Depending upon when to apply the fault tolerance techniques, fault tolerance has been classified as follows (7):

Proactive fault tolerance

This is an important technique, which predicts the fault and eliminates recovery from faults and failures by substituting the alleged component. It is able to detect the problem before it arises. This is a perception that prevents compute node failures from running parallel applications by pre-emptively migrating parts of an application.

Reactive fault tolerance

This mechanism enables us to reduce the effort of failures when they take place. The reactive fault-tolerance technique facilitates the system’s more robust or on-demand fault tolerance.

Adaptive fault tolerance

The fault-tolerance of an application depends on its existing position and the range of control inputs that can be applied efficiently. Therefore, in adaptive fault tolerance, entire procedures are performed automatically as per the condition. It can guarantee reliability of critical modules under resource constraints and temporal constraints by allocating redundancy to less critical modules. It can be affordable by minimizing resource requirements.

In paper, we aim to develop a systematic solution to improve the predictability of the fault-tolerant system in cloud computing environment using the machine learning approach which has been proved to be a well-adapted model on different domains.

Using the reactive fault-tolerant technique increases the clock execution time of the system. While rule-based fault detection may underfit the problem, model-based system are generally complex and computationally intensive, which requires a large amount of skilled work to develop a model for a particular system.

A systematic solution is required to improve the observability of the system. A machine learning algorithm, such as a support vector machine may be a well-efficient solution to solve this particular issue.

The paper is organized as follows: Section “2. Related Works” presents the related works where the various existing techniques have been pointed out as well as a comparison chart is provided to provide an analytical view. Section “3. Framework” introduces the proposed proactive fault tolerance framework; Section “4. Method Development” demonstrates how the proposed technique works in detail; and the last section concludes the paper and discusses future work.

In check pointing, after making a change in system, a checkpoint is created. Whenever a task fails, the job is restarted from the recently added checkpoint.

If the job cannot be executed on a particular machine, it is migrated to another machine where it can be continued.

It permits several copies of tasks to run on different resources for their effective implementation and to receive the expected result.

Different instances of an application are allowed to run on virtual machines and the failure of instances is handled repeatedly.

The command does not meet the safety properties and is likely to be vulnerable.

This is less turbulent than normal stream processing and is based on rollback recovery.

This retires the failed task on the identical resource, which was implemented repeatedly.

The task is resubmitted either to a similar or different resource for execution whenever a failed task is detected.

This technique, with critical function, can be performed by a watch dog.

It enables the workflow to continue until it becomes unimaginable to move forward without catering to the failed task.

It allows frequent reboots of the system. It assists the system with a clean start and a fresh start.

This is regularly monitored and analyzed using a feedback-loop control appliance.

A new state is recognized as a transformed state after the employment of error recovery. If this process is applied thoroughly in the absence of effective error provision, the user is error masking.

A faulty element from the system is removed.

A resource is allocated for execution of task.

User defines the treatment for a task on its failure.

Based upon these techniques, a number of models is implemented. Table 1 summarizes different models based on protection against fault and procedure.

It is equipped with the lm-sensors, which are used to evaluate the system, affecting parameters that are used for forecasting as attributes along with the prediction model that has been purposed in this study.

Node-monitoring module monitors the following:

A failure predictor module is run in each node. It uses the model trained using the support vector machine algorithm to predict failure. The model is trained by using the logs captured in the past and with some seed values. The parameters captured by the node-monitoring modules are used as input by the model, which is used for predictions (Figure 2) (10, 11).

The objective of this module is to decrease the influence of failure of the execution. The policies that will be implemented by the proposed architecture are as follows:

The log is maintained by noting the incident of happening of fault.

The controller module is the one that implements the policies listed earlier. In every node, a controller module is installed. This node is responsible for the action to be performed by the node that is about to fail (12). Once the fault in the system is predicted by the model, the controller module takes an action based upon the policies and records the incident in the log table.

Support vector machine are the one of the well-known supervised approaches. This is usually used for classification purposes. The support vector machine divides the data provided by the hyperplane. Using the support vector model involves the following two phases:

The supervised learning support vector machine must first be trained. Different instances are captured from the log table using seed values and are labeled into two classes:

(a) Normal and

(b) Fault

These instances are further used for training the model. Once the model is trained, it is tested using different testing methods, such as the cross-validation test. Optimal Hyper Plane separating Fault and Normal.

The trained and tested model is implemented and used for the purpose of prediction. The model takes the data captured in step (a) and then predicts the class for the data. If the predicted data class is found to be fault, the alarm is set.

Once the occurrence of a fault is predicted, the controller module takes an appropriate action based upon the policies set and maintains the log.