Random workflow performance optimization method for server-free platform

Abstract

The invention discloses a server-free platform-oriented random workflow performance optimization method, which comprises a task random workflow sorting stage: according to a basic principle of a queuing theory model, performing arrival sorting on tasks of random workflows on the basis of an arrival sequence and a topological structure of the random workflows; in the instance parameter calculation stage, the blocking probability is calculated through an Erland B formula, the cold start probability is deduced, modeling is conducted through the semi-Markov process in the queuing theory, and performance indexes under the steady state condition are calculated through the parameters so as to evaluate the performance of the platform; and a system performance analysis stage: according to different task parameters, comprehensively analyzing the influence of the different task parameters on the performance of the function task, and based on the influence, giving out estimated resource overhead so as to realize cost minimization. According to the method, a processor distribution mode and constraint conditions are studied, and a demand processor combination prediction model is constructed by using a queuing theory model M / N / m / infinity for the problem of inaccurate prediction caused by strong uncertainty of a cloud center.

Description

Technical Field

[0001] This invention relates to a performance optimization method for random workflows on serverless platforms, belonging to the technical fields of serverless computing, reinforcement learning, load prediction, and random workflow optimization. Background Technology

[0002] In cloud center task processing, unbalanced task configuration seriously affects hardware utilization, power consumption and task response time. At the same time, due to the strong uncertainty of cloud centers, it is difficult to accurately predict its requests. In response to the problem of processor allocation mode in cloud center systems, many improvement measures have been proposed. Ogura et al.[1] optimized the performance of multi-core processors by studying appropriate virtualization configuration. Li et al.[2] implemented a memory pooling technology to maintain low access latency and used a machine learning model to predict possible workloads, while building a machine learning model to predict the required resource allocation. Aladiyan et al.[3] studied advanced data structures and optimized graph structures suitable for cloud applications, and deeply analyzed key technologies such as data partitioning, load balancing and parallel processing to comprehensively improve the operating efficiency and reliability of cloud infrastructure. Zhang et al.[4] proposed a solution to solve the resource scheduling and management challenges in cloud computing environment by using machine learning (especially deep reinforcement learning) optimization technology. Xie et al.[5] used reinforcement learning to develop a power limit strategy suitable for cloud computing nodes based on real-time power consumption data and instantaneous application performance indicators. These methods have demonstrated the excellent effect of machine learning-based methods in improving the hardware utilization of cloud centers.

[0003] Predicting the task load in the cloud center is a key technology for studying processor allocation patterns and constraints. In response to this problem, scholars at home and abroad have proposed a variety of prediction methods, which include various machine learning models. The accuracy and efficiency of prediction are constantly improving. Early studies mainly used statistical methods and traditional machine learning models for load prediction. Chauhan et al. [6] proposed to use the ant colony optimization algorithm to optimize resource allocation in the distributed architecture, thereby shortening the optimization time of task scheduling. Chen [7] used a prediction method based on recurrent neural networks to classify tasks in the cloud cluster into different categories by collecting resource usage measurement data or performance data. Jaros et al. [8] predicted and evaluated the job cycle and cost of candidate execution plans through a cluster simulator, and optimized the resource allocation between tasks through a genetic algorithm. Priyadarshini et al. [9] proposed the Dipper Throated optimization algorithm (DTO-CNN) under convolutional neural networks. The method uses CNN to test historical data for predicting task scheduling, and uses the DTO algorithm to enhance the spatial search capability of CNN, thereby optimizing task scheduling and resource allocation in the cloud environment. These methods have achieved load prediction to a certain extent, but they do not make strong use of historical processor scheduling data and have failed to build a combined prediction model.

[0004] [1]Ogura DR, Midorikawa ET, "Characterization of Scientific andTransactional Applications under Multi-core Architectures on Cloud ComputingEnvironment", 2010 13th IEEE International Conference on ComputationalScience and Engineering, 314-320.

[0005] [2]Li H, Berger D S, Hsu L, Ernst D and Zardoshti P, “Pond: CXL-BasedMemory Pooling Systems for Cloud Platforms”, Proceedings of the 28th ACMInternational Conference on Architectural Support for Programming Languagesand Operating Systems, Volume 2, 2023, 574-587.

[0006] [3]ALADIYAN A, “Efficient Data Structures and Algorithms for CloudComputing Platforms”, 2024 4th International Conference on Advance Computingand Innovative Technologies in Engineering (ICACITE). 2024: 1717-1721.

[0007] [4]ZHANG Y, LIU B, GONG Y, Huang, J, Xu J and Wang W, “Application ofMachine Learning Optimization in Cloud Computing Resource Scheduling andManagement”, Proceedings of the 5th International Conference on ComputerInformation and Big Data Applications. New York, NY, USA: Association forComputing Machinery, 2024: 171-175.

[0008] [5]Xie G, Zeng G, Li R, Li K, “Energy-Aware Processor MergingAlgorithms for Deadline Constrained Parallel Applications in HeterogeneousCloud Computing”, IEEE Transactions on Sustainable Computing, 2017, 2(2): 62-75.

[0009] [6]Chauhan N, Bansal A, Matam R, “Optimized Distributed ResourceAllocation for Cloud Computing Applications”, International Journal onEmerging Technologies, 2020, 11(2): 929-934.

[0010] [7]Chen X, “Failure analysis and prediction in compute clouds”,University of British Columbia, 2014.

[0011] [8]Jaros M, Jaros J, “Optimization of Execution Parameters ofMoldable Ultrasound Workflows Under Incomplete Performance Data”, KLUSÁČEKD, JULITA C, RODRIGO G P. Job Scheduling Strategies for Parallel Processing,Cham: Springer Nature Switzerland, 2023: 152-171.

[0012] [9]Priyadarshini P V A, Ajzan Balassem Z, Seetha A, Nandhini P and Shreyas A V, “Dipper Throated Optimization with Convolutional Neural Network Based Task Scheduling and Resource Allocation in the Cloud Computing”, 2024 International Conference on Integrated Intelligence and Communication Systems (ICIICS), 2024: 1-5. Summary of the Invention

[0013] Technical issues: Current research solutions mainly focus on hardware utilization optimization and load prediction and task scheduling. Regarding hardware utilization, existing solutions optimize scenarios such as distributed multi-virtual machine architectures and single-virtual machine multi-core configurations by optimizing virtualization configurations; memory pooling techniques effectively utilize idle memory resources, reducing DRAM usage by 7%; parallel processing optimization solutions utilize scalable Bloom filters, distributed hash tables, and optimized graph structures to improve the data processing efficiency and reliability of cloud infrastructure; reinforcement learning solutions combine deep learning with cloud computing scheduling through experience optimization, improving the performance and scalability of cloud servers, and achieving performance levels comparable to traditional methods such as ant colony optimization and simulated annealing. In terms of load prediction and task scheduling, traditional optimization algorithms have improved resource allocation on cloud services through a bio-inspired antlion optimization algorithm. Compared with traditional distributed methods, this solution can improve host migration efficiency by 11.30%, while also improving CPU utilization while maintaining the same memory consumption. The neural network prediction solution uses recurrent neural networks to analyze various attribute features and performance time series data to predict the termination status of tasks and jobs in cluster tracking records, achieving a true positive rate of 84% and a false positive rate of less than 20%, while saving 6%-10% of resource consumption. The fusion algorithm solution combines convolutional neural networks with the Dipper Throated optimization algorithm, using a convolutional neural network classifier for task scheduling prediction and resource allocation decisions, and combining it with the DTO optimization algorithm to improve the algorithm's spatial search capability, ultimately improving task throughput and reliability, and reducing model response time. While the aforementioned optimization schemes have achieved significant results in specific scenarios, their generality and adaptability are still limited by static resource assumptions and deterministic workflow models. In serverless platform environments, the unique nature of stochastic workflows makes it impossible to accurately predict workflow arrival times and quantities in advance. Furthermore, the dependencies between tasks within a workflow are complex, and the execution order and dependencies between tasks can significantly impact overall performance. Simultaneously, the execution time and resource requirements of tasks are highly uncertain, not only between different tasks but also potentially varying significantly across different execution instances of the same task. Existing optimization schemes are difficult to directly apply and effectively address serverless platforms, and traditional resource allocation strategies and load forecasting methods also have minimal impact on serverless environments.

[0014] To address the prediction inaccuracies caused by the uncertainty of random workflows in serverless computing platform task scenarios, this invention characterizes performance metrics such as cold start probability, steady-state instance count, and average response time, and proposes an M / N / m / ∞ model for tasks. Furthermore, given the limited function instance resources on serverless platforms, a multi-scenario random workflow task ranking algorithm is proposed. Based on this, performance analysis is performed and system resource costs are calculated according to the impact of different parameters on function tasks. Finally, to address the issue of significant storage and computing resource costs caused by uneven processor configuration in cloud centers, a performance optimization method for random workflows on serverless platforms is provided.

[0015] Technical Solution: To achieve the above objectives, the technical solution adopted by this invention is: a performance optimization method for random workflows on serverless platforms, comprising the following stages:

[0016] Task A: Random Workflow Sorting Stage: Based on the fundamental principles of queuing theory, tasks are sorted according to their arrival order and topology in the random workflow. At the same time, the system modeling is described under the condition that multiple random workflows arrive at Poisson speed and the task response time follows a Gaussian distribution.

[0017] In the B-parameter calculation phase: the cold start probability and blocking probability of the task are derived using the Irish B formula, the task termination rate is analyzed, and a semi-Markov process model is used in queuing theory. The number of states in the model represents the number of instances in the hot instance pool between 0 and the maximum concurrent instance count. When an instance is added to the hot instance pool, the instance is cold-started, and the model changes from state to state. To state The model transitions from state 0 to state 0 when an instance is idle for a certain period and terminates. Switch to status Based on this model, performance metrics under steady-state conditions are calculated to evaluate the platform's performance.

[0018] In the C system performance analysis phase, based on different task parameters, the impact on the performance of the function task is comprehensively analyzed, taking into account the storage and computing overhead in the cloud, and based on this, an estimated total task resource overhead is given to minimize costs. This invention studies processor allocation patterns and constraints. Addressing the problem of inaccurate predictions caused by strong uncertainties in cloud centers, a demand processor combination prediction model is constructed using the queuing theory model M / N / m / ∞. To address the low utilization and high overhead problems caused by unbalanced processor configuration in cloud centers, a demand processor combination prediction model is constructed using historical data from cloud center requests and processor scheduling, thereby enabling the prediction of task resource overhead in cloud centers.

[0019] The specific steps of the random workflow preprocessing stage are as follows:

[0020] A1. Sort the tasks according to the arrival order of the random workflow and the task priority relationship;

[0021] A2. Once the random workflow has fully arrived, move the tasks from the random workflow into a queuing queue and sort them according to a specified sorting method (such as FCFS, HRRN, etc.). The random workflow... Arrival time It follows a Poisson distribution, that is: ;

[0022] A3. The sorted task queue is then passed to the algorithm model below, which represents the tasks in the queue. (Random workflow) The first in Task start time ,in, Represented as a random workflow The first in Task completion time Represented as a random workflow The first in The actual arrival and execution delay of each task.

[0023] The specific steps for the instance parameter calculation stage are as follows:

[0024] B1. Calculate the cold start rate, which includes the following steps:

[0025] B1.1. Calculate the amount of heat in the example pool using the Irish formula B. The probability of being blocked when there are only one instance ,in This indicates that there are instances in the hot instance pool. The probability of blocking per instance This is represented as the workload of the job;

[0026] B1.2. The blocking probability calculated in B1.1 is used when a request is blocked by the hot instance pool and the sum of all current function instances does not exceed the maximum concurrency value. In this case, a new instance will be called and a cold start will be performed on the function task, thereby calculating the cold start probability.

[0027] B2. Calculate the arrival rate of requests to each instance, specifically including the following steps;

[0028] B2.1. Based on the blocking probability, calculate the number of instances in the hot instance pool that are greater than [a certain value]. functions When an instance is used, the function After the task request arrives, it is placed in the hot instance pool. service probability ;

[0029] B2.2. Based on the function obtained in B2.1 Instances in the task hot instance pool The probability of a service can be used to calculate the arrival rate of a request to each instance. When there is... When calling a function for an instance. Task arrival instance rate satisfy: ,in, Indicates that there is When there are one instance, the function The actual arrival rate of the task.

[0030] B3. Calculate the termination rate of the instance, which includes the following steps:

[0031] B3.1. Since the arrival patterns of all function requests follow a Poisson distribution, i.e. Therefore, the arrival of every function request follows a Poisson arrival pattern, i.e. Following an exponential distribution, based on the function obtained above... The arrival rate of requests to each instance ,get ;

[0032] B3.2. Since an instance will be terminated upon expiration when the time interval between two arriving requests exceeds the instance's maximum idle time, this function can be calculated using the probability density function obtained in B3.1. Request for example The probability of the last request before termination This allows us to derive the functions processed by each server before it was terminated. The average number of requests is The reciprocal of. Among them satisfy ;

[0033] B3.3. Calculate the expected value of the arrival time interval and compare it with the function processed before each instance is terminated. Multiply the average number of requests, and then add the idle time T required for instance termination to obtain the result for each function. Instance lifespan ;

[0034] B3.4. Based on the above content, the example can be derived. The termination rate is an example The reciprocal of the survival time can be used to calculate the number of instances in the hot instance pool. When there are only a few instances, the rate at which instances are terminated in the instance pool. The sum of the termination rates for each instance, i.e. .in, Indicated as having When there are one instance, the instance The expiration rate.

[0035] B4. Calculate the steady-state probability of multi-task operation under steady-state conditions, specifically including the following steps:

[0036] B4.1. Based on the semi-Markov process, construct the transition probability matrix Q of its transition states, with the following parameters:

[0037] ,in, Indicating hot instance pool arrive Conversion rate per instance;

[0038] B4.2. Using the state transition matrix obtained in B4.1, obtain the steady-state probability matrix of the transition states. ,in satisfy ,and .

[0039] Compared to traditional models, semi-Markov processes can better represent the irregularity and non-instantaneous nature of system state transitions. At the same time, semi-Markov processes allow for arbitrarily distributed dwell times, which is closer to the non-exponential characteristics of state duration in real systems.

[0040] The specific steps of the system performance analysis phase are as follows:

[0041] C1. Calculate performance using steady-state probability, and perform performance analysis based on the impact of different parameters on the functional task. This includes the following steps:

[0042] C1.1. Calculation Function The probability that a task will be rejected by the system. ,in ;

[0043] C1.2. Calculation Functions The probability that a task will be rejected by the hot start pool ,in ,in, Representation function Average task runtime;

[0044] C1.3. Calculation Functions Probability of using cold start ,in ;

[0045] C1.4. Calculate the average response time ,in , Indicates calling a function Average response time for a task request to be warmed up. Indicates calling a function Average response time for cold start of task requests;

[0046] C2. Construct a system resource cost model .

[0047] in, Representation function Task memory usage Indicates calling a function Average response time of task requests Indicates calling serverless platform functions Number of tasks Representation function CPU usage of the task Indicates the unit price of memory overhead. Indicates the unit price of CPU overhead. This indicates the unit price of hard disk storage overhead. Represented as a random workflow The first in The actual arrival and execution delay of each task Represents a random workflow The first in Disk storage usage of task data results.

[0048] The established model divides server resource utilization into idle state and dynamic load state, which more accurately reflects the actual energy consumption characteristics of heterogeneous server clusters. At the same time, the model uses elastic combination allocation of CPU / memory resources, providing more fine-grained overhead optimization options and realizing deep coupling between the overhead model and the elastic scaling characteristics of containers.

[0049] Beneficial Effects: The random workflow performance optimization method for serverless platforms provided by this invention has the following beneficial effects compared with existing technologies:

[0050] (1) This invention studies the historical processing data records of cloud center processors and constructs a dynamic monitoring model of cloud center processor load to balance the configuration of cloud center processors;

[0051] (2) This invention uses the queuing theory model M / N / m / ∞ to describe the system behavior in a multi-task arrival and multi-server processing environment, and models task arrival through a Poisson process to simulate the randomness and independence of user requests in the real world.

[0052] (3) Based on the historical data of cloud center requests and processor scheduling, this invention introduces a hybrid modeling method based on Markov chains to construct a performance prediction model that can accurately characterize the dynamics of task execution. Compared with traditional static models based on queuing theory (such as PMSC), the constructed model is more in line with the elastic characteristics of serverless platforms and provides accurate help for cloud center random workflow load prediction.

[0053] (4) This invention studies the processor allocation mode and constraints, and based on this, constructs a system resource cost overhead model to evaluate the system performance and resource energy consumption cost of the cloud center, thereby assisting in cost optimization under the serverless platform. Attached Figure Description

[0054] Figure 1 This is a flowchart of the algorithm of the present invention;

[0055] Figure 2 This is a business scenario diagram in an embodiment of the method of the present invention;

[0056] Figure 3 This is a schematic diagram of the overall process in an embodiment of the present invention;

[0057] Figure 4 This is a response time diagram of the task in the embodiment of the method of the present invention. Detailed Implementation

[0058] The present invention will be further illustrated below with reference to the accompanying drawings and specific implementation examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0059] Example:

[0060] like Figure 1 The diagram shows the algorithm flowchart for a performance optimization method for random workflows on serverless platforms. Figure 2 This is a business scenario diagram in the embodiment of the method of the present invention. The blocking probability is calculated by the Irish B formula, the cold start probability is derived, and the performance index under steady state is calculated by the semi-Markov process in queuing theory to predict the performance of the platform. Then, based on different task parameters, the impact on the performance of the function task is comprehensively analyzed. Figure 3 This is a schematic diagram of the overall process in an embodiment of the present invention.

[0061] Suppose that at this time, the platform contains, for example, Figure 3 The machine learning model tasks on the left, where f1 to fd represent the tasks generated during model preprocessing, training, testing, and evaluation, respectively. The future arrival status of tasks on the platform is as follows: Figure 3 As shown on the right. The specific operation of this invention in an embodiment includes the following stages:

[0062] A. Task Random Workflow Sorting Stage: Based on the fundamental principles of queuing theory, tasks are sorted according to the arrival order and topology of the random workflow. At the same time, the system modeling is described under the condition that multiple random workflows arrive at Poisson and the task response time follows a Gaussian distribution.

[0063] B. Task Random Workflow Sequencing Phase: The cold start probability and blocking probability of tasks are derived using the Irish B formula. The task termination rate is analyzed, and a semi-Markov process model is used in queuing theory. The number of states in the model represents the number of instances in the hot instance pool between 0 and the maximum concurrent instance count. For example... Figure 3 As shown in the bottom right corner, in this example, the maximum allowed number of function instances is 6. When an instance is added to the hot instance pool, the instance is cold-started, at which point the model changes from state to state. To state The model transitions from state 0 to state 0 when an instance is idle for a certain period and terminates. Switch to status Based on this model, performance metrics under steady-state conditions are calculated to evaluate the platform's performance.

[0064] C. Instance parameter calculation stage: Based on different machine learning task parameters, comprehensively analyze their impact on the performance of the function task, take into account the storage and computing overhead in the cloud, and give an estimated total resource overhead for the task to minimize costs.

[0065] The specific steps of the random workflow preprocessing stage are as follows:

[0066] A1. The tasks are sorted according to their arrival order and priority relationship within the random workflow. Since the tasks in this example involve machine learning preprocessing, training, testing, and evaluation, the task topology in a random workflow within a machine learning scenario is relatively well-defined, and the arrival order and topology of the random workflow are intuitive. Therefore, it is assumed that the tasks in multiple random workflows also follow a Poisson arrival process, as follows: Figure 4 As shown, the response times of the same system tasks follow a Gaussian distribution.

[0067] A2. Once the random workflow has fully arrived, the tasks of the random workflow are moved into the queuing queue. For machine learning tasks with a clear topology, a first-come, first-served scheduling algorithm can be used to move the tasks of the random workflow into the queuing queue and sort the tasks that have entered the queuing queue.

[0068] A3. Pass the sorted task queue to the algorithm model below, and the model will represent the tasks in the queue.

[0069] The specific steps for the instance parameter calculation stage are as follows:

[0070] B1. Calculate the cold start rate. Since this example discusses the case of multiple function tasks, it is necessary to consider the function tasks run by instances in the hot instance pool. This includes the following steps:

[0071] B1.1. Calculate the amount of heat in the example pool using the Irish formula B. The probability of being blocked when there are only one instance ;

[0072] B1.2. If the sum of all current function instances does not exceed 6, call a new instance and perform a cold start for that function task, and calculate the cold start probability;

[0073] B2. Calculate the arrival rate of requests to each instance. This is for calculating the function arrival rate in the hot instance pool. The termination rate of an instance needs to be calculated first using the function. The arrival rate of requests to each instance specifically includes the following steps:

[0074] B2.1. Based on the blocking probability, calculate the number of instances in the hot instance pool that are greater than [a certain value]. functions When an instance is used, the function After the task request arrives, it is placed in the hot instance pool. service probability ;

[0075] B2.2. Functions Instances in the task hot instance pool The probability of service is used to calculate the arrival rate of requests to each instance. When calling a function for an instance. Task arrival instance The rate is ;

[0076] B3. Calculate the termination rate of the instance, which includes the following steps:

[0077] B3.1. The arrival of every function request follows a Poisson arrival pattern, i.e. Following an exponential distribution, we can obtain... ;

[0078] B3.2. When the time interval between two requests exceeds the instance's maximum idle time, the instance is terminated upon expiration, and the calculation function is executed. Request for example The probability of the last request before termination The function processed before each server is terminated. The average number of requests is The reciprocal of;

[0079] B3.3. Calculate the expected value of the arrival time interval and compare it with the function processed before each instance is terminated. Multiply the average number of requests, and then add the idle time T required for instance termination to obtain the result for each function. Instance lifespan ;

[0080] B3.4. When there are hot instance pools When there are only a few instances, the rate at which instances are terminated in the instance pool. The sum of the termination rates for each instance, i.e. .

[0081] B4. Calculate the steady-state probability of multi-task operation under steady-state conditions. The steady-state probability represents the number of hot instances in the hot instance pool at steady state. The probability of an instance specifically includes the following steps:

[0082] B4.1. For functions When there are hot instance pools functions When an instance is executed, the average termination rate of the instance is... Functions are generated in the instance pool. The rate of the instance is Based on the semi-Markov process, construct the transition probability matrix Q of the transition states;

[0083] B4.2. Obtain the steady-state probability matrix of the transition states from the state transition matrix. .

[0084] The specific steps of the system performance analysis phase are as follows:

[0085] C1. Performance analysis using the model includes parameters such as rejection rate after function arrival, cold start probability, and average response time. Based on the impact of different parameters on the function task, performance analysis is performed, specifically including the following steps:

[0086] C1.1. Calculation Function The probability that a task will be rejected by the system. ;

[0087] C1.2. Calculation Functions The probability that a task will be rejected by the hot start pool ;

[0088] C1.3. Calculation Functions Probability of using cold start ;

[0089] C1.4. Calculate the average response time ;

[0090] C2. Construct a system resource cost model. Based on the above performance and resource cost unit price, the system resource cost is calculated as follows: In this example, the unit price of memory overhead is... The CPU overhead per unit is USD 0.0000000167 per second. The unit price of hard disk storage overhead is USD0.0000166667 per GB for computation time per second. It costs USD 0.0000000309 per GB per second.

[0091] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for optimizing the performance of random workflows on serverless platforms, characterized in that: Includes the following stages: A. Task Random Workflow Sorting Stage: Based on the fundamental principles of queuing theory, tasks are sorted according to the arrival order and topology of the random workflow. At the same time, the system modeling is described under the condition that multiple random workflows arrive at Poisson and the task response time follows a Gaussian distribution. B. Instance Parameter Calculation Phase: The cold start probability and blocking probability of the task are derived using the Irish B formula. The task termination rate is analyzed, and a semi-Markov process model is used in queuing theory. The number of states in the model represents the number of instances in the hot instance pool between 0 and the maximum concurrent instance count. When an instance is added to the hot instance pool, it is cold-started. At this point, the model changes from state to state. To state The model transitions from state 0 to state 0 when an instance is idle for a certain period and terminates. Switch to status Based on this model, the performance index under steady-state conditions is calculated to evaluate the platform's performance. C. System Performance Analysis Phase: Based on different task parameters, comprehensively analyze their impact on the performance of the function task, take into account the storage and computing overhead in the cloud, and give an estimated total resource overhead for the task to minimize costs.

2. The performance optimization method for random workflows on serverless platforms according to claim 1, characterized in that: In the task random workflow sorting stage, the specific steps of the random workflow preprocessing stage are as follows: A1. Sort the tasks according to the arrival order of the random workflow and the task priority relationship; A2. Once the random workflow has fully arrived, move the tasks of the random workflow into the queue and sort the tasks in the queue according to the specified sorting method. A3. The sorted task queue is passed to the subsequent algorithm model, and the model represents the tasks in the queue.

3. The performance optimization method for random workflows on serverless platforms according to claim 2, characterized in that: The instance parameter calculation phase includes two parts: random workflow missing value completion and load prediction. The specific steps are as follows: B1. Calculate the cold start rate; B2. Calculate the arrival rate of requests to each instance; B3. Calculate the termination rate of the instance; B4. Calculate the steady-state probability of multi-task operation under steady-state conditions.

4. The performance optimization method for random workflows on serverless platforms according to claim 3, characterized in that: The system performance analysis phase includes two parts: cluster modeling and feature engineering characterization. The specific steps are as follows: C1. Calculate performance using steady-state probability and perform performance analysis based on the impact of different parameters on the functional task; C2. Construct a system resource cost model.

5. The method for optimizing the performance of random workflows for serverless platforms according to claim 2, characterized in that: A3. The sorted task queue is passed to the algorithm model below, which then represents the tasks in the queue, creating a randomized workflow. The first in Task start time ,in, Represented as a random workflow The first in Task completion time Represented as a random workflow The first in The actual arrival and execution delay of each task.

6. The method for optimizing the performance of random workflows for serverless platforms according to claim 2, characterized in that: The specific steps for the instance parameter calculation phase are as follows: B1. Calculate the cold start rate, which includes the following steps: B1.

1. Calculate the amount of heat in the example pool using the Irish formula B. The probability of being blocked when there are only one instance ,in This indicates that there are instances in the hot instance pool. The blocking probability when there are 100 instances, and , The workload, represented as a job, can be expressed using the average number of instances required, where... This means that under the Poisson distribution assumption, exactly The state weights of each instance that are occupied simultaneously As a normalization factor, the final normalized sum of probabilities is limited to 1; B1.

2. The blocking probability calculated in B1.1 is used when a request is blocked by the hot instance pool and the sum of all current function instances does not exceed the maximum concurrency value. In this case, a new instance will be called and a cold start will be performed on the function task, thereby calculating the cold start probability. B2. Calculate the arrival rate of requests to each instance, specifically including the following steps; B2.

1. Based on the blocking probability, calculate the number of instances in the hot instance pool that are greater than [a certain value]. functions When an instance is used, the function After the task request arrives, it is placed in the hot instance pool. service probability ; B2.

2. Based on the function obtained in B2.1 Instances in the task hot instance pool The probability of service is used to calculate the arrival rate of requests to each instance. When calling a function for an instance. Task arrival instance rate satisfy: ,in, Indicates that there is When there are one instance, the function The actual arrival rate of the task. B3. Calculate the termination rate of the instance, which includes the following steps: B3.

1. Since the arrival patterns of all function requests follow a Poisson distribution, i.e. Therefore, the arrival of every function request follows a Poisson arrival pattern, i.e. Following an exponential distribution, based on the function obtained above... The arrival rate of requests to each instance ,get ; B3.

2. Since an instance will be terminated upon expiration when the time interval between two arriving requests exceeds the instance's maximum idle time, the probability density function obtained in B3.1 is used to calculate this function. Request for example The probability of the last request before termination This allows us to derive the functions processed by each server before it was terminated. The average number of requests is The reciprocal of, among which satisfy ; B3.

3. Calculate the expected value of the arrival time interval and compare it with the function processed before each instance is terminated. Multiply the average number of requests, and then add the idle time T required for instance termination to obtain the result for each function. Instance lifespan ; B3.

4. Derive the example based on the above content. The termination rate is an example The reciprocal of the survival time, and the number of hot instances in the pool. When there are only a few instances, the rate at which instances are terminated in the instance pool. The sum of the termination rates for each instance, i.e. ,in, Indicated as having When there are one instance, the instance The expiration rate, B4. Calculate the steady-state probability of multi-task operation under steady-state conditions, specifically including the following steps: B4.

1. Based on the semi-Markov process, construct the transition probability matrix Q of its transition states, with the following parameters: ， in, Indicating hot instance pool arrive Conversion rate per instance; B4.

2. Using the state transition matrix obtained in B4.1, obtain the steady-state probability matrix of the transition states. ,in satisfy ,and .

7. The performance optimization method for random workflows on serverless platforms according to claim 1, characterized in that: The specific steps of the system performance analysis phase are as follows: C1. Calculate performance using steady-state probability, and perform performance analysis based on the impact of different parameters on the functional task. This includes the following steps: C1.

1. Calculation Function The probability that a task will be rejected by the system. ,in ; C1.

2. Calculation Functions The probability that a task will be rejected by the hot start pool ,in ,in, Representation function Average task runtime; C1.

3. Calculation Functions Probability of using cold start ,in ; C1.

4. Calculate the average response time ,in , Indicates calling a function Average response time for a task request to be warmed up. Indicates calling a function Average response time for cold start of task requests; C2. Construct a system resource cost model ; in, Representation function Task memory usage Indicates calling a function Average response time of task requests Indicates calling serverless platform functions Number of tasks Representation function CPU usage of the task Indicates the unit price of memory overhead. Indicates the unit price of CPU overhead. This indicates the unit price of hard disk storage overhead. Represented as a random workflow The first in The actual arrival and execution delay of each task Represents a random workflow The first in Disk storage usage of task data results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the random workflow performance optimization method for serverless platforms as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the random workflow performance optimization method for serverless platforms as described in any one of claims 1-7.

Images