Joint optimization method for DNN partition and resource allocation

Abstract

The invention discloses a joint optimization method for DNN partition and resource allocation, which realizes collaborative reasoning and resource optimization of a deep neural network in an environment of a single edge server and a plurality of mobile devices, and comprises the following steps of: determining DNN hierarchy segmentation points of each device and the number of computing resource blocks and wireless resource blocks purchased from the edge server; decoupling modeling is performed on a DNN structure and an edge environment through a heterogeneous graph attention network, an optimal partition point and a resource deployment strategy are determined through Actor-Critic under the constraint of an M / M / 1 queuing theory model, and an edge-end cooperative reasoning scheme with price-demand-time delay closed-loop optimization is formed.

Description

A joint optimization method for DNN partitioning and resource allocation Technical Field

[0001] This invention belongs to the field of mobile edge computing technology, specifically relating to a joint optimization method for DNN partitioning and resource allocation. Background Technology

[0002] The demand for low latency and high reliability inference is growing rapidly in mobile AI applications (such as AR / VR, vehicle-to-everything (V2X) and industrial IoT).

[0003] Traditional methods such as all-local inference are limited by terminal computing power and energy consumption, resulting in high tail latency; all-edge inference causes severe congestion and queuing when there are multiple users concurrently and uplink is limited. Collaborative inference balances computing and communication overhead by partitioning the DNN between the device and the edge.

[0004] In the existing joint optimization framework of single ES and multiple MDs, edge servers sell computing / spectrum resources according to CRB / RRB, and MDs select the location of partitioning points and the number of CRBs / RRBs purchased based on DNN layer characteristics, channel conditions, task arrival rate, and price. In this scenario, the price set by the edge server will affect the procurement of mobile devices and the partitioning of the DNN, thereby changing the queue and throughput, which in turn affects the price and supply relationship.

[0005] Existing optimization methods are often limited to optimizing single-stage processes, such as optimizing only partitioning strategies or resource allocation strategies, ignoring the coupling relationships between these factors. At the same time, due to the increasingly complex structure of DNNs, existing DNN networks include residuals and branching structures, and their computation graphs are typical directed acyclic graphs (DAGs), with complex inter-layer dependencies, making it difficult to model the intermediate activation scale and execution time in a closed loop.

[0006] Most current optimization techniques are designed for chain-structured DNNs and cannot adapt to DNNs with directed acyclic graph (DAG) structures. At the same time, as the scale of mobile devices and DNN models increases, the cost of searching for the globally optimal deployment scheme is unacceptable for latency-sensitive applications. Summary of the Invention

[0007] The technical problem that the present invention needs to solve is: 1. How to uniformly calculate the computational amount on both sides of a partition and the data transmission amount across partitions in a complex DAG structure and dynamic heterogeneous environment, so as to avoid the distortion and surge in communication overhead caused by single-point partitioning.

[0008] 2. How to unify partitioning decisions, CRB / RRB pricing mechanisms, and procurement strategies into a closed-loop framework when multiple users share a single edge server, so as to minimize end-to-end latency and optimize the utility of servers and mobile devices.

[0009] 3. How to ensure rapid decision convergence while satisfying queue stability and resource constraints under conditions of random task arrival and dynamic channel fluctuations, and achieve robustness and scalability in non-stationary environments.

[0010] 4. How to solve problems such as latency instability, bandwidth / computing power bottlenecks, and resource pricing imbalance caused by concurrent intelligent application tasks in multi-user scenarios.

[0011] To address the aforementioned technical problems, this invention provides a joint optimization method for DNN partitioning and resource allocation, comprising the following steps: constructing a DNN graph using DNN layers as nodes and data between layers as edges; assigning values ​​to each node based on local execution and edge execution to obtain local subgraphs and edge subgraphs; calculating the amount of intermediate data transmitted across the network; and calculating the intermediate data volume based on task arrival rate, local computing power, edge computing power, and node... The computational load, system bandwidth, and channel signal-to-noise ratio are considered to obtain the local computation rate, edge computation rate, and achievable data transmission rate per RRB. Based on the achievable data transmission rate per RRB and the amount of intermediate data transmitted across the network, the outgoing task rate per unit time is obtained. The local average latency and edge latency are calculated, and the total latency is obtained by treating the return result time as a constant. The utility and constraints for users and the edge are defined. Features are constructed for each node, and neighbor node information is aggregated using attention on the DNN graph. Pooling is performed on "candidate cuts" and "currently assigned subgraphs," and then the "candidate nodes" and "currently assigned subgraphs" are pooled. The allocated subgraph is integrated into a supernode, and the resource requirements gathered by the supernode are concatenated with the server status to obtain the complete decision state. Based on the actions of each device and the complete decision state, an Actor-Critic network is used for updating, and a lightweight entropy regularization is added to maintain exploration. A "centralized commentator, decentralized execution" approach is adopted to improve stability, and "action masking" is used to directly exclude schemes that violate queue stability and resource limits. The edge-side pricing is obtained using a two-sided perturbation gradient, forming an edge-end collaborative inference scheme with price-demand-latency closed-loop optimization, realizing the joint optimization of DNN partitioning and resource allocation.

[0012] Preferably, the node characteristics of the node include layer type, number of parameters, and multiply-accumulate operation, and the edge characteristics include the size of transmitted data.

[0013] Preferably, the formula for the amount of intermediate data transmitted across the network is as follows:

[0014] In the formula, For nodes The amount of output data, For local subgraphs, This is the edge subgraph.

[0015] This invention provides a joint optimization method for DNN partitioning and resource allocation. By implementing collaborative inference and resource optimization of deep neural networks in a single edge server and multiple mobile device environments, it includes determining the DNN layer partitioning point for each device and the number of computing and wireless resource blocks purchased from the edge server. The DNN structure and edge environment are decoupled and modeled through a heterogeneous graph attention network. Under the constraints of the M / M / 1 queuing theory model, the optimal partitioning point and resource deployment strategy are determined through Actor-Critic, forming an edge-end collaborative inference scheme with price-demand-latency closed-loop optimization. Attached Figure Description

[0016] Figure 1 shows a mobile edge computing system model with a single edge server and multiple mobile devices executing multiple DNN partitions; Figure 2 shows a flowchart of the joint optimization method. Detailed Implementation

[0017] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0018] As shown in Figures 1 and 2, an embodiment of the present invention provides a joint optimization method for DNN partitioning and resource allocation, comprising the following steps: 1) DAG modeling and feasible partitioning: representing the DNN as a graph Here, nodes represent layers in a DNN, and edges represent data dependencies between layers. Node features include layer type, number of parameters, and multiply-accumulate (MAC) operations, while edge features include the size of the transmitted data. For each node... Assignment:

[0019] To ensure forward executability, it is required that Based on this, the local subgraph is obtained. With edge subgraph The amount of intermediate data that needs to be transferred across the network is:

[0020] in, For nodes The amount of output data.

[0021] 2) Delay and utility modeling assumes the task arrival rate is Local computing power Edge computing power ,node The computational cost is System bandwidth Channel signal-to-noise ratio The purchased share of edge computing and wireless resources were respectively... This refers to the amount of "available quota" that a mobile device can obtain from the edge server.

[0022] The calculation rate of each step can be defined as:

[0023] The achievable data transfer rate for each RRB is defined as follows:

[0024] Therefore, the outgoing task rate per unit time is:

[0025] The average time delay of each stage can be approximated using M / M / 1:

[0026] Time of return result Let be a constant, then the total time delay can be defined as:

[0027] User and edge utility ( As weight, For price, (for cost)

[0028]

[0029] constraint: .

[0030] 3) Graph attention state representation constructs features for each node. Aggregate neighbor node information on the graph using attention:

[0031] in, To activate, pooling is performed on the "candidate cut" and the "currently assigned subgraph" to obtain... Then, integrate the "candidate nodes" and "currently allocated subgraphs" into a supernode, and concatenate the resource requirements aggregated in the supernode with the server status to obtain the complete decision state:

[0032] This reflects both the model structure and the real-time environmental state.

[0033] 4) Multi-agent Actor-Critic and pricing: the actions of each device are... ,in Forward feasibility must be satisfied. To ensure differentiability and feasibility, Gumbel-Softmax sampling is used first, followed by projection onto... Updates are performed using the Actor-Critic network:

[0034] Lightweight entropy regularization is incorporated to maintain exploration. Training employs a "centralized commentator, distributed execution" approach to improve stability; "action masking" is used to directly exclude schemes that violate queue stability and resource limits.

[0035] Edge-side pricing uses a two-sided perturbation gradient:

[0036] Prices adjust smoothly according to demand, avoiding fluctuations.

[0037] The beneficial effects of the embodiments of the present invention are as follows: 1. It can be applied to more general DNN networks, avoiding high transmission overhead at large feature maps.

[0038] 2. Graph attention mechanisms can improve the clarity of structure-environment states, and the Actor-Critic algorithm can quickly find high-quality solutions in a large action space.

[0039] 3. Joint optimization of pricing and strategies can maintain user experience and edge server revenue under complex and fluctuating channel conditions.

[0040] This invention implements collaborative inference and resource optimization of deep neural networks (DNNs) in a single edge server (ES) and multiple mobile devices (MDs) environment. This includes determining the DNN layer splitting point for each device and the number of compute resource blocks (CRBs) and wireless resource blocks (RRBs) purchased from the edge server. The DNN structure and edge environment are decoupled and modeled using a heterogeneous graph attention network (HGAT). Under the constraints of the M / M / 1 queuing theory model, the optimal partitioning point and resource deployment strategy are determined by the Actor-Critic model, forming an edge-end collaborative inference scheme with price-demand-latency closed-loop optimization.

Claims

1. A joint optimization method for DNN partitioning and resource allocation, characterized in that, The process includes the following steps: Constructing a DNN graph using layers as nodes and data between layers as edges; assigning values ​​to each node based on local and edge execution to obtain local and edge subgraphs; calculating the amount of intermediate data transmitted across the network; and calculating the intermediate data volume based on task arrival rate, local computing power, edge computing power, and node... The computational load, system bandwidth, and channel signal-to-noise ratio are used to obtain the local computation rate, edge computation rate, and data transmission rate achievable per RRB. Based on the data transmission rate achievable per RRB and the amount of intermediate data transmitted across the network, the outgoing task rate per unit time is obtained. The local average latency and edge latency are calculated. The total latency is obtained by treating the return result time as a constant. The utility and constraints of users and the edge are defined. Features are constructed for each node, and neighbor node information is aggregated using attention on the DNN graph. Pooling is performed on "candidate cuts" and "currently allocated subgraphs," and then "candidate nodes" and "currently allocated subgraphs" are integrated into supernodes. The resource requirements aggregated by the supernodes are concatenated with the server status to obtain the complete decision state. Based on the actions of each device and the complete decision state, an Actor-Critic network is used for updating, and lightweight entropy regularization is added to maintain exploration. "Centralized commentators and distributed execution" are adopted to improve stability. "Action masking" is used to directly exclude schemes that violate queue stability and resource limits. The edge-side pricing uses a two-sided perturbation gradient to form an edge-end collaborative inference scheme with price-demand-latency closed-loop optimization, realizing the joint optimization of DNN partitioning and resource allocation.

2. The joint optimization method for DNN partitioning and resource allocation as described in claim 1, characterized in that, The node characteristics include layer type, number of parameters, and multiply-accumulate operations, and the edge characteristics include the size of the transmitted data.

3. The joint optimization method for DNN partitioning and resource allocation as described in claim 1, characterized in that, The formula for the amount of intermediate data transmitted across the network is as follows: In the formula, For nodes The amount of output data, For local subgraphs, This is the edge subgraph.

Images