- Mentor: Seong Soo Cho, Openstack Korea Group
- Seok Hwan Lee, Kyungpook National University
- Tae Gon Lee, Kyungpook National University
- ji Pyo Hong, Kyungpook National University
- Hee Rim Hong, Kyungpook National University
In this project, we have adopted a novel approach focusing on the efficient management of the cloud, diverging from the perspective commonly found in existing visualization dashboards such as Grafana. While conventional tools primarily emphasize the visualization of computing resource usage, our goal extends beyond mere information provision to offer insights into optimizing and managing cloud resources. Consequently, the new dashboard introduces a perspective that not only visualizes usage over time but also provides information on usage patterns by comparing current data with historical data. To achieve this, users are presented with computing resource usage data based on 24 hours, 7 days, and 30 days, enabling a comprehensive understanding of usage trends across the entire cloud environment. Such information is expected to provide tangible value in fine-tuning and optimizing cloud resources.
In order to analyze cloud usage patterns, it is essential to collect metrics on the components that make up the cloud. Clouds are primarily based on computing resources, and the most relevant metric for our ultimate goal of a low-carbon private cloud is electricity usage. However, collecting electricity usage requires direct access to physical resources, so in this project, we built an experimental environment that focuses on collecting three metrics that can indirectly estimate electricity usage: CPU, Memory, and Disk Usage.
In performance monitoring, various key metrics such as CPU utilization, average multi- core load, per-core usage, context switches, idle threads, queue length, interrupts, and system calls are considered. It is crucial to note that, despite sharing the term "CPU utilization," the significance of these metrics varies. CPU utilization, measured in a frequency-based manner, represents the ratio of time a logical processor spends executing threads over a specified time period. On the other hand, CPU usage, measured in a time-based manner, reflects the percentage of time the logical processor spends executing threads within a given time frame. When conducting performance monitoring, the fundamental metric to reference is the CPU core usage (Usage), which indicates the percentage of time a logical processor dedicates to executing threads within the measurement window. This is because CPU core utilization (Utilization) represents the theoretical performance achievable by the logical processor, assuming it continuously executes without entering an idle state. In such cases, the calculated result may exceed 100%, leading to potential inaccuracies if relied upon as the sole metric for performance monitoring. Therefore, for accurate performance monitoring, it is essential to consider CPU core usage metrics within the specified measurement interval, providing a more reliable representation of the actual workload executed by logical processors.
The process of collecting and monitoring metrics in the cloud.
Monitoring in the cloud is basically collecting metrics to visualize performance trends and provide them to users and administrators. This service uses Zabbix servers and agents to monitor the resources of instances and the processes running inside them. The monitoring metrics collected from Zabbix can be analyzed to provide users with CPU, memory, and disk usage patterns.
Instance resource monitoring is performed as follows. Zabbix agent are installed on the instance server to measure metrics, and they are collected by Zabbix agent using the active method.
Metrics can provide cloud users with three main pieces of information that can help them optimize resources and reduce carbon emissions (see Table 1).
Monitoring the CPU, memory, and disk usage of each process within an instance enables the identification of resource-intensive processes, facilitating their efficient management. This crucial step contributes to resource optimization and the reduction of unnecessary energy consumption.
CPU usage compared to the previous day(day/week/month). By meticulously examining processes that exhibit a rise in CPU usage when juxtaposed with their previous day's metrics, one can adeptly discern and address performance bottlenecks within those specific processes. This insightful analysis not only facilitates the efficient allocation of resources but also empowers the implementation of targeted performance optimization strategies.
By monitoring the CPU, memory, and disk usage of instances, we can identify instances that either provide excessive performance or remain underutilized. This enables the prevention of unnecessary hardware resource consumption and facilitates more efficient management of cloud resources. Through these methods, cloud users can contribute to reducing their carbon footprint and improving energy efficiency. Furthermore, enhanced resource management aids in cost savings and promotes sustainable cloud operations.
Hypervisor resource monitoring is performed as follows. Libvirt Exporter and Node Exporter are installed on the hypervisor server to measure metrics, and they are collected by Prometheus using the pull method.
The optimization of overall cloud hypervisor count by consolidating instances with low resource utilization into hypervisors with higher overcommit ratios contributes to reducing energy consumption. By collecting usage metrics at the hypervisor level, this approach aims to efficiently allocate resources, thereby minimizing the number of hypervisors required while effectively managing the workload distribution across the cloud infrastructure (see Table 2). 6 With metrics on the resource usage of individual instances and the resource usage of the hypervisor, administrators can gain the following benefits.
- control Overcommitting and Under Committing: For Over committing, you can make the best use of resources by allocating more virtual machines than available memory and paging out or sharing memory with other virtual machines as needed . For Under committing , you can also allocate a smaller number of virtual CPU cores for each virtual machine to minimize competition for resources and ensure reliability for a particular virtual machine . This enables efficient resource allocation in the hypervisor environment by considering resource usage versus availbale reources.
- Leverage preparedness metrics to predict future resource needs: Preparedness metrics can predict future resource needs, which can be used to optimize resource allocation.
- Detect and respond to abnormal situations early: By continuously monitoring preparedness metrics, you can detect unexpected changes in resource usage and detect and respond to overload situations early.
We can see the process of how the system responds to a client's request (see Fig. 2). The measured metrics are collected by the Data Integration System and stored in MongoDB. On the server, the metrics stored in MongoDB are analyzed and processed into meaningful data to be displayed in a dashboard for users to manage resources efficiently.
