| dc.description.abstract | In recent years, cloud computing has become increasingly popular and mature. But at the same time, the extension of downtime of cloud service has become more and more common in recent years, and the cost caused by it has also increased year by year. Virtual Machines (VMs) are the foundation of most cloud services. During cloud system recovery management, the VM needs to be restarted. However, the time to restart the VM varies in different situations. If the VM boot time can be predicted more accurately, the VM placement method that takes the least time to start the service can be found, and the recovery time will be shorter, thereby shortening the downtime of the cloud service.
There has been little research on VM boot time because VM boot time is often considered constant. However, previous studies show this is not correct. Lee proposed five models to predict the VM boot time in the environment of four hosts, and the VM background did not run the program that increase the host CPU loading. The results show that the (Random Forest) RF model is the model with the highest accuracy, but the amount of data it requires grows exponentially with the number of hosts, so it is recommended to be used in a small-scale cloud environment.
However, Lee did not verify if the number of hosts increases, the ML-based model can maintain accuracy; after increasing host CPU loading, the ML-based model can still maintain accuracy. Therefore, this study will address the above two issues. The results show that after increasing the number of hosts, the RF model accuracy does not decrease because it can adapt to a more complex environment; however, after increasing the host CPU loading, the RF model accuracy decreases significantly. In addition, the time cost is high due to the collection of data for ML-based models. Therefore, this study suggests that YLL′s rule-based model should be used in a cloud environment with more than 10 hosts. Its advantage is that it only needs to collect a small amount of data, and the time required is very short compared to ML-based models.
| en_US |