Considerable number of researches has been conducted for VM consolidation using various methods based on heuristics. In this research, we have worked on two problems, 1) VM selection method, and 2) Overload detection method. Here, we will discuss about various VM selection methods. For Overload detection methods, there are two types of algorithms, a) Threshold based methods, and b) Prediction based method. In threshold based method, researchers apply different statistical or heuristic methods to calculate threshold for a host to be identified as overloaded. On the other hand, there are some predictive methods where researchers use approaches/techniques to predict the future load of a host.
In [1, 2, 4] Beloglazov et al. proposed heuristic based approach to deduce thresholds through different statistical measures. VM Consolidation problem is divided into sub-problems and algorithms for each sub-problem had been designed. Heuristic based algorithms are designed for each sub problems and designed in such a way that they act, adapt and keep their threshold changing based on different scenario in different time so that they can still provide the functionality and consolidation decision in the changed environment. This adaption process allows the system to be dynamic. They designed threshold based (e.g. IQR) and prediction based (e.g. LR) host overload detection mechanisms. We have shown in result section, our proposed algorithms outperformed the algorithms proposed in literature. References [5, 6] describe CloudSim which provides various functionalities of a cloud environment and facilitates in cloud simulation. Reference [1, 2, 4] have also used CloudSim for simulation. The main components of CloudSim are datacenter, Virtual Machine (VM) and cloudlet. Cloudlet can be data from real cloud. The simulator creates datacenter, Virtual Machine and cloudlet based on the defined parameters. When the simulation starts, Virtual Machines are placed in the datacenter for processing. Sub-problems (i–iv) are already developed in CloudSim. To extend it further, one needs to create new class and develop new methods to test it. The VM selection methods and Overload detection methods are compared in this research and the proposed algorithm performs better in all metrics defined in CloudSim. In the previous section we have compared our proposed algorithm with both thresholds based (MAD, IQR and THR) and prediction based approaches (LR and LRR). We have discussed several approaches which are proposed in the literature.
Farahnakian et al. [9] used ant colony system to deduce a near-optimal VM placement solution based on the specified objective function. In [3] VM consolidation with migration control is introduced. Here VMs with steady usage are not migrated and not steady VMs are migrated to ensure better performance. The migrations are triggered and done by heuristic approaches. But this research has not been used the other sub problem rather focused on only the VM placement problem. We have considered all the sub-problems together.
Farahnakian et al. [18] proposed a Reinforcement Learning-based Dynamic Consolidation method (RL-DC) to minimize the number of active host considering the resource requirement. The RL-DC utilizes an agent to learn the optimal policy. The agent uses the past knowledge to take intelligent decision whether to keep the host in active or sleep mode and improves itself as the workload changes. It also dynamically adapts changes. In [12] linear regression has been used to predict CPU utilization by the same author. These researches are developed in CloudSim and follow the distributed architecture. From result and comparison, it is evident that our proposed algorithm performs better in regression based host overload detection method.
Cao et al. [15] proposed a redesigned energy-aware heuristic framework for VM consolidation to achieve a better energy-performance tradeoff. They designed a Service Level Agreement (SLA) violation decision algorithm which is used to decide whether a host is overloaded with SLA violation or not. SLA violation is determined if the allocated CPU of a particular VM is less than the requested CPU of that VM. In other words, if a host is not capable of assigning CPU resource to all the VMs as per demand, then the SLA violation occurs. There is another type of SLA violation which is SLA violation due to migration. If a particular VM is in migration, at the time of migration, the VM is not capable of serving users need hence it is counted as SLA violation. This research is based on CloudSim and algorithms are developed in CloudSim and they have used mean and standard deviation as the prediction method, whereas in our research we used median and standard deviation derived from median. In the result section we depicted that how our method outperformed. Duy et al. [21] proposed a neural network predictor in a Green scheduling algorithm to predict future resource requirements based on historical data. Based on the prediction, decision is taken to keep unused servers in sleep mode and keep the high utilized servers in active mode. There are also similar works that can be found in [22, 23]. These works provide a host utilization prediction and one sub problem is discussed which is overload detection. Srikantaiah et al. [24] have studied the interrelationships between energy consumption, resource utilization, and performance of consolidated workloads. The study shows the energy performance trade-off for consolidation. That research did not use all the sub problems of VM consolidation, rather considered the whole problem as a single one.
Mastroianni et al. [19] presented ecoCloud, a self-organizing and adaptive approach for the consolidation of VMs on CPU and RAM. Assignment and migration decisions are driven by probabilistic processes and based on local information. Focusing on the VM placement problem, they experimented in real datacenter. However, all the sub-problems are not addressed. Madani et al. [17] focused on an architecture configuration to manage virtual machines in a data center to optimize the consumption of energy and meet SLA by grafting a tracing component of multiple consolidation plans which ensure minimum number of servers is switched on. In this research, the problem is seen as scheduling problem and sub problems are not discussed.
Sheng at al. [11] designed a method based on Bayes model to predict the mean load over a long-term time interval. Prevost et al. [10] introduced a framework by combining load demand prediction and stochastic state transition models. They used neural network and autoregressive linear prediction algorithms to forecast loads in cloud data center applications. These works used statistical and neural network to predict the host utilization and only focused on overload detection of a host.
In [7] and [8] we worked with basic VM selection algorithm and introduced migration control in the built in CloudSim VM selection methods. In [26] a preliminary study was carried out using fuzzy logic in VM selection. As the initial findings are encouraging, in this research we incorporate all our previous methodologies, e.g., migration control and fuzzy logic together and study the performance on VM selection. Besides, we have introduced a new overload detection algorithm based on mean, median and standard deviation of utilization of VMs. In this research, we study the performance of our proposed VM selection algorithm coupled with the newly designed overload detection algorithm. A comparative study has been also reported to present the performance against previous VM selection algorithms found in CloudSim.