Server-free workflow dynamic resource configuration system based on resource decoupling

Abstract

The invention discloses a server-free workflow dynamic resource configuration system based on resource decoupling, and relates to the field of cloud computing. Comprising a workflow sampling scheduling module, a configuration decoupling search module and a dynamic configuration module. The dynamic configuration module shunts the workload in combination with the input characteristic data, the scheduling workflow sampling scheduling module obtains the decoupled optimal resource configuration, and a mapping relation between the input characteristic and the decoupled resource configuration is established by training a random forest model; the workflow sampling scheduling module analyzes a workflow structure through input, generates a weighted directed acyclic graph and identifies a key path, preferentially searches for optimal configuration for the key path, and then iteratively optimizes a sub-path on the premise of not violating the consistency of the key path; the configuration decoupling search module allocates decoupled CPU / memory resources based on a critical path priority policy. According to the method, the cost is minimized while the service level target is met through an automatic searching method for decoupling the memory and the CPU.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing, and particularly to a serverless workflow dynamic resource allocation system based on resource decoupling. Background Art

[0002] With the rapid development of cloud computing technology, serverless computing, as an emerging computing paradigm, has gradually become the mainstream way to build and deploy modern applications. Serverless computing allows developers to focus on business logic and entrust the management of underlying infrastructure to cloud service providers, greatly simplifying the application development and operation and maintenance processes. Serverless workflow, as an important part of serverless computing, provides high modularity and flexibility by decomposing complex tasks into multiple fine-grained function calls and is widely used in scenarios such as data processing, machine learning, and video analysis.

[0003] However, serverless workflows face problems such as low resource utilization, high costs, and unstable performance in practical applications, especially significant challenges in resource allocation and configuration. Current serverless platforms usually adopt a resource bundling allocation strategy, that is, resources such as memory, CPU, and network bandwidth are tightly coupled and allocated together. Although this strategy simplifies resource management, it cannot meet the diverse needs of different types of workflows, resulting in resource waste and performance imbalance. For example, memory-intensive workflows and compute-intensive workflows have significantly different resource requirements, and a single resource configuration strategy is difficult to meet the needs of both at the same time. In addition, functions in serverless workflows usually have high dynamism and input sensitivity, and different input characteristics (such as data scale, video length, etc.) have a greater impact on resource requirements. Traditional static resource configuration methods cannot dynamically adjust resource allocation according to input characteristics, resulting in resource waste or insufficient performance.

[0004] Commercial cloud providers (such as AWS Lambda Power Tuning) provide resource configuration suggestions by testing the function execution performance and cost under different memory configurations. However, these tools are usually limited to a single resource type (such as memory) and are difficult to handle complex and variable workflow loads. The academic community has proposed various resource configuration optimization methods, mainly divided into two categories: one is the offline method based on prior analysis, which executes functions under different resource configurations and generates performance profiles to select the optimal configuration for the function; the other is the online method based on runtime monitoring, which dynamically adjusts the resource configuration by collecting performance data in real time. However, the offline method depends on the definition of "typical workloads" and is difficult to handle functions with highly variable inputs; the online method requires additional resource overhead, increasing the operating cost and facing performance bottlenecks in complex workflows. The proposed methods all rely on existing serverless platforms, so the resource configuration is also based on the underlying logic of bundled allocation and cannot allocate appropriate memory and CPU resources well according to the characteristics of each function in the workflow.

[0005] In addition, when dealing with input-sensitive workflows, existing research usually only relies on a single attribute such as input size and ignores key features such as data complexity and type, resulting in inaccurate resource configuration. At the same time, existing methods are insufficient in dealing with cross-function dependencies and global resource optimization and are difficult to achieve the optimal resource configuration for the overall workflow. The cold start problem is also an important challenge for the performance optimization of serverless workflows. Although existing methods introduce cold start-aware scheduling, their prediction and mitigation mechanisms still rely on heuristic methods and fail to completely eliminate the impact of cold start on performance.

[0006] Therefore, technicians in this field are committed to developing a dynamic resource configuration system for serverless workflows based on resource decoupling. Recognizing the cost and performance benefits brought by decoupling memory and CPU resources, an automatic configuration method based on resource decoupling is proposed, breaking the tightly coupled configuration mode of memory and CPU resources in traditional serverless platforms. By independently adjusting memory and CPU resources, resource waste and operating costs are significantly reduced. The present invention also proposes a dynamic resource configuration method based on a random forest model, which realizes dynamic adjustment of resource allocation according to input characteristics by capturing the complex relationship between input characteristics and the best resource configuration, maximizing resource utilization and cost efficiency. Summary of the Invention

[0007] In view of the above defects of the prior art, the technical problem to be solved by the present invention is resource waste and performance imbalance in serverless computing; by decoupling CPU and memory resource allocation and dynamically adjusting input characteristic perception, cost optimization of serverless workflows is achieved under the premise of meeting SLO.

[0008] To achieve the above object, the present invention provides a serverless workflow dynamic resource configuration system based on resource decoupling, including a workflow sampling scheduling module, a configuration decoupling search module, and a dynamic configuration module;

[0009] The workflow sampling scheduling module performs global resource optimization through critical path analysis and sub-path optimization;

[0010] The configuration decoupling search module performs decoupling search for CPU / memory resources, manages resource adjustment operations through a dynamic priority queue, and minimizes resource costs while ensuring SLO;

[0011] The dynamic configuration module dynamically adjusts the resource allocation strategy according to input characteristics.

[0012] Further, the workflow sampling scheduling module converts the workflow into a weighted DAG (directed acyclic graph) and identifies the critical path.

[0013] Further, the workflow sampling scheduling module adopts a two-stage optimization strategy. First, it ensures the basic SLO through resource over-allocation, and then performs refined resource allocation through dynamic pruning.

[0014] Further, the workflow sampling scheduling module includes the following steps:

[0015] Step 1.1, Weighted DAG construction: Execute the workflow using virtual inputs and record the running time of each function as the weight of the DAG node;

[0016] Step 1.2, Critical path identification: Extract the critical path through the longest path algorithm;

[0017] Step 1.3, Sub-path generation: Traverse the DAG to find parallel sub-paths connected to the critical path nodes;

[0018] Step 1.4, SLO decomposition: Allocate the end-to-end SLO to the critical path and sub-paths according to the time ratio;

[0019] Step 1.5, Iterative optimization: Use the greedy algorithm to reduce the resources of non-critical paths until the SLO threshold is triggered.

[0020] Further, the configuration decoupling search module independently adjusts the memory and CPU resources and configures them according to the resource affinity of the workflow.

[0021] Further, the configuration decoupling search module adopts a two-dimensional progressive search strategy.

[0022] Further, the configuration decoupling search module includes the following steps:

[0023] Step 2.1. Initialization of the operation queue: Create two priority operations of CPU reduction and memory reduction for each function, and the initial priority is calculated based on the resource unit price and performance sensitivity;

[0024] Step 2.2. Elastic adjustment mechanism: Monitor the delay change after each adjustment. If the SLO is violated, trigger a rollback and lower the operation priority;

[0025] Step 2.3. Optimal cost determination: Define the cost-benefit ratio = (original cost - new cost) / delay increase, and only accept adjustments with a ratio greater than the threshold;

[0026] Step 2.4. Parallel optimization: Perform concurrent adjustments on non-conflicting functions and avoid resource competition through a lock mechanism.

[0027] Furthermore, the dynamic configuration module dynamically configures resources based on a random forest model.

[0028] Furthermore, the dynamic configuration module groups workflows with significantly different input features through a workload shunting mechanism; for each shunted group, obtain the decoupled CPU / memory configuration and performance data as a training set, construct a prediction model of input feature-resource configuration, and dynamically adjust resource allocation according to the input features using the prediction model.

[0029] Furthermore, a custom interface is also included.

[0030] Existing serverless platforms typically adopt a bundled configuration mode for memory and CPU resources. This mode cannot be finely adjusted according to the resource requirements of the workflow, resulting in resource waste and performance imbalance. Moreover, existing resource configuration methods usually ignore the resource affinity differences of the workflow, leading to sub-optimal performance and cost-effectiveness during the automatic configuration process. The present invention proposes a serverless workflow resource configuration framework AARC based on resource decoupling. By breaking the tightly coupled configuration mode of memory and CPU resources in traditional serverless platforms and independently adjusting memory and CPU resources, it can perform fine-grained configuration according to the resource affinity of the workflow, significantly reducing resource waste and operating costs. In the resource decoupled configuration mode of the present invention, memory and CPU resources can be independently adjusted, and the resource adjustment order is dynamically arranged through a priority scheduling queue, so as to perform fine-grained configuration according to the actual requirements of the workflow. By analyzing the resource affinity of the workflow, the framework adopts a priority scheduling mechanism based on real-time performance feedback, which can allocate the most suitable resource combination for each function: giving priority to executing the resource configuration operation that has the greatest impact on the SLO compliance rate, while intelligently avoiding resource conflicts, ensuring that while meeting the service level objective (SLO), the resource adjustment benefits are maximized through queued scheduling, and resource waste is minimized. The framework of the present invention realizes a double breakthrough in resource efficiency and service quality through the dynamic orchestration mechanism of resource decoupled configuration and priority scheduling queue. Experiments show that in three typical workflow scenarios of Chatbot, ML Pipeline, and Video Analysis, compared with the traditional bundled resource configuration method, the framework achieves average cost reduction rates of 44.1%, 49.9%, and 34.9% respectively, while maintaining the SLO requirements compliance for all requests. This achievement stems from the intelligent sorting of resource adjustment operations by the priority scheduling queue, which dynamically adjusts the CPU / memory configuration order by monitoring the function execution status. The experimental results verify its core advantage of balancing resource efficiency and service quality in complex scenarios.

[0031] In input-sensitive workflows, different input characteristics (such as video length, data scale, etc.) can lead to significant differences in the resource requirements of workflow functions. Traditional resource allocation methods cannot be dynamically adjusted, further exacerbating the problems of resource waste and performance degradation. For input-sensitive workflows, this framework proposes a dynamic resource allocation method based on the random forest model. By capturing the complex relationship between input characteristics and the optimal resource allocation, this method establishes a resource allocation prediction model and dynamically adjusts resource allocation according to input characteristics to address the significant differences in resource requirements for different inputs, maximizing resource utilization and cost efficiency. The dynamic resource allocation method based on the random forest model in this invention groups workflows with significantly different input characteristics through a workload shunting mechanism. For each shunted group, the workflow sampling scheduling module is called to obtain the decoupled CPU / memory configuration and performance data as the training set, and a prediction model of input characteristics-resource allocation is constructed. This method trains the model with historical data to capture the complex relationship between input characteristics and resource requirements, and specifically establishes a configuration prediction channel for key characteristics such as video resolution and dataset scale. During actual operation, the system automatically matches the feature groups according to the current input characteristics, uses the trained model to predict the optimal resource allocation, and dynamically adjusts the resource allocation after verifying the feasibility of the configuration through the sampling scheduling module. This method can effectively respond to the changes in resource requirements of input-sensitive workflows, achieve a closed-loop optimization from feature recognition → configuration prediction → scheduling verification, and ensure cost minimization while meeting the SLO. To address the challenges of dynamic resource allocation for input-sensitive workflows, this framework innovatively introduces a dynamic configuration mechanism based on the random forest model. By analyzing key input characteristics such as video duration and resolution, and combining historical configuration data for intelligent shunting, the random forest model can accurately establish a feature-configuration mapping relationship. Experimental verification shows that in the video processing scenario, this module reduces the processing cost of long videos (6 minutes) by 34.9% compared with the Bayesian optimization method and by 45.7% compared with the MAFF gradient descent method. Especially in the stress test of the mixed-length video stream, the module automatically identifies the resource allocation differences between short videos (20 seconds) and long videos, achieving an overall cost reduction of 45.6%, proving the effectiveness and robustness of the feature-driven dynamic configuration strategy in complex input scenarios.

[0032] This invention proposes AARC (Automated Affinity-aware Resource Configuration), a dynamic resource allocation framework for workflows based on resource decoupling, aiming to optimize the dynamic resource allocation of input-sensitive workflows through an automated search method for decoupling memory and CPU, and minimize costs while meeting the service level objective (SLO). As Figure 1As shown in the figure, AARC consists of three major modules: workflow sampling scheduling, configuration decoupling search, and dynamic configuration. After the user uploads the workflow definition and SLO requirements, the system combines the input characteristic data (such as video duration, complexity) to split the workload, driving the workflow sampling scheduling module to obtain the decoupled optimal resource configuration based on the split workload. Then, it uses the sampled configuration information to train a random forest model to establish the mapping relationship between input characteristics and decoupled resource configuration, generating the minimum resource benchmark configuration table that meets the SLO. The workflow sampling scheduling module analyzes the workflow structure through input, generates a weighted DAG and identifies the critical path, preferentially searches for the optimal configuration for the critical path, and then iteratively optimizes the sub-path on the premise of not violating the critical path consistency. The configuration decoupling search module allocates decoupled CPU / memory resources based on the critical path priority strategy, preferentially taking out the operations that may reduce costs to the greatest extent. If violated, it adopts an exponential back-off strategy to quickly recover.

[0033] Compared with the prior art, the present invention has the following obvious substantial features and remarkable advantages:

[0034] 1. Technical advantages: The serverless workflow dynamic resource configuration framework (AARC) based on resource decoupling realizes refined resource configuration by decoupling the bundled allocation of memory and CPU resources. For input-sensitive workflows, AARC proposes a dynamic resource configuration method based on a random forest model, which can dynamically adjust the resource allocation strategy according to input characteristics (such as video length, resolution, data scale, etc.), significantly reducing costs while meeting the service level objective (SLO). In addition, AARC optimizes the resource allocation of the critical path and sub-paths through the workflow sampling scheduling module and the priority scheduling algorithm, solving the problem of resource waste caused by static resource configuration in traditional methods.

[0035] 2. Performance indicators: In the tests of three types of workflows, namely Chatbot, ML Pipeline, and VideoAnalysis, the AARC framework reduces the average running cost by 45.6% compared with Bayesian optimization (BO) and the MAFF gradient descent algorithm, and shortens the configuration search time by 85.8%. Its dynamic configuration module predicts input characteristics through a random forest model, making the SLO violation rate of long video processing approach 0, and reducing resource waste by 55.5%.

[0036] 3. Production implementation: The automated resource search and scheduling mechanism of AARC reduces the dependence on manual intervention. Developers only need to submit the workflow and SLO to complete the optimization deployment. In addition, AARC provides flexible custom interfaces, allowing developers to replace the default random forest model (such as using a reinforcement learning or Bayesian optimization model) to meet the personalized needs of different scenarios.

[0037] 4. Industrial Application Prospect: AARC has broad application potential in the fields of cloud computing, edge computing, and industrial intelligence. First of all, AARC can provide resource decoupling and dynamic configuration capabilities for public cloud providers (such as Alibaba Cloud and Tencent Cloud), helping customers reduce the operating costs of serverless workflows and improve the resource utilization rate of cloud platforms. Secondly, in scenarios with significantly different input characteristics such as video processing (such as content review on short video platforms), medical image analysis, and Internet of Things data processing, the dynamic configuration module of AARC can significantly improve resource allocation efficiency and reduce enterprise IT costs. In addition, by reducing resource waste and improving energy efficiency, AARC can help data centers reduce carbon emissions. Generally speaking, AARC solves the problems of resource waste and cost control in the serverless workflow scenario through resource decoupling and dynamic configuration, and is expected to become the core optimization component of the next-generation serverless platform.

[0038] The following will further illustrate the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings to fully understand the purpose, features, and effects of the present invention. Brief Description of the Drawings

[0039] Figure 1 is the AARC framework of a preferred embodiment of the present invention. Detailed Embodiments

[0040] The following introduces multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0041] In the accompanying drawings, components with the same structure are denoted by the same numerical labels, and components with similar structures or functions are denoted by similar numerical labels. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. To make the drawings clearer, the thickness of some parts in the drawings is appropriately exaggerated.

[0042] The present invention proposes AARC (Automated Affinity-aware Resource Configuration) - a dynamic resource configuration framework for workflows based on resource decoupling, aiming to achieve cost optimization of serverless workflows under the premise of meeting SLO through decoupling CPU and memory resource allocation and dynamic adjustment of input characteristic perception. AARC consists of a workflow sampling scheduling module, a configuration decoupling search module, and a dynamic configuration module. The following are specific embodiments:

[0043] 1: Workflow Sampling Scheduling Module

[0044] As the core scheduling engine of AARC, this module achieves global resource optimization through critical path analysis and sub-path optimization. Its core function is to transform the workflow into a weighted DAG and identify the critical path, providing topological dependencies for subsequent resource allocation. The module adopts a two-stage optimization strategy. First, it initializes over-allocation to ensure the basic SLO, and then realizes refined resource allocation through dynamic pruning. The specific process is as follows:

[0045] 1. Weighted DAG construction: Execute the workflow using virtual inputs and record the running time of each function as the weight of the DAG node.

[0046] 2. Critical path identification: Extract the critical path (such as the "decoding → object detection → encoding" link in the video analysis workflow) that determines the overall workflow delay through the longest path algorithm.

[0047] 3. Sub-path generation: Traverse the DAG to find parallel sub-paths (such as non-critical tasks like logging and metadata processing) connected to the critical path nodes.

[0048] 4. SLO decomposition: Allocate the end-to-end SLO to the critical path and sub-paths according to the time ratio. For example, if the total SLO is 10 seconds and the critical path takes 8 seconds, the sub-path gets a 2-second margin.

[0049] 5. Iterative optimization: Use the greedy algorithm to gradually reduce the resources of non-critical paths until the SLO threshold is triggered.

[0050] 2: Configuration decoupling search module based on priority scheduling

[0051] This module innovatively realizes the decoupled search of CPU / memory resources. It manages resource adjustment operations through a dynamic priority queue and minimizes the resource cost while ensuring the SLO. The core algorithm adopts a two-dimensional progressive search strategy:

[0052] 1. Operation queue initialization: Create two priority operations (CPU reduction, memory reduction) for each function, and the initial priority is calculated based on the resource unit price and performance sensitivity.

[0053] 2. Elastic adjustment mechanism: Monitor the latency change after each adjustment. If the SLO is violated, trigger a rollback and lower the priority of this operation. Introduce an exponential backoff strategy, and the adjustment step size starts from 128MB / 0.5vCPU and gradually shrinks to 4MB / 0.1vCPU after failure.

[0054] 3. Cost-optimal determination: Define the cost-benefit ratio = (original cost - new cost) / latency increase, and only accept adjustments with a ratio greater than the threshold.

[0055] 4. Parallel optimization: Implement concurrent adjustments for non-conflicting functions (located on different paths), and use a lock mechanism to avoid resource competition.

[0056] 3: Dynamic Configuration Module for Input-Sensitive Workflows

[0057] This module addresses the resource efficiency issues caused by differences in input characteristics of workflows through input feature analysis and dynamic resource configuration. First, a feature processing pipeline is constructed: multi-dimensional features such as resolution, frame rate, and dataset size are extracted from the input data, and key features are selected using information gain or principal component analysis (e.g., retaining the resolution feature and discarding the encoding format in a video scenario). Subsequently, the workload is split into multiple groups according to feature differences (such as high / low resolution video groups), and each group calls the sampling scheduling module to independently search for the minimum resource configuration that meets the SLO, establishing a feature-configuration mapping library.

[0058] In the dynamic configuration stage, a random forest model is used to learn the correlation between features and resource configuration (supporting replacement with models such as Bayesian optimization customized by developers). When a new workflow arrives, its input features are parsed and matched to the corresponding group, and the CPU / memory quota required for each function is predicted through the trained model. The system dynamically allocates resources based on the prediction results and monitors the achievement of the SLO during real-time operation to achieve the goal of minimizing resource costs.

[0059] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of this application based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. A serverless workflow dynamic resource configuration system based on resource decoupling, characterized in that, It includes a workflow sampling scheduling module, a configuration decoupling search module, and a dynamic configuration module; The workflow sampling scheduling module performs global resource optimization through critical path analysis and sub-path optimization; The configuration decoupling search module performs decoupling search for CPU / memory resources, manages resource adjustment operations through a dynamic priority queue, and minimizes resource costs while ensuring SLO; The dynamic configuration module dynamically adjusts the resource allocation strategy according to input characteristics.

2. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The workflow sampling scheduling module converts the workflow into a weighted DAG and identifies the critical path.

3. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The workflow sampling scheduling module adopts a two-stage optimization strategy. First, it ensures the basic SLO through resource over-allocation, and then performs refined resource allocation through dynamic pruning.

4. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The workflow sampling scheduling module includes the following steps: Step 1.1, Weighted DAG construction: Execute the workflow using virtual inputs and record the running time of each function as the weight of the DAG node; Step 1.2, Critical path identification: Extract the critical path through the longest path algorithm; Step 1.3, Sub-path generation: Traverse the DAG to find parallel sub-paths connected to the critical path nodes; Step 1.4, SLO decomposition: Allocate the end-to-end SLO to the critical path and sub-paths according to the time ratio; Step 1.5, Iterative optimization: Use the greedy algorithm to reduce the resources of non-critical paths until the SLO threshold is triggered.

5. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The configuration decoupling search module independently adjusts the memory and CPU resources and configures them according to the resource affinity of the workflow.

6. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The configuration decoupling search module adopts a two-dimensional progressive search strategy.

7. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein, The configuration decoupling search module includes the following steps: Step 2.1, Operation queue initialization: Create two priority operations for CPU reduction and memory reduction for each function, and the initial priority is calculated based on the resource unit price and performance sensitivity; Step 2.2, Elastic adjustment mechanism: Monitor the latency change after each adjustment. If the SLO is violated, trigger a rollback and reduce the operation priority; Step 2.3, Cost-optimal determination: Define the cost-benefit ratio = (original cost - new cost) / latency increase, and only accept adjustments with a ratio greater than the threshold; Step 2.4, Parallel optimization: Perform concurrent adjustments on non-conflicting functions and avoid resource competition through a lock mechanism.

8. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The dynamic configuration module dynamically configures resources based on a random forest model.

9. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein The dynamic configuration module, through a workload shunting mechanism, groups workflows with significantly different input characteristics for processing; obtains the decoupled CPU / memory configuration and performance data for each shunting group as a training set, constructs a prediction model of input characteristics-resource configuration, and dynamically adjusts the resource allocation according to the input characteristics using the prediction model.

10. The serverless workflow dynamic resource configuration system based on resource decoupling according to claim 1, wherein, It also includes a custom interface.

Images