Dynamic system allocation and application of cloud computing virtual resources based on system architecture

2.1 Cloud computing technology

Cloud computing is a computing method based on the Internet, which refers to decomposing huge data computing processing programs into countless small programs through the network “cloud,” and then processing and analyzing these small programs through a system composed of multiple servers. The result is returned to the user. Users can access on demand simply and quickly. For users, cloud computing is the most transparent. When the users send task requests on demand, the cloud data center will respond and process accordingly in the shortest time [5].

2.1.2 Classification

According to the various cloud infrastructure deployment methods of cloud computing, cloud computing systems can be divided into the following types.

1. Public cloud: Cloud service providers will build cloud infrastructure and provide cloud services to the public or large enterprises through open networks.

2. Private cloud: Cloud infrastructure can be used by individual organizations or individuals to build or lease cloud services.

3. Hybrid cloud: One or more personal cloud-based hybrid usage standards. The cloud models that make up a hybrid cloud remain independent.

According to the type of service, bottom-up cloud computing can be divided into three types: cloud infrastructure, cloud platform as a service, and cloud software as a service [8].

1. Infrastructure as a Service (IaaS): Users obtain cloud services, including physical servers, virtual servers, memory, Internet broadband, and additional basic computing resources. Customers do not need to manage and maintain cloud infrastructure and can execute all software including operating systems and other applications.

2. Platform as a Service (PaaS): According to the inherent rules of the platform, manufacturers will use the provided development tools or languages to develop their own applications. Or buy the app directly from the vendor [9]. PaaS provides a platform for developers to build applications and services on the global Internet, and it provides an on-demand development environment for developing, testing, and managing software applications.

3. Cloud software as a service (SaaS): Providers will provide users with ready-to-use applications based on cloud infrastructure. Users obtain application programs from various clients through interfaces such as web browsers. Users do not need to manage and maintain cloud infrastructure [10].

The relationship between the three cloud computing service models is shown in Figure 1. Each service model can be executed independently or in conjunction with cloud infrastructure.

Figure 1

Relationship between cloud computing service models.

In a technical point of view, it is not a simple inheritance relationship. SaaS can be based on PaaS or directly deployed on top of IaaS. PaaS can be built on top of IaaS or directly on physical resources.

2.2 Virtualization technology

In general, a virtualized environment may include virtual appliances, virtual networks, virtual communities, and virtual computing environments. Virtualization technology is the foundation of cloud computing environment. Cloud integration management includes user identification, query, and use of various information resources and services in virtual storage [14]. The system can be accessed directly from middleware software and software. The software no longer requires complex setups based on metal handling machines. The virtual network abstracts the basic structure of the network, enabling it to organize and utilize information effectively in the form of a comprehensive network. The development of the network is an important means of cooperation. Through the joint development community on the Internet, various levels of service software can be realized. The software itself is constantly evolving and is designed for ordinary people. Virtual computing environment provides end users with an integrated, fast method, and interface, and corresponding cloud application resources [15].

2.3 Cloud computing virtual resource allocation

Virtual resources are based on computers and networks, and comprehensively use virtual reality technology, simulation technology, geographic information system technology, remote sensing technology, visualization, and other technical means to simulate the temporal and spatial changes in the resource research objects. The introduction of virtual resources in cloud computing can be divided into two aspects: utilization level and facility level. Either way, the focus is on specific applications and infrastructures [16]. Scaling at the task level means that the system distributes the tasks of the application to resources within the executing virtual cluster (programs that require more computing resources, etc.). At the facility level, development includes basic system resources and allocating infrastructure to users, such as host servers in cloud data centers.

VM resources involve multiple aspects: CPU, memory, network, and disk. The relationship between these resources should be considered when planning VMs. In the allocation of cloud computing resources, the allocation of virtual resources follows a specific strategy: (1) Packing. Allocate the fewest servers for resource requests of tasks, maximize the available capacity of servers in the service cluster, reduce the number of shards, and reduce energy consumption. (2) Split. Split task requests so that they are distributed across as many physical servers as possible. This can reduce the impact of host failures and improve the application performance of tasks [17]. (3) Load distribution. Distribute tasks from the least loaded physical servers to improve overall application performance. (4) High availability balance. Start with the most available server, allocate resources to suit the requirements and costs of the server, and adjust the availability level to use the server more efficiently [18]. (5) Affinity balance. For the most guaranteed application performance, you will configure workloads that are closest to swap resources, such as storage devices [19]. (6) Balance of server model. In order to maximize the utilization of higher performance and more expensive resources, task resources will be allocated according to the type of server model [20].

The cloud resource management platform is a platform for centralized management and unified configuration of basic IT resources such as data center computing resources, storage resources, network resources, and application support resources. The main technologies of cloud resource scheduling mainly include the following aspects:

1. Optimization objective: In order to estimate massive scheduling policies, the scheduling center needs to define various objective functions, such as minimum cost, maximum resource utilization, maximum benefit, and maximum satisfaction of user needs [21].

2. Scheduling algorithm: It can achieve the optimization goal according to the objective function, and save its own resources to the greatest extent in a very short time. Scheduling algorithms are usually NP problems, computationally intensive, and not general. In the industry, each application uses a different optimal scheduling algorithm.

3. Scheduling system architecture: At present, the scheduling system architecture of the cloud environment is mainly distributed in multiple levels, which are closely related to the infrastructure of the data center [22].

4. Analysis of business characteristics of statistical center: Learning business service characteristics helps to optimize the scheduling algorithm [23].

2.4 Virtual resource allocation algorithm

If there is a physical server P that maintains resource D, then the d-dimensional resource utilization of the physical server is 1 ≤ d ≤ D , and then the state vector can be expressed as follows: S → ( p ) = ( γ 1 , γ 2 , … , γ D ) .

Average resource efficiency refers to the average rate or mathematical expectation of the d-dimensional resource utilization of a physical server, denoted by the symbol γ ¯ p [24]. If the state vector of the physical server P is S → ( p ) = ( γ 1 , γ 2 , … , γ D ) , the calculation formula is as follows:

The change value of resource utilization rate can reflect the balance of resource utilization rate by objectively evaluating the change range of each resource utilization rate. By definition, the variation value of resource utilization is equal to the mean variance of the resource state vector of the physical server P, represented by the symbol ξ p [25,26]. The calculation is as follows:

In terms of density, the resource utilization of a physical controller is the current usage of servers as a percentage of the total size of the facility. It is measured as follows:

For a given state vector point S → ( γ 1 , γ 2 , … γ D ) , the domain function is F ( S ) defined as follows:

The order of requested resources for virtual server V is A v = ( A 1 , A 2 , … , A D ) . After the virtual server V is deployed to the physical server P, the state vector of P changes. A state vector is defined as the latent state vector shown as S → p v = ( γ 1 ' , γ 2 ' , … , γ D ' ) . The formula for calculating resource usage γ d ' is as follows:

Whether it is appropriate to deploy the VM V in the physical server P can be judged according to the context vector. That is, the adaptability of the physical server can be judged according to the domain of the potential vector in the vector space. If multiple physical servers selected in the same subdomain have the same priority, the best physical server will be selected [27,28]. The value of the potential state vector at a certain moment is called the state of the system at that moment. The value of the potential state vector at time t = 0 is called the initial state or initial state of the system.

The data center has M subtask declarations, and each service is assigned a J number. The time span required by the data center to complete the task is expressed as: time cost

Based on the utilization of the above three sub-indicators, the utilization formula of VM resource nodes is as follows:

In a cloud computing environment, users acquire resources through metering and billing. Therefore, the goal of virtual resource scheduling in cloud computing environment should not only consider the performance of the application, but also the utilization of resources, the performance of the application, and the energy consumption of the cloud data center [29].

3.1 Experiment content and parameter settings

Test content for test and scenario test segments. Here the service test can verify the efficiency and stability of the system, and the scenario test can verify the function of the system and the accuracy of the scheduling algorithm. The performance test mainly includes three parts: runtime test, CPU test execution, and memory test. The scenario test mainly tests two scenarios: load balancing, energy-saving integration, and hybrid scenarios [30,31].

The experimental parameter settings and VM parameter configuration are shown in Tables 1 and 2. The experiments consider three resources, namely, CPU, RAM, and internet connection flexibility. Physical machine CPU has a storage of 30,000 MIPS, memory of 5,120 MB, and network speed of 2,000 MB/s [32].

3.2 Experimental environment

Cluster-available channels are shared by all users with automatic channel selection. It is a multipurpose, high-performance wireless scheduling communication system for resource sharing, cost sharing, and channel equipment and service sharing. The cluster network topology is shown in Figure 2. In a real cloud environment, a cluster usually consists of multiple racks. Switches are used to connect each physical server in the same rack, and routers are used to interconnect servers in different racks. To bring the results closer to practical application, the 11 PCSs were divided into two groups, and the two switches were connected to the routers [33].

Figure 2

Cluster network topology.

3.3 Experimental process

3.4 Experimental results

The system runtime is about 22–27 ms, and the number of cold-spot physical machines usually accounts for about 1/4 to 1/5 of all physical machines. The CPU utilization test results show that most of the data are around 90–95%. Memory utilization remains low during system operation. On average, it takes up about 3.5 MB of memory.

The total runtime is about 22–27 ms, and the number of cold-spot physical machines is about one-third of all physical machines. Most numbers are around 90–95% based on the results of the CPU usage test. When the system is running, the average memory usage is around 3.5 MB, which is kept at a low level.

4.1 Analysis of experimental results

4.1.1 Total operating time

The running time of the system in the energy-saving integration scenario and the load balancing scenario is shown in Figures 3 and 4. As the cluster size becomes larger, the processing time of each indicator increases, but at the current scale, the time increase is not large. Even at the maximum scale shown in the figure, the total runtime of the system is only around 22–27 ms, so the execution time shown by the energy saving integration algorithm is still low. The load balancing algorithm takes longer to execute than other algorithms. The reason is that the balancing algorithm requires an iterative process to find the optimal placement strategy. This process consumes the most time.

Figure 3

Runtime changes in energy saving integration scenarios.

Figure 4

Runtime changes in a load balancing scenario.

4.2 Analysis of virtual resource utilization

First-Fit algorithm is a continuous physical memory allocation algorithm. It concatenates free memory blocks according to their addresses from small to large. The specific implementation uses a doubly linked list. The resource utilization of the First-Fit algorithm and the MS series algorithm is shown in Figure 5. Both First-Fit and MS-SF algorithms are suitable for static service models, while MS-DF and MS-V algorithms are suitable for dynamic models. It can be seen from the figure that when the number of VM requests is small, the resource utilization of the two algorithms is equivalent. When the number of VM requests increases, the resource utilization of the MS-SF algorithm is higher. This is because these two algorithms are actually greedy algorithms. When the VM request size is small, the decision difference is not large, but when the VM request size becomes large, First-Fit cannot guarantee the most robust algorithm solution. The MS-SF algorithm can ensure that the optimal configuration pair is found to optimize the target when the algorithm solution deviates from the optimal solution. Therefore, as the scale of VM requests increases, the resource utilization of MS-SF will be higher. Given the same request, MS-V algorithm can provide higher resource utilization than MS-DF, because more physical machine types lead to more scheduling options, resulting in less scheduling fragmentation and higher resources utilization. Compared with the static model algorithm, the resource utilization rate of the two is lower, because the mobility of the VM is large, and the physical resources are not always occupied by the VM, so the resource utilization rate under the dynamic model will be relatively low.

Figure 5

Comparison of resource utilization between First-Fit algorithm and MS series algorithm.

When scheduling virtual resources, it is necessary to ensure that the sum of various resources of the VM on the physical host does not exceed the total resources provided by the physical machine, so that various resources have different utilization rates. In this combination, disk is a relatively surplus resource because disk utilization has been relatively low. Therefore, memory utilization is selected as a parameter for evaluating resource utilization, and only physical machine occupancy is counted. By comparison, it can be found that in the case of fewer VM requests, the physical machine resource utilization of the First-Fit and MS-SF algorithms are equivalent. The MS-SF algorithm is a “segmentation” or “clustering” algorithm that groups examples into categories that contain similar features. It identifies relationships in a dataset and generates a series of classifications based on those relationships.

4.3 Analysis of virtual resource allocation strategy

As a resource management scheduler, the physical server PM6 is responsible for selecting appropriate physical resources according to user requirements and providing them to the cloud controller PM1. PM1 is responsible for sending a resource allocation message to the selected physical server, calling the physical resource to deploy the virtual server required by the user, and the selected physical server feeds back the current available resource status of resource monitoring to the resource scheduling manager PM6 device through physical resources, so as to do the next resource scheduling. The resource scheduling manager PM6 receives the virtual server request traffic flow in a certain time slot. A service flow consists of multiple virtual server request queues. Requests may come from different users. The type and number of request queues that make up the traffic vary. Then, the resource scheduling manager analyzes the resources of the service queue and obtains the resource requirements of the virtual server. The 60-slot resource scheduling is performed using the FF policy and the AM-R policy, that is, when the FF policy and the AM-R policy are used respectively, 60 VMs respond to requests for service flows. When the AM-R strategy is adopted, the resource fragmentation generated by the physical server is much less than that of the FF strategy, and the number of resource fragments has been relatively stable. This shows that in the process of deploying virtual servers, the usage status of physical servers is relatively stable. The resource sharding curve of the FF strategy increases linearly with time, and the number of resource shards is much larger than that of the AM-R strategy. This shows that the usage status of the physical machine under the scheduling policy has been changing and may be new. When a physical server is turned on, a large number of resource fragments are generated. For the same virtual server request service flow, the AM-R strategy uses significantly less number of physical servers than the FF strategy. At the same time, we can find that the number of physical servers used by both the strategies increases with the number of traffic flows processed.

In the early stage of virtual server deployment, the physical server utilization of the two strategies fluctuated greatly, and then tended to be stable. However, the physical server utilization of the FF strategy is higher than that of the AM-R algorithm, indicating that the AM-R strategy uses fewer physical servers when the physical resources are sufficient. The reason why the physical server utilization of the FF strategy tends to be stable is that the physical resources in the cluster under the current strategy are saturated and can no longer accept new VM requests. From the physical server utilization curve of the AM-R strategy, it can be seen that in the initial stage of virtual server deployment, that is, when the cloud platform transitions from the initialization stage to the working stage, multiple physical servers will be turned on to deploy virtual server machines. After that, most of the VM requests received were deployed on physical servers in existing working state. If the deployment conditions are not met, a new physical server will be started, so the utilization curve of the physical server is relatively stable. With the passage of time, more and more VM request traffic is received, the used physical resources cannot be released in time, and more and more new physical servers are created. Therefore, the curve is still showing an overall upward trend and will eventually align with the FF strategy’s physical server utilization curve.