End-side cloud storage computing resource collaborative optimization method and device

Abstract

The invention provides an end-side cloud storage and computing resource collaborative optimization method and device, and the method comprises the steps: constructing a uniform resource description model based on a dynamic quasi-matrix theory for heterogeneous storage resources and heterogeneous computing resources of end nodes, edge nodes and cloud nodes; based on the uniform resource description model, taking total reasoning time delay of task execution on a resource set, total system energy consumption consumed by task execution and a resource utilization rate as optimization objectives, and introducing a sub-modular function to construct a multi-objective optimization model; allocating the task among the end node, the edge node and the cloud node according to the calculation-intensive or storage-intensive characteristics of the task and the resource state of each node reflected by the dynamic weight mechanism; and after the tasks are allocated to the corresponding nodes, according to a solving result of the multi-objective optimization model, allocating storage resources and computing resources of the nodes for processing the tasks. According to the method, the dynamic allocation efficiency of the storage and calculation resources in the end-side cloud collaborative environment is improved.

Description

Technical Field

[0001] This invention relates to the field of edge-cloud technology, and in particular to a method and apparatus for collaborative optimization of edge-cloud storage and computing resources. Background Technology

[0002] With the rapid development of edge computing and cloud computing technologies, edge-cloud collaborative architectures have been widely deployed and explored in complex application scenarios such as the Industrial Internet, intelligent manufacturing, and smart cities. Edge-cloud storage-computing integration is a new computing paradigm that has emerged with the development of IoT, 5G, and AI technologies. It aims to achieve efficient integration and dynamic optimization of storage and computing resources through the collaboration of terminal devices, edge nodes, and cloud data centers. Traditional cloud computing struggles to meet the demands for low latency and high real-time performance due to centralized processing, while edge computing, although alleviating latency pressure, is limited by resource scale. Terminal devices, on the other hand, lack sufficient computing and storage capabilities to independently handle complex tasks. The core of edge-cloud storage-computing integration architecture lies in breaking down the inherent boundaries between storage and computing. In the traditional model, data is frequently moved between storage and computing units, generating significant I / O overhead. Storage-computing integration technology, by deeply integrating storage and computing functions, can reduce time and energy consumption during data transmission and improve computing efficiency. Edge-cloud collaboration further complements the advantages of these three elements: the terminal handles real-time sensing and lightweight computing, edge nodes provide near-end processing and caching, and the cloud supports massive storage and global analysis. However, the edge-cloud integrated storage and computing architecture still faces many challenges in practical applications. In particular, the dynamic allocation of storage and computing resources is a pressing issue. Due to the large number of terminal devices and the complex and ever-changing network environment, how to dynamically adjust the allocation of storage and computing resources according to real-time load conditions to ensure maximum resource utilization efficiency has become an urgent problem to be solved.

[0003] Existing collaborative optimization strategies mostly rely on static resource modeling and centralized scheduling mechanisms, which are insufficient to effectively address real-world challenges such as continuously changing task requirements and frequent network state fluctuations. Specifically, in the multi-layered architecture of edge-cloud with significant resource heterogeneity, there are significant differences in attributes such as computing power, storage, and network connectivity. This makes it difficult for existing solutions to achieve high-precision matching and dynamic load balancing of cross-layer resources, often resulting in resource imbalances where some nodes are idle while others are overloaded. Summary of the Invention

[0004] To address the problem that existing technologies lack optimization capabilities under conditions of strong resource heterogeneity and changing task requirements, this invention proposes a collaborative optimization method and device for edge-cloud storage and computing resources. This method improves the dynamic allocation efficiency of storage and computing resources in an edge-cloud collaborative environment.

[0005] To achieve the above objectives, the present invention provides a method for collaborative optimization of edge-cloud storage and computing resources, comprising:

[0006] For heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes, a unified resource description model is constructed based on dynamic matroid theory. This unified resource description model includes:

[0007] A set of resource elements, which includes the storage and computing resources of the end nodes, edge nodes, and cloud nodes;

[0008] Independent clusters, defined by the resource allocation constraints of the end nodes, edge nodes, and cloud nodes;

[0009] A dynamic weighting mechanism, wherein the dynamic weighting mechanism assigns time-varying weight values ​​based on the real-time resource status of the end nodes, edge nodes, and cloud nodes;

[0010] Based on the unified resource description model, a multi-objective optimization model is constructed by introducing a submodular function with the optimization objectives of the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate.

[0011] The upper-layer scheduler allocates the task among the end nodes, edge nodes, and cloud nodes based on the task's computationally intensive or storage-intensive characteristics and the resource status of each node as reflected by the dynamic weighting mechanism, thereby balancing the load on each node.

[0012] After the task is assigned to the corresponding node, the lower-level scheduler allocates storage and computing resources to that node to process the task based on the solution results of the multi-objective optimization model.

[0013] In one embodiment of the present invention, the construction of the resource element set includes: traversing the hardware resource lists of the end nodes, edge nodes and cloud nodes, extracting storage resource units that meet a preset storage capacity threshold and computing resource units that meet a predetermined computing capability threshold, and forming the resource element set.

[0014] In one embodiment of the present invention, the resource allocation constraint of the independent set family is as follows: for any task, when the total storage resources of the resource subset are not less than the storage requirement threshold of the task, and the total computing resources are not less than the computing requirement threshold of the task, the resource subset belongs to the independent set family.

[0015] In one embodiment of the present invention, the dynamic weighting mechanism calculates the time-varying weight value in the following manner:

[0016] ω_ e(t) =w1·(1-load rate) _e(t) )+w2·(remaining capacity) _e(t) Total capacity _e(t) )

[0017] Where w1 and w2 are preset weighting coefficients, and the load rate is... _e(t) Let t be the CPU or GPU utilization parameter of resource element e at the corresponding node, and the remaining capacity. _e(t) Let t be the parameter of the remaining storage space of resource element e in the corresponding node at time t.

[0018] In one embodiment of the present invention, the submodular function is used to characterize the marginal benefit of each resource in the task scheduling process;

[0019] Combining the marginal gain with the optimization objective, the multi-objective optimization model is constructed as follows:

[0020]

[0021] And satisfy Δ f (e|S)=f(S∪{e})-f(S), f(S)∈{f1(S), f2(S), f3(S)};

[0022] Where α, β, and γ are weighting coefficients, S is the current resource subset, e is the newly added resource, and Δ f (e|S) is the marginal gain function of the resource subset S and satisfies Δ f (e|S) decreases as the size of S increases; f1(S), f2(S), and f3(S) represent the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate, respectively. The latency reduction benefit brought by resource e is α·(f1(S)-f1(S∪{e})), the energy consumption reduction benefit brought by resource e is β·(f2(S)-f2(S∪{e})), and the utilization rate improvement benefit brought by resource e is γ·(f3(S∪{e}-f3(S)).

[0023] In one embodiment of the present invention, computationally intensive tasks are preferentially allocated to edge nodes or cloud nodes where the proportion of GPU resources is not less than a preset threshold; and / or,

[0024] Storage-intensive tasks are preferentially assigned to edge nodes or end nodes with storage capacity not less than a preset capacity threshold.

[0025] In one embodiment of the present invention, when allocating the storage resources, a predetermined percentage threshold is met, where the virtual memory percentage does not exceed the physical memory capacity.

[0026] When allocating computing resources, time-slice round-robin scheduling is used, and the CPU time slice ratio of a single task does not exceed a predetermined threshold of the total number of cores in the node.

[0027] In one embodiment of the present invention, the method further includes:

[0028] The time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model are updated by updating the time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model through a time sliding window, thereby adjusting the task allocation of the upper-level scheduler and the resource allocation of the lower-level scheduler.

[0029] In one embodiment of the present invention, the update scheduling decision mechanism adopted in each period T is as follows:

[0030]

[0031] in, This is the objective function value in the next state.

[0032] In another aspect, the present invention provides an edge-cloud storage and computing resource collaborative optimization device, comprising:

[0033] A resource description model construction module is used to construct a unified resource description model based on dynamic matroid theory for heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes. The unified resource description model includes: a set of resource elements containing the storage and computing resources of the end nodes, edge nodes, and cloud nodes; independent sets defined by resource allocation constraints of the end nodes, edge nodes, and cloud nodes; and a dynamic weighting mechanism that assigns time-varying weight values ​​based on the real-time resource status of the end nodes, edge nodes, and cloud nodes.

[0034] The multi-objective optimization model construction module is used to construct a multi-objective optimization model based on the unified resource description model, with the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate as optimization objectives, and introduces a submodular function.

[0035] The upper-layer scheduling module is used by the upper-layer scheduler to allocate the task among the end nodes, edge nodes, and cloud nodes according to the computationally intensive or storage-intensive characteristics of the task and the resource status of each node reflected by the dynamic weighting mechanism, so as to balance the load of each node.

[0036] The lower-level scheduling module is used to allocate storage and computing resources to the corresponding node to process the task after the task is assigned to the corresponding node, based on the solution results of the multi-objective optimization model.

[0037] As can be seen from the above solutions, the advantages of the present invention are:

[0038] The edge-cloud collaborative optimization method for storage and computing resources provided by this invention firstly constructs a unified resource description model based on dynamic matroid theory for heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes. This dynamic matroid modeling achieves a unified abstraction and constraint representation of heterogeneous storage and computing resources. Secondly, with the total inference latency of tasks executed on the resource set, the total system energy consumption consumed by task execution, and resource utilization as optimization objectives, a submodular function is introduced to construct a multi-objective optimization model, ensuring an approximate optimal solution for multi-objective collaborative scheduling. Furthermore, based on two-layer scheduling and combined with the state changes of edge-cloud resources, it supports adaptive migration of tasks between multiple layers of nodes and resource allocation by nodes based on task requirements. Overall, this invention can significantly improve the scheduling efficiency of storage and computing resources, task processing performance, and the overall response resilience and robustness of the system in complex scenarios facing high-concurrency task flows, highly dynamic network states, and multi-source heterogeneous resource constraints in edge-cloud collaborative systems. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the edge-cloud storage and computing resource collaborative optimization method provided in Embodiment 1 of the present invention.

[0040] Figure 2 This is a flowchart illustrating the edge-cloud storage and computing resource collaborative optimization method provided in Embodiment 2 of the present invention.

[0041] Figure 3 A schematic diagram of the structure of the edge-cloud storage and computing resource collaborative optimization device provided in another embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the structure of an electronic device;

[0043] in:

[0044] 300: Edge-Cloud Storage and Computing Resource Collaborative Optimization Device;

[0045] 310: Resource Description Model Construction Module;

[0046] 320: Multi-objective optimization model construction module;

[0047] 330: Upper-layer scheduling module;

[0048] 340: Lower-level scheduling module;

[0049] 350: Scheduling update module;

[0050] 400: Electronic devices;

[0051] 410: Processor;

[0052] 420: Memory. Detailed Implementation

[0053] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0054] In edge-cloud collaborative computing scenarios, characterized by high concurrency and strong dynamism, traditional static resource allocation strategies suffer from low resource utilization and slow scheduling response when dealing with real-time changing task loads and heterogeneous resources. This invention draws inspiration from dynamic matroid theory, a combinatorial mathematical tool that can flexibly describe the dynamic changes in resource structure and constraints, which aligns perfectly with the needs of in-memory computing collaborative optimization. First, combining the characteristics of storage and computing resources at the edge, cloud, and endpoint layers, a unified description model of heterogeneous resources based on dynamic matroid is constructed, enabling a structured expression and schedulable constraint modeling of heterogeneous resources. Subsequently, a multi-objective collaborative optimization model is constructed, unifying multi-dimensional optimization objectives such as latency, energy consumption, and resource utilization into the optimization model. By introducing a submodular function to characterize the marginal benefits and complementarity of resources during task scheduling, the nonlinear impact of in-memory computing resources on overall performance is accurately reflected. This multi-objective dynamic optimization strategy can improve the overall system's task processing efficiency and save on storage and computing resource overhead. Furthermore, based on a two-layer scheduling optimization algorithm, the upper layer is responsible for task allocation among nodes in the edge, cloud, and endpoint, while the lower layer is responsible for the collaborative management of storage and computing resources. Ultimately, it can optimize scheduling decisions in real time according to the dynamic changes in task requirements and resource status, enabling flexible task migration and elastic scheduling of storage and computing resources among the endpoint, edge, and cloud layers.

[0055] For details, please refer to Figure 1 As shown, Figure 1 The diagram shows a flowchart of the edge-cloud storage and computing resource collaborative optimization method provided in Embodiment 1 of the present invention.

[0056] A method for collaborative optimization of edge-cloud storage and computing resources includes the following steps:

[0057] Step S1: For heterogeneous storage and computing resources of end nodes, edge nodes and cloud nodes, construct a unified resource description model based on dynamic matroid theory.

[0058] The unified resource description model includes: a set of resource elements, independent clusters, and a dynamic weighting mechanism. The set of resource elements includes storage and computing resource units of the endpoint nodes, edge nodes, and cloud nodes; the independent clusters are defined by resource allocation constraints of the endpoint nodes, edge nodes, and cloud nodes; and the dynamic weighting mechanism assigns time-varying weight values ​​based on the real-time resource status of the endpoint nodes, edge nodes, and cloud nodes.

[0059] Step S2: Based on the unified resource description model, with the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate as optimization objectives, a submodular function is introduced to construct a multi-objective optimization model.

[0060] Step S3: The upper-layer scheduler allocates the task among the end nodes, edge nodes, and cloud nodes based on the computationally intensive or storage-intensive characteristics of the task and the resource status of each node as reflected by the dynamic weighting mechanism, thereby balancing the load of each node.

[0061] Step S4: After the task is assigned to the corresponding node, the lower-level scheduler allocates storage and computing resources to the node to process the task based on the solution results of the multi-objective optimization model.

[0062] In this embodiment, by innovatively introducing dynamic matroid theory into resource modeling and task scheduling, the problem of traditional static resource allocation being unable to adapt to dynamic changes is effectively solved. First, dynamic matroid modeling achieves a unified abstraction and constraint representation of heterogeneous in-memory computing resources. Second, submodular optimization theory guarantees an approximate optimal solution for multi-objective collaborative scheduling. Furthermore, based on two-layer scheduling and combined with the state changes of edge-cloud resources, adaptive migration and joint scheduling of tasks across multiple layers of nodes are supported. Overall, this invention significantly improves the scheduling efficiency of in-memory computing resources, task processing performance, and the overall response resilience and robustness of the system in complex scenarios involving high-concurrency task flows, highly dynamic network states, and multi-source heterogeneous resource constraints within the edge-cloud collaborative system.

[0063] In a preferred embodiment, in step S1, considering the characteristics of heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes in an edge-cloud collaborative computing scenario, a dynamic matte M = (E, I) is defined, and the resource element set E of the matte is determined. This resource element set includes all storage resource units and computing resource units in the system. For example, storage resources include memory and hard disk space on different nodes, while computing resources include CPU cores and GPUs. By traversing the hardware resource lists of the end nodes, edge nodes, and cloud nodes, storage resource units that meet a preset storage capacity threshold and computing resource units that meet a predetermined computing capability threshold are extracted to form the resource element set.

[0064] Simultaneously, a family of independent sets I for matroids is defined. Defined by the resource allocation constraints of the end nodes, edge nodes, and cloud nodes, an independent set represents a subset of resources that satisfies certain allocatability constraints. For example, for a task, certain combinations of storage and computing resources can meet its operational requirements; these combinations constitute an independent set. Using the independent set theory of matroids, the allocatability of storage and computing resources is described through a series of rules and conditions. For example, for any task, when the total storage resources of a resource subset are not less than the storage requirement threshold of the task, and the total computing resources are not less than the computing requirement threshold of the task, the resource subset belongs to the family of independent sets. Assume that task i's storage resource requirement is... The demand for computing resources is Resource subset If the total amount of storage resources in A is greater than or equal to And the total amount of computing resources is greater than or equal to Then A is task T i An independent set.

[0065] Furthermore, the dynamic weighting mechanism assigns time-varying weight values ​​based on the real-time resource status of the end nodes, edge nodes, and cloud nodes. To reflect the real-time resource status of different nodes, each resource element e∈E is assigned a dynamic weight w(e,t), where t represents time. The weight can be determined based on factors such as resource load and remaining capacity. Specifically, the dynamic weighting mechanism calculates the time-varying weight value in the following way: ω_ e(t) =w1·(1-load rate) _e(t) )+w2·(remaining capacity) _e(t) Total capacity _e(t) Where w1 and w2 are preset weighting coefficients, and the load rate is... _e(t) Let t be the CPU or GPU utilization parameter of resource element e at the corresponding node, and the remaining capacity. _e(t) Let be the remaining storage space parameter of resource element e at time t in the corresponding node. For example, for a CPU core, if its current load is low and its remaining computing power is high, its weight is relatively large, indicating that the resource is more suitable to be allocated at the current time. Through the dynamic weight mechanism, the model can perceive the changes in the resource status of each node in real time, thereby making resource scheduling more accurate.

[0066] In a preferred embodiment, in step S2, based on the unified resource description model constructed above, a multi-objective optimization model is constructed by introducing a submodular function, with the optimization objectives being the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate. Let f1(S) be the total inference latency of the task execution on the resource set, f2(S) be the total system energy consumption consumed by the task execution, and f3(S) be the resource utilization efficiency. The optimization objective is to find a resource scheduling strategy that simultaneously achieves or is close to the optimal for these objectives. The submodular function f(X) is used to characterize the marginal benefit of resources in the task scheduling process. For each resource e∈E, the marginal gain function is defined as: Δ f (e|S)=f(S∪{e})-f(S). This property reflects that resources have greater marginal value as they become scarcer, making it suitable for expressing the complementarity of heterogeneous resources and the phenomenon of dynamic yield decay. For example, for a task, initially increasing computing resources significantly reduces inference latency, but once the computing resources have increased to a certain level, adding the same amount of computing resources results in a smaller reduction in latency. The submodular function can quantify the impact of different resource scheduling strategies on global inference efficiency. By combining the marginal gain with the optimization objective, the multi-objective optimization model is constructed as follows:

[0067]

[0068] And satisfy Δ f (e|S)=f(S∪{e})-f(S), f(S)∈{f1(S), f2(S), f3(S)};

[0069] Further revision to:

[0070]

[0071] Where α, β, and γ are weighting coefficients, S is the current resource subset, e is the newly added resource, and Δ f (e|S) is the marginal gain function of the resource subset S and satisfies Δ f (e|S) decreases as the size of S increases; f1(S), f2(S), and f3(S) represent the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate, respectively. The latency reduction benefit brought by resource e is α·(f1(S)-f1(S∪{e})), the energy consumption reduction benefit brought by resource e is β·(f2(S)-f2(S∪{e})), and the utilization improvement benefit brought by resource e is γ·(f3(S∪{e}-f3(S)). By solving this multi-objective optimization model, the optimal resource scheduling strategy can be obtained, which minimizes the task inference latency and system energy consumption while maximizing resource utilization rate, thus meeting the task requirements.

[0072] In a preferred embodiment, in step S3, the upper-layer scheduler is responsible for task allocation among the nodes of the edge, cloud, and endpoint. Based on the computationally intensive or storage-intensive characteristics of each task and the resource status of each node reflected by the dynamic weighting mechanism, the task is allocated among the endpoint, edge, and cloud nodes, ensuring that the task is assigned to the most suitable node and balancing the load on each node. In a specific embodiment, computationally intensive tasks are preferentially allocated to edge nodes or cloud nodes where the GPU resource ratio is not less than a preset threshold; and / or, storage-intensive tasks are preferentially allocated to edge nodes or endpoint nodes where the storage capacity is not less than a preset capacity threshold. When allocating tasks, it is necessary to consider the load balancing of nodes to avoid some nodes being overloaded while others are idle.

[0073] In a preferred embodiment, in step S4, after the task is assigned to a node, the lower-level scheduler is responsible for allocating computing and storage resources to the node that received the task, according to the task's requirements. Sufficient memory space and CPU time slices are allocated to the task to ensure efficient operation. In a specific embodiment, when allocating storage resources, the virtual memory ratio does not exceed a predetermined threshold of physical memory capacity. When allocating computing resources, time-slice round-robin scheduling is used, and the CPU time slice ratio of a single task does not exceed a predetermined threshold of the total number of cores on the node. Simultaneously, the lower-level scheduler also needs to consider the dynamic changes in node computing and storage resources, such as storage resource usage and computing load changes, and adjust the resource allocation scheme in a timely manner to adapt to the task's operational needs.

[0074] In another embodiment, reference Figure 2 As shown, Figure 2 The diagram shows a flowchart of the edge-cloud storage and computing resource collaborative optimization method provided in Embodiment 2 of the present invention.

[0075] A method for collaborative optimization of edge-cloud storage and computing resources includes the following steps:

[0076] Step S1: For heterogeneous storage and computing resources of end nodes, edge nodes and cloud nodes, construct a unified resource description model based on dynamic matroid theory.

[0077] Step S2: Based on the unified resource description model, with the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate as optimization objectives, a submodular function is introduced to construct a multi-objective optimization model.

[0078] Step S3: The upper-layer scheduler allocates the task among the end nodes, edge nodes, and cloud nodes based on the computationally intensive or storage-intensive characteristics of the task and the resource status of each node as reflected by the dynamic weighting mechanism, thereby balancing the load of each node.

[0079] Step S4: After the task is assigned to the corresponding node, the lower-level scheduler allocates storage and computing resources to the node to process the task based on the solution results of the multi-objective optimization model.

[0080] Step S5: Update the time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model through a time sliding window, and adjust the task allocation of the upper-level scheduler and the resource allocation of the lower-level scheduler.

[0081] This embodiment, based on the above embodiment, further updates the time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model through a time-sliding window, adjusting the task allocation of the upper-level scheduler and the resource allocation of the lower-level scheduler. It senses changes in task requirements and fluctuations in resource status, and dynamically updates the weights using a time-sliding window. Simultaneously, it updates the unified resource description model and the multi-objective optimization model, thereby ensuring the adaptability of the optimization model. Within each period T, the updated scheduling decision mechanism is as follows: Among them, S t The objective function value for the current state. This represents the objective function value in the next state. By updating the scheduling decision mechanism, scheduling decisions can be optimized in real time based on dynamic changes in task requirements and resource status. For example, when a new task appears in the system, the upper-level scheduler will re-evaluate the task allocation scheme and assign the new task to a suitable node; when the lower-level scheduler detects a change in the resource status of a node, it will promptly adjust the resource allocation of tasks on that node to ensure the normal operation of the tasks. Through continuous online evolution and optimization, the system can adapt to various complex application scenarios and dynamically changing resource requirements, improving the overall performance and adaptability of the system.

[0082] Furthermore, the specific technical details of each process from S1 to S4 are the same as those disclosed in the above embodiments, and will not be repeated in this embodiment.

[0083] Reference Figure 3 , Figure 3 An edge-cloud storage and computing resource collaborative optimization device 300 is shown, which can realize the following through... Figure 1 , Figure 2 The device provided in this embodiment of the invention can realize each process of the above-described edge-cloud storage and computing resource collaborative optimization method.

[0084] An edge-cloud storage and computing resource collaborative optimization device 300 includes:

[0085] The resource description model construction module 310 is used to construct a unified resource description model based on dynamic matroid theory for heterogeneous storage resources and heterogeneous computing resources of end nodes, edge nodes, and cloud nodes. The unified resource description model includes: a set of resource elements, which contains storage resource units and computing resource units of the end nodes, edge nodes, and cloud nodes; an independent set family, which is defined by the resource allocation constraints of the end nodes, edge nodes, and cloud nodes; and a dynamic weighting mechanism, which assigns time-varying weight values ​​according to the real-time resource status of the end nodes, edge nodes, and cloud nodes.

[0086] The multi-objective optimization model construction module 320 is used to construct a multi-objective optimization model based on the unified resource description model, with the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate as optimization objectives, and introduces a submodular function.

[0087] The upper-layer scheduling module 330 is used by the upper-layer scheduler to allocate the task among the end nodes, edge nodes, and cloud nodes according to the computationally intensive or storage-intensive characteristics of the task and the resource status of each node reflected by the dynamic weighting mechanism, so as to balance the load of each node.

[0088] The lower-level scheduling module 340 is used to allocate storage and computing resources to the node to process the task after the task is assigned to the corresponding node, based on the solution results of the multi-objective optimization model.

[0089] The scheduling update module 350 is used to update the time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model through a time sliding window, and adjust the task allocation of the upper-level scheduler and the resource allocation of the lower-level scheduler.

[0090] It should be understood that the descriptions of the above-described edge-cloud storage and computing resource collaborative optimization methods also apply to the edge-cloud storage and computing resource collaborative optimization apparatus according to embodiments of the present invention. To avoid repetition, they will not be described in detail again.

[0091] Furthermore, it should be understood that in the edge-cloud storage and computing resource collaborative optimization device according to the embodiments of the present invention, the above-described division of functional modules is only used as an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the edge-cloud storage and computing resource collaborative optimization device can be divided into functional modules different from the modules shown above to complete all or part of the functions described above.

[0092] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0093] like Figure 4As shown in the figure, the present invention also provides an electronic device 400, including a processor 410, a memory 420, and a program or instructions stored in the memory 420 and executable on the processor 410. When the program or instructions are executed by the processor 410, they implement the steps of the above-mentioned edge-cloud storage and computing resource collaborative optimization method and achieve the same technical effect.

[0094] This invention also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the above-described edge-cloud storage and computing resource collaborative optimization method and achieve the same technical effect.

[0095] Another embodiment of the present invention also provides a computer program product, comprising a computer program that, when executed by a processor, implements the steps of the above-described edge-cloud storage and computing resource collaborative optimization method.

[0096] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0097] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be applied, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0099] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for collaborative optimization of edge-cloud storage and computing resources, characterized in that, include: For heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes, a unified resource description model is constructed based on dynamic matroid theory. This unified resource description model includes: A set of resource elements, which includes the storage and computing resources of the end nodes, edge nodes, and cloud nodes; Independent clusters, defined by the resource allocation constraints of the end nodes, edge nodes, and cloud nodes; A dynamic weighting mechanism, wherein the dynamic weighting mechanism assigns time-varying weight values ​​based on the real-time resource status of the end nodes, edge nodes, and cloud nodes; Based on the unified resource description model, a multi-objective optimization model is constructed by introducing a submodular function with the optimization objectives of the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate. The upper-layer scheduler allocates the task among the end nodes, edge nodes, and cloud nodes based on the task's computationally intensive or storage-intensive characteristics and the resource status of each node as reflected by the dynamic weighting mechanism, thereby balancing the load on each node. After the task is assigned to the corresponding node, the lower-level scheduler allocates storage and computing resources to that node to process the task based on the solution results of the multi-objective optimization model.

2. The method according to claim 1, characterized in that, The construction of the resource element set includes: traversing the hardware resource lists of the end nodes, edge nodes, and cloud nodes, extracting storage resource units that meet a preset storage capacity threshold and computing resource units that meet a predetermined computing capability threshold, and forming the resource element set.

3. The method according to claim 1, characterized in that, The resource allocation constraint for the independent set family is as follows: for any task, when the total storage resources of the resource subset are not less than the storage requirement threshold of the task, and the total computing resources are not less than the computing requirement threshold of the task, the resource subset belongs to the independent set family.

4. The method according to claim 1, characterized in that, The dynamic weighting mechanism calculates time-varying weight values ​​in the following manner: ω_ e(t) =w1·(1-load rate) _e(t) )+w2·(remaining capacity) _e(t) Total capacity _e(t) ) Where w1 and w2 are preset weighting coefficients, and the load rate is... _e(t) Let t be the CPU or GPU utilization parameter of resource element e at the corresponding node, and the remaining capacity. _e(t) Let t be the parameter of the remaining storage space of resource element e in the corresponding node at time t.

5. The method according to claim 1, characterized in that, The marginal benefit of each resource in the task scheduling process is characterized by the submodular function. Combining the marginal gain with the optimization objective, the multi-objective optimization model is constructed as follows: And satisfy Δ f (e|S)=f(S∪{e})-f(S), f(S)∈{f1(S), f2(S), f3(S)}; Where α, β, and γ are weighting coefficients, S is the current resource subset, e is the newly added resource, and Δ f (e|S) is the marginal gain function of the resource subset S and satisfies Δ f (e|S) decreases as the size of S increases; f1(S), f2(S), and f3(S) represent the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate, respectively. The latency reduction benefit brought by resource e is α·(f1(S)-f1(S∪{e})), the energy consumption reduction benefit brought by resource e is β·(f2(S)-f2(S∪{e})), and the utilization rate improvement benefit brought by resource e is γ·(f3(S∪{e}-f3(S)).

6. The method according to claim 1, characterized in that, Computationally intensive tasks are preferentially allocated to edge nodes or cloud nodes where GPU resources account for no less than a preset threshold; and / or, Storage-intensive tasks are preferentially assigned to edge nodes or end nodes with storage capacity not less than a preset capacity threshold.

7. The method according to claim 1, characterized in that, When allocating the storage resources, the virtual memory ratio shall not exceed a predetermined threshold of the physical memory capacity; When allocating computing resources, time-slice round-robin scheduling is used, and the CPU time slice ratio of a single task does not exceed a predetermined threshold of the total number of cores in the node.

8. The method according to claim 5, characterized in that, Also includes: The time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model are updated by updating the time-varying weight values ​​of the dynamic weighting mechanism and the parameters of the multi-objective optimization model through a time sliding window, thereby adjusting the task allocation of the upper-level scheduler and the resource allocation of the lower-level scheduler.

9. The method according to claim 8, characterized in that, Within each period T, the update scheduling decision mechanism adopted is as follows: in, This is the objective function value in the next state.

10. A device for collaborative optimization of edge-cloud storage and computing resources, characterized in that, include: A resource description model construction module is used to construct a unified resource description model based on dynamic matroid theory for heterogeneous storage and computing resources of end nodes, edge nodes, and cloud nodes. The unified resource description model includes: a set of resource elements containing the storage and computing resources of the end nodes, edge nodes, and cloud nodes; independent sets defined by resource allocation constraints of the end nodes, edge nodes, and cloud nodes; and a dynamic weighting mechanism that assigns time-varying weight values ​​based on the real-time resource status of the end nodes, edge nodes, and cloud nodes. The multi-objective optimization model construction module is used to construct a multi-objective optimization model based on the unified resource description model, with the total inference latency of the task execution on the resource set, the total system energy consumption consumed by the task execution, and the resource utilization rate as optimization objectives, and introduces a submodular function. The upper-layer scheduling module is used by the upper-layer scheduler to allocate the task among the end nodes, edge nodes, and cloud nodes according to the computationally intensive or storage-intensive characteristics of the task and the resource status of each node reflected by the dynamic weighting mechanism, so as to balance the load of each node. The lower-level scheduling module is used to allocate storage and computing resources to the corresponding node to process the task after the task is assigned to the corresponding node, based on the solution results of the multi-objective optimization model.

Images