A Serverless Computing Resource Elastic Scaling Method and System Based on Probability Distribution

Abstract

This invention discloses a method and system for elastic scaling of serverless computing resources based on probability distribution, belonging to the field of serverless computing elastic scaling technology. This invention deploys multiple user service functions in a Kubernetes cluster. Users call these deployed functions through the corresponding APIs or HTTP requests provided by the cluster. The elastic scaler expands and shrinks function instances according to the volume of user requests to quickly respond to user requests. This invention calculates the number of preheated function instances by constructing a probability distribution and combines the dispersion of data to address the problem of sudden traffic surges in cloud environments for elastic scaling. This not only ensures timely responses to user requests but also reduces the computational cost and time overhead in calculating the number of preheated function instances while maintaining low latency. By predicting the probability distribution using a deep learning model, request response time can be significantly reduced, effectively mitigating the impact of cold start problems on serverless computing.

Description

Technical Field

[0001] This invention relates to a method and system for elastic scaling of serverless computing resources based on probability distribution, and belongs to the field of serverless computing elastic scaling technology. Background Technology

[0002] With the development of container technology and the popularization of microservice frameworks, developers are looking for a cloud computing paradigm that eliminates the need to manage server operations and hardware resources. Traditional server computing requires developers to rent servers on a subscription-based basis and manage the server's hardware resources. Furthermore, in traditional server computing, users pay for virtual machines or containers that are frequently idle. To meet the needs of developers who don't need to manage servers and to save costs for users, researchers have proposed the concept of serverless computing. According to the Cloud Native Computing Foundation (CNCF), serverless computing is a computing paradigm that eliminates the need to manage servers when building and running applications. It describes a fine-grained deployment model where applications consisting of one or more functions can be uploaded to a platform and executed, scaled up and down, and billed based on actual runtime resource consumption. According to the Cloud Native Computing Foundation's definition of serverless computing, the main advantages of serverless computing are: scalability, serverless computing increases and releases computing resources without relying on the operations and maintenance team, and deployed services can quickly and automatically scale up when traffic peaks arrive, and conversely, can automatically release resources to reduce costs when traffic decreases; and a pay-as-you-go billing model, in serverless computing, users only pay for the computing resources they use, that is, they only pay when executing cloud functions.

[0003] While serverless computing offers many unique features and advantages, its differences from traditional server computing present numerous challenges. These challenges can significantly impact system performance and security. In serverless computing, functions must undergo initialization before execution on the serverless platform. Once a function is invoked, the first step is preparing a function instance; whether or not a new function instance is created to respond to the function request is the key difference between cold and warm starts. Generally, cold starts have significantly higher latency than warm starts. Serverless computing aims to provide users with a high-quality, low-cost service; therefore, minimizing the occurrence of cold starts at the lowest possible cost is a major performance challenge for serverless computing.

[0004] While existing cold start optimization techniques (such as caching optimization, instance reuse, and instance preheating) are effective in most cases, they still encounter performance issues when dealing with cold start problems. On one hand, these optimization techniques introduce additional resource consumption or computational latency. On the other hand, some latencies are significantly higher than the average in concurrent applications, known as tail latency, caused by function initialization, storage access, and bursts of function call traffic. Instance preheating techniques typically employ time series forecasting methods, such as moving averages, exponential smoothing, and ARIMA. Furthermore, due to the rapid development of deep learning and reinforcement learning in recent years, many researchers have applied deep learning and reinforcement learning algorithms to instance preheating techniques. However, the burst traffic issues present in cloud environments greatly increase the difficulty of predicting the number of function instances to be preheated. This can lead to request congestion when function instances in the cluster cannot respond to request traffic in a timely manner during bursts of traffic, which is fatal for services with low latency requirements. Summary of the Invention

[0005] To address the cold start latency issue in serverless computing's elastic scaling, pre-warming function instances is used to reduce the frequency of cold starts and prevent them from occurring, effectively reducing request response time and enabling function instances to respond quickly to requests. This invention, considering the characteristics of serverless computing scenarios, proposes a probability distribution-based elastic scaling method for serverless computing resources. This method utilizes historical data to construct the probability distribution of future requests, aiming to minimize request response time. Furthermore, it calculates the optimal number of pre-warmed function instances based on the data's dispersion, better addressing traffic bursts and effectively reducing serverless computing request response latency, thus mitigating the impact of cold start issues on serverless computing.

[0006] The serverless computing resource elastic scaling method based on probability distribution provided by this invention is as follows:

[0007] The participants include user requests, a Kubernetes cluster, and an elastic scaler. Multiple user service-related functions are deployed in the Kubernetes cluster. Users invoke these deployed functions via APIs or HTTP requests provided by the Kubernetes cluster. The elastic scaler scales function instances according to the volume of user requests to respond to those requests. The elastic scaling method includes:

[0008] Step 1: Establish a network model based on the user requests from the participating entities, the Kubernetes cluster, and the elastic scaler;

[0009] Step 2: Construct a probability distribution of future requests using the historical data of user requests, and then calculate the number of warm-up function instances based on the dispersion of the data;

[0010] Step 3: Based on the number of preheating function instances and the number of ready function instances, the elastic scaler performs the corresponding scaling operation.

[0011] Optionally, step two includes:

[0012] Step 21: The user request is routed to the Kubernetes cluster through the gateway. The cluster service responds to the user request and reports the request volume to the monitoring module of the elastic scale. The monitoring module of the elastic scale collects and saves the user request data.

[0013] Step 22: The analysis module of the elastic expander analyzes the historical data of user requests, calculates the corresponding parameter μ for constructing the probability distribution of future requests and the dispersion cv of the historical data, constructs the corresponding probability distribution through parameter μ and calculates the number of preheating function instances by combining the dispersion cv.

[0014] Step 23: The planning module of the elastic expander compares the number of preheating function instances with the number of ready function instances. If the number of preheating function instances is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module, and vice versa.

[0015] Step 24: The execution module of the elastic scaler will create or release function instances according to the received commands;

[0016] Step 25: Repeat steps 22-25 at the beginning of each expansion cycle to perform elastic expansion.

[0017] Optionally, the method for calculating the number of preheating function instances in step 22 includes:

[0018] For the i-th service, let service i∈S, calculate the parameter λ that constructs the Poisson distribution at time t, i.e.:

[0019]

[0020] in, This indicates that historical data of the i-th service is collected at time t; S represents the set of services.

[0021] The probability density function of future requests is constructed using the parameter λ, namely:

[0022]

[0023] The cumulative probability distribution can be obtained from the probability density function, i.e.:

[0024]

[0025] Then, the observed value OV is obtained using the cumulative probability distribution, that is:

[0026] OV = F -1 (p)

[0027] Where p represents the confidence probability:

[0028] p = α + (1-α) × tanh(cv)

[0029] Where α represents when The confidence probability at that time;

[0030] Finally, the optimal number of warm-up function instances is calculated based on the user-defined target value (TV):

[0031]

[0032] NewOrderInstances represents the optimal number of warm-up function instances.

[0033] Optionally, the expression for the dispersion cv of the historical data is:

[0034]

[0035] Here, var() represents the variance, and mean() represents the mean.

[0036] Optionally, step 22 uses a deep learning model to calculate the corresponding parameter μ for constructing the probability distribution of future requests.

[0037] The second objective of this invention is to provide a serverless computing resource elastic scaling system, the system comprising: a Kubernetes cluster and an elastic scaler; the elastic scaler comprising: a monitoring module, an analysis module, a planning module, and an execution module;

[0038] The Kubernetes cluster receives and responds to user requests, and reports the request volume to the monitoring module. The monitoring module collects and saves user request data.

[0039] The analysis module analyzes the historical data of user requests, calculates the corresponding parameter μ for constructing the probability distribution of future requests and the dispersion cv of the historical data, constructs the corresponding probability distribution through the parameter μ and calculates the number of preheating function instances by combining the dispersion cv.

[0040] The planning module compares the number of preheating function instances with the number of ready function instances. If the number is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module. If the number is higher than the number of ready function instances, it sends a release command and the corresponding number to the execution module.

[0041] The execution module creates or releases function instances based on the received commands.

[0042] Optionally, when Revision is inactive, client requests are forwarded to Activator through the gateway, and Activator receives and caches the requests and reports the request volume to the monitoring module of the elastic scale. When Revision is active, client requests are forwarded directly to the User Container of the Pod. The monitoring module of the elastic scale collects metrics from the Queue-Proxy Container. The analysis module uses historical data to calculate the probability distribution of requests and calculates the observation value OV based on the degree of dispersion. The planning module calculates the number of Pods to be warmed up according to the observation value OV. The execution module performs corresponding scaling operations based on the number of Pods to be warmed up.

[0043] Optionally, the system abstracts the serverless computing elastic scaling process into a MAPE control loop, adjusting the number of function instances in real time to respond to dynamic load changes.

[0044] Optionally, the analysis module uses a deep learning model to calculate the corresponding parameter μ for constructing the probability distribution of future requests.

[0045] A third objective of the present invention is to provide a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the serverless computing resource elastic scaling method according to any one of the preceding claims.

[0046] The beneficial effects of this invention are:

[0047] Due to the burst traffic issues present in cloud environments, time series forecasting methods, such as moving average algorithms, exponential smoothing, and ARIMA, struggle to accurately predict the number of warm-up function instances. This can lead to request congestion when function instances in the cluster cannot respond to requests in a timely manner during burst traffic events, which is fatal for services with low latency requirements. Therefore, this invention addresses these issues by studying data characteristics under burst traffic conditions, aiming to minimize request response time and effectively reduce the impact of cold start problems on serverless computing.

[0048] Furthermore, due to the changing nature of time, user requests also change accordingly. Therefore, the elastic scaling mechanism of serverless computing requires an adaptive strategy to provide automated decision-making by monitoring the cloud environment. Thus, reinforcement learning and deep learning algorithms are well-suited for this field. Classical reinforcement learning algorithms require discrete environments and action spaces, but in serverless computing elastic scaling, environmental states such as memory, CPU utilization, and call times are continuous. If deep learning algorithms are used to predict the number of warm-up instances, due to the burstiness of the cloud environment, it is difficult to achieve high accuracy with simple models, while using complex deep learning models leads to high computational costs and time overhead. Therefore, this invention proposes to use probability distribution combined with the discreteness of data to address the timeliness of user request demands. Moreover, the serverless computing resource elastic scaling method based on probability distribution in this invention has low time complexity, thus ensuring that it does not lead to high computational costs and time overhead.

[0049] The serverless computing resource elastic scaling method based on probability distribution according to the present invention calculates the number of warm-up function instances by constructing a probability distribution and addresses the problem of traffic bursts in cloud environments by combining the dispersion of data, and performs elastic scaling. This not only protects user requests from being responded to in a timely manner, but also reduces the computational cost and time overhead required to calculate the number of warm-up function instances while ensuring low latency. Furthermore, by predicting the probability distribution through a deep learning model, the present invention can significantly reduce request response time, avoid function cold starts, and effectively reduce the impact of cold start problems on serverless computing. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a network model structure diagram of the elastic scaling method according to the present invention.

[0052] Figure 2 This is a diagram of a deep learning model for predicting probability distribution parameters in the elastic scaling method according to the present invention.

[0053] Figure 3 This is a flowchart illustrating the elastic stretching method according to the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0055] Example 1:

[0056] This embodiment provides a serverless computing resource elastic scaling method. The participants include user requests, a Kubernetes cluster, and an elastic scaler. Multiple user service-related functions are deployed in the Kubernetes cluster. Users invoke these deployed functions via APIs or HTTP requests provided by the Kubernetes cluster. The elastic scaler scales function instances according to the volume of user requests to respond to those requests. The elastic scaling method includes:

[0057] Step 1: Establish a network model based on the user requests from the participating entities, the Kubernetes cluster, and the elastic scaler;

[0058] Step 2: Construct a probability distribution of future requests using the historical data of user requests, and then calculate the number of warm-up function instances based on the dispersion of the data;

[0059] Step 3: Based on the number of preheating function instances and the number of ready function instances, the elastic scaler will perform the corresponding scaling operations.

[0060] Example 2:

[0061] This embodiment provides a method for elastic scaling of serverless computing resources based on probability distribution. The specific technical solution of this method is as follows:

[0062] The serverless computing resource elastic scaling method in this embodiment involves user requests, a Kubernetes cluster, and an elastic scaler. Multiple user service functions are deployed in the Kubernetes cluster. Users invoke these functions via APIs or HTTP requests provided by the Kubernetes cluster. The elastic scaler scales function instances according to the volume of user requests to quickly respond to user requests. The specific steps include:

[0063] Step 1: Establish a network model based on the user requests from the participating entities, the Kubernetes cluster, and the elastic scaler;

[0064] Step 2: Construct a probability distribution of future requests using historical user request data, and then calculate the number of warm-up function instances based on the data's dispersion.

[0065] Step 3: Based on the number of preheating function instances and the number of ready function instances, the elastic scaler will perform the corresponding scaling operations.

[0066] The serverless computing resource elastic scaling method based on probability distribution according to this embodiment further includes, in step two, effectively combining a deep learning model to predict the probability distribution of future requests.

[0067] Reference Appendix Figure 2 According to the serverless computing resource elastic scaling method based on probability distribution in this embodiment, the method combined with the deep learning model is as follows: the elastic scaler records the historical data of user requests and provides the corresponding parameters for the deep learning model to calculate and construct the probability distribution of future requests. The obtained parameters are used to construct the corresponding probability distribution. The number of warm-up function instances is calculated in combination with the dispersion of the data. The number of warm-up function instances is compared with the number of ready function instances. If the number of warm-up function instances is lower than the number of ready function instances, the corresponding number of function instances are created. Otherwise, the corresponding number of function instances are released.

[0068] In one implementation, for ease of description, x = {x t-1 ,x t-2 ,x t-3 ,…,x t-n} is the historical data of user requests collected, x t Let ∈N represent the number of user requests at time t. Calculate the parameters μ = DL(x) for constructing the probability distribution of future requests and the dispersion of historical data. DL(.) is a deep learning network model. It constructs the corresponding probability distribution through the parameter μ and calculates the number of warm-up function instances by combining the degree of discreteness cv. It compares the number of warm-up function instances with the number of ready function instances. If the number of warm-up function instances is lower than the number of ready function instances, it creates the corresponding number of function instances, and vice versa.

[0069] According to the serverless computing resource elastic scaling method based on probability distribution in this embodiment, the specific steps for constructing the probability distribution of future requests using historical data and calculating the optimal number of warm-up function instances based on the dispersion of the data are as follows:

[0070] S1. User requests are routed through the gateway to the Kubernetes cluster. The cluster service responds to the user request and reports the request volume to the elastic scale's monitoring module. The elastic scale's monitoring module collects and saves the user request data.

[0071] S2. The analysis module of the elastic expander analyzes historical data of user requests, calculates the corresponding parameters μ = DL(x) for constructing the probability distribution of future requests, and the degree of dispersion of the historical data. The corresponding probability distribution is constructed using parameter μ, and the number of preheating function instances is calculated by combining the degree of discreteness cv.

[0072] S3. The planning module of the elastic expander compares the number of warm-up function instances with the number of ready function instances. If the number of warm-up function instances is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module, and vice versa.

[0073] S4. The execution module of the elastic scaler will create or release function instances according to the received commands;

[0074] S5. Repeat steps S2-S5 at the beginning of each time period to perform elastic scaling;

[0075] Reference Appendix Figure 3 The calculation method for the number of preheating function instances adopts the following calculation rules:

[0076] This indicates that historical data for the i-th service is collected at time t; S represents the set of services; P i This represents the set of function instances for the i-th service;

[0077] The number of warm-up function instances is calculated for all services in the cluster as follows:

[0078] For the i-th service, let service i∈S, calculate the parameter λ that constructs the Poisson distribution at time t, i.e.:

[0079]

[0080] The probability density function of future requests is constructed using the parameter λ, namely:

[0081]

[0082] The cumulative probability distribution can be obtained from the probability density function, i.e.:

[0083]

[0084] Then, the observed value (OV) is obtained using the cumulative probability distribution, that is:

[0085] OV = F -1 (p)

[0086] Here, p represents the confidence probability, which is the probability that all future data values ​​fall within a certain range. This indicates that the calculated number of Pods is highly likely to handle the maximum future request load, i.e.:

[0087] p = α + (1-α) × tanh(cv)

[0088] The tanh function, with its range [0, 1) over its domain [0, +∞), exhibits monotonically increasing properties, ensuring a positive correlation between p and cv, and guaranteeing that p's range is [0, 1). The tanh function formula is as follows:

[0089]

[0090] α represents when The confidence probability at that time, when all historical data are constant, the obtained OV value will be equal to that value, that is:

[0091] α=F(λ)

[0092] cv is a coefficient representing the degree of dispersion, that is, the ratio of variance to mean, i.e.:

[0093]

[0094] Here, var() represents the variance.

[0095] The optimal number of warm-up function instances is calculated based on the user-defined target value (TV), i.e.:

[0096]

[0097] This embodiment of the serverless computing resource elastic scaling method based on probability distribution calculates the number of warm-up function instances by constructing a probability distribution and addresses traffic bursts in cloud environments by considering the dispersion of data. Elastic scaling not only ensures timely responses to user requests but also reduces the computational cost and time overhead in calculating the number of warm-up function instances while maintaining low latency. Furthermore, by predicting the probability distribution using a deep learning model, this embodiment can significantly reduce request response time, avoid function cold starts, and effectively mitigate the impact of cold start issues on serverless computing.

[0098] Example 3:

[0099] This embodiment provides a serverless computing resource elastic scaling system, including: a Kubernetes cluster and an elastic scaler; the elastic scaler includes: a monitoring module, an analysis module, a planning module, and an execution module;

[0100] The Kubernetes cluster receives and responds to user requests, and reports the request volume to the monitoring module. The monitoring module collects and saves user request data.

[0101] The analysis module analyzes the historical data of user requests, calculates the corresponding parameter μ for constructing the probability distribution of future requests and the dispersion cv of the historical data, constructs the corresponding probability distribution through the parameter μ, and calculates the number of warm-up function instances by combining the dispersion cv. In this embodiment, a deep learning model is used to calculate the corresponding parameter μ for constructing the probability distribution of future requests.

[0102] The planning module compares the number of warm-up function instances with the number of ready function instances. If the number is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module. If the number is higher than the number of ready function instances, it sends a release command and the corresponding number to the execution module.

[0103] The execution module creates or releases function instances based on the received commands.

[0104] When there are no ready Pods in a Revision (inactive state), client requests are forwarded to the Activator through the gateway. The Activator then receives and caches the requests and reports the request volume to the elastic scaler's monitoring module. Conversely, when a Revision is active, client requests are forwarded directly to the Pod's User Container. The elastic scaler's monitoring module collects metrics from the Queue-Proxy Container, the analysis module calculates the probability distribution of requests using historical data, and then calculates the Open Value (OV) based on the dispersion. The planning module calculates the number of Pods to be warmed up according to the OV value, and the execution module performs corresponding scaling operations based on the number of warmed-up Pods. See [link to documentation]. Figure 1 .

[0105] This embodiment abstracts the serverless computing elastic scaling process into a MAPE (Monitor, Analyze, Plan, and Execute) control loop, which can adjust the number of function instances in real time to respond to dynamic load changes.

[0106] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for elastic scaling of serverless computing resources, comprising user requests, a Kubernetes cluster, and an elastic scaler, wherein multiple user service-related functions are deployed in the Kubernetes cluster, and users invoke these deployed functions via APIs or HTTP requests provided by the Kubernetes cluster. The elastic scaler scales function instances according to the volume of user requests to respond to the user requests, characterized in that... The elastic stretching method includes: Step 1: Establish a network model based on the user requests from the participating entities, the Kubernetes cluster, and the elastic scaler; Step 2: Construct a probability distribution of future requests using the historical data of user requests, and then calculate the number of warm-up function instances based on the dispersion of the data; Step 3: Based on the number of preheating function instances and the number of ready function instances, the elastic scaler performs the corresponding scaling operations; Step two includes: Step 21: The user request is routed to the Kubernetes cluster through the gateway. The cluster service responds to the user request and reports the request volume to the monitoring module of the elastic scale. The monitoring module of the elastic scale collects and saves the user request data. Step 22: The analysis module of the elastic expander analyzes the historical data of user requests and calculates the corresponding parameters for constructing the probability distribution of future requests. and the degree of dispersion of historical data Through parameters Construct the corresponding probability distribution and combine it with the degree of discreteness The number of preheating function instances is calculated as follows: For the i This service makes the service Calculated in t Parameters for constructing the Poisson distribution at any time ,Right now: in, Indicates in t Time collection i Historical data for each service; Represents a collection of services; Using parameters Construct the probability density function for future requests, i.e.: The cumulative probability distribution can be obtained from the probability density function, i.e.: Then, the observed values ​​are obtained using the cumulative probability distribution. OV ,Right now: in, p Represents confidence probability: in, Indicates when The confidence probability at that time; Finally, combined with the user-set target value TV Calculate the optimal number of preheating function instances: in, Indicates the number of instances of the optimal preheating function; Step 23: The planning module of the elastic expander compares the number of preheating function instances with the number of ready function instances. If the number of preheating function instances is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module, and vice versa. Step 24: The execution module of the elastic scaler will create or release function instances according to the received commands; Step 25: Repeat steps 22-25 at the beginning of each expansion cycle to perform elastic expansion.

2. The serverless computing resource elastic scaling method according to claim 1, characterized in that, The dispersion of the historical data The expression is: in, For variance, This is the mean.

3. The serverless computing resource elastic scaling method according to claim 1, characterized in that, Step 22 uses a deep learning model to calculate the corresponding parameters for constructing the probability distribution of future requests. .

4. A serverless computing resource elastic scaling system based on the serverless computing resource elastic scaling method as described in any one of claims 1-3, characterized in that, The system includes a Kubernetes cluster and an elastic scaler; the elastic scaler includes a monitoring module, an analysis module, a planning module, and an execution module. The Kubernetes cluster receives and responds to user requests, and reports the request volume to the monitoring module. The monitoring module collects and saves user request data. The analysis module analyzes historical data of user requests and calculates the corresponding parameters for constructing the probability distribution of future requests. and the degree of dispersion of historical data Through parameters Construct the corresponding probability distribution and combine it with the degree of discreteness Calculate the number of preheating function instances; The planning module compares the number of preheating function instances with the number of ready function instances. If the number is lower than the number of ready function instances, it sends an expansion command and the corresponding number to the execution module. If the number is higher than the number of ready function instances, it sends a release command and the corresponding number to the execution module. The execution module creates or releases function instances based on the received commands.

5. The serverless computing resource elastic scaling system according to claim 4, characterized in that, When Revision is inactive, client requests are forwarded through the gateway to the Activator, which then receives and caches the requests and reports the request volume to the elastic scale's monitoring module. When Revision is active, client requests are forwarded directly to the Pod's User Container. The elastic scale's monitoring module collects metrics from the Queue-ProxyContainer, and the analysis module uses historical data to calculate the probability distribution of requests, then combines this with the degree of dispersion to calculate the observed values. OV The planning module calculates using the observed values. OV The execution module performs corresponding scaling operations based on the number of preheated Pods.

6. The serverless computing resource elastic scaling system according to claim 4, characterized in that, The system abstracts the serverless computing elastic scaling process into a MAPE control loop, adjusting the number of function instances in real time to respond to dynamic load changes.

7. The serverless computing resource elastic scaling system according to claim 4, characterized in that, The analysis module uses a deep learning model to calculate the corresponding parameters for constructing the probability distribution of future requests. .

8. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the serverless computing resource elastic scaling method according to any one of claims 1 to 3.

Images