay of the task are further extracted from the user / task set and the target edge network. The communication delay can be derived from the output workload of the vertex and the expected transmission rate of the link;

[0050] Due to the uncertainty of tasks and competition with other subtasks from different users, tasks will experience queuing delays. For each service and its task queue, its queuing delay problem is modeled as an M / M / 1 queuing problem; for this queuing problem, the queuing delay is solved based on queuing theory and Poisson process.

[0051] Step 3.2), service placement plan update.

[0052] When the service expected latency decreases compared to the placement in the previous iteration, in order to avoid direct updates causing local optimality, the entire placement solution has the opportunity to be updated to the placement settings in the current iteration. When the overall task latency decreases, the probability that the placement solution is updated to the latest generated location is related to the difference in expected latency between the current iteration and the previous iteration, and the greater the latency reduction, the greater the probability value.

[0053] The above iterations will continue until a pre-set timer is triggered. The timer is set based on the requirements of the network operator. If the placement is for long-term maintenance, the timer can be set very large, allowing thousands of iterations. Otherwise, for short-term placements, the timer should be set much smaller and the resulting solution may not achieve a significant improvement.

[0054] Step S3: According to the placement plan obtained by the placement method in the previous stage, the placement of the distributed learning task service in the target wireless edge network is completed.

[0055] In the above scheme, the contents that are not specifically introduced, such as modeling as an M / M / 1 queuing problem, solving based on queuing theory and Poisson process, etc., themselves (but not the entire scheme constituted by them) are implemented as conventional technical means in the field, and the present invention will not elaborate on them.

[0056] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A task-aware service placement method for distributed learning in a wireless edge network, characterized in that: The following steps are involved: Step S1: According to all task directed acyclic graphs, global service directed acyclic graph information is obtained by merging task directed acyclic graphs; Step S2: searching for a placement solution for the merged directed acyclic graph, that is, searching for a suitable location for each service in the wireless edge network. Step S2 includes the following two stages: The first stage is to search for the server nodes to be placed in the directed acyclic graph of each independent service after the merger; The second phase is to evaluate the overall latency of a given service placement solution; Step S3: According to the placement plan obtained by the placement method in the previous stage, the placement of the distributed learning task service in the target wireless edge network is completed.

2. The method for task-aware service placement for distributed learning in a wireless edge network according to claim 1, characterized in that: The merging of the directed acyclic graphs is performed by accumulating ownership values ​​of the same vertices and edges in the directed acyclic graphs into a merged service directed acyclic graph.

3. The method for task-aware service placement for distributed learning in a wireless edge network according to claim 1, characterized in that: In the first stage, a server that hosts a root service is first selected, and then a server that hosts subsequent services in the directed acyclic graph is selected.

4. The method for task-aware service placement for distributed learning in a wireless edge network according to claim 3, characterized in that: In the first stage, a server is randomly selected from the root server set for the root service; after the root node is determined, for each subsequent service, a greedy scheme is used to select an edge server with the maximum transmission rate with the server hosting the root service.

5. The method for task-aware service placement for distributed learning in a wireless edge network according to claim 4, characterized in that: In the first phase, after all subsequent services are placed, if there are multiple independent service directed acyclic graphs after merging, move to the next root service and repeat the iteration.

6. The method for placing task-aware services for distributed learning in a wireless edge network according to any one of claims 3 to 5, characterized in that: In the second stage, by estimating the overall task delay of a specific placement solution, the placement in the iteration is updated to obtain a placement solution with the minimum overall task delay, which specifically includes the following steps: Step 3.1, evaluate the overall expected delay of a given placement solution; specifically, calculate the task queuing delay, processing time, and communication delay of each path delay in the target edge network to obtain the longest path delay in the directed acyclic graph; Step 3.2: Service placement solution update. To avoid local optimality caused by direct update, the entire placement solution has a chance to be updated to the placement setting in the current iteration according to probability. The above iteration will continue until a pre-set timer is triggered.

7. The method for task-aware service placement for distributed learning in a wireless edge network according to claim 6, characterized in that: In step 3.1, due to the uncertainty of the task and the queues caused by competition with other subtasks from different users, for each service and its task queue, the queuing delay is calculated, specifically: the queuing delay problem is modeled as an M / M / 1 queuing problem; for this queuing problem, the queuing delay is solved based on queuing theory and Poisson process.

Images