Optimize Cloud-server Resources for Comfortable Web Conferencing

2.1 Technological superiority

The cloud-server resource-optimization technology uses an artificial intelligence (AI) model that predicts various types of performance on the basis of system load and service-usage conditions and calculates the minimum amount of resources that satisfy the requirements (intents) of service providers and users (Fig. 1).

Fig. 1. Technical overview.

It has the following advantages over conventional resource-control technology that focus only on system performance.

By combining multiple AI models that predict various types of performance, this technology can calculate the optimum amount of resources not only for system-performance requirements but also for intents including QoE requirements. The AI models used are trained to use different log data to predict different types of performance. Basically, a regression analysis model is used. However, some QoE indicators are difficult to predict numerically through regression analysis. In this case, by using a classification analysis model, the technology is able to predict the performance level with high precision (Fig. 1). These features make it possible not only to avoid quality failures and improve resource efficiency but also to provide services tailored to users’ usage conditions and objectives.

Various efforts have also been made to train AI models to more reliably meet QoE requirements. The loss function is unique, for example, with modifications to lower intent violation risk (Fig. 2). The loss function is also defined separately not only for the degree of violation but also for the period and number of violations, which are important for performance management in network operation.

Fig. 2. Requirement-violation risk-control function.

By using both the service (application) and system (infrastructure) workloads (iii. in Fig. 1) as input to the AI, it is possible to calculate resources in accordance with the service usage status (including predictions) in addition to the system load status. This makes it possible to proactively control the service in accordance with the demand forecast of the service provider. This advantage is especially effective in cases such as a web-conference service, where it is not easy to scale up the host server (virtual machine instance) of a conference or change the host server while providing the service once the meeting started. By calculating the optimal resource in advance on the basis of the meeting-reservation information, the system can scale out in advance when the meeting starts.

2.2 Application of cloud-server resource-optimization technology

This technology can be used not only in the resource-design phase when a service is introduced but also for dynamic resource control in the service-operation phase. This technology can be used not only for SaaS (Software as a Service) but also for IaaS (Infrastructure as a Service) infrastructure. In the future, we plan to expand the scope of application scenarios to virtual network services.

3.1 Requirements with web-conferencing services and their solutions

Web-conferencing services have the following requirements due to the changes in social conditions described above.

We can satisfy these requirements by applying this technology, as shown in Fig. 3.

Fig. 3. Image of problem solving.

With the conventional resource operation, it is common to prepare sufficient server resources in advance for the expected maximum service usage (number of users, etc.) on the basis of the experience of engineers and verifications. In this case, when there are only a few users, unnecessary resources are allocated, and when more users access than expected, QoE degradation occurs. It is also difficult to increase the number of central processing unit (CPU) cores while running a service such as web conferencing, which can lead to a fatal situation.

By combining this technology with the control functions of cloud services, it is possible to solve this problem by operating services with the minimum amount of resources when the number of users is small and scaling out in advance servers as the number of users increases.

3.2 Constructing a performance-prediction AI model for web conferencing

QoE indicators that are important for web-conferencing services are generally throughput and jitter, which are indicators of the cleanliness and stability of video and audio. Because these values fluctuate constantly, they are difficult to predict with high accuracy (Fig. 4(a)). Therefore, we focus on the fact that these values do not deteriorate when the CPU usage is below a certain level, and we can derive the optimal resource amount to maintain QoE by using an AI model (Fig. 4(b)) that predicts the CPU usage from the workload and resource amount.

Fig. 4. AI-model evaluation results.

3.3 Peripheral function

As mentioned above, to provide this technology to a web-conferencing service, it is necessary to have the function of determining the amount of resources required and the allocation of the conference among various server instances before the conference starts. Therefore, it is necessary to provide a function to supplement the workload information (the number of video and audio streams) from the meeting-reservation information before an actual meeting occurs (Fig. 5(a)). A function is also required to predict the QoE indicator of each meeting combination in accordance with the supplemented workload and determine whether it can be accommodated in the existing instances with predetermined instance type (number of CPU cores) and execute scale-out in advance (Fig. 5(b)). By operating this resource-calculation function on the basis of these usage conditions (predictions) periodically (from a few minutes to a few hours), proactive and automatic resource control can always be achieved during service operation, and the maximum effect can be obtained in both maintaining QoE and improving resource efficiency.

Fig. 5. Application to a web-conferencing service.