1 Introduction

Users now have immediate access to pooled resources thanks to a new technology called cloud computing. The core of a cloud computing (CC) data center has changed as a result of the rapid growth of cloud computing technology. Cloud computing and supercomputing are made feasible by increasingly dispersed computing resources, such as high-bandwidth networks, massive data centers, and volumes of storage that are immense [1]. The daily rise in the number of cloud-based services that meet customer demands is made possible by the performance of the cloud environment. Some of the stages that the CC goes through include grid computing, utility computing, parallel computing, and distributed computing [2].

Among the many issues in cloud computing, the fundamental process at the Software as a service (SaaS) level is task scheduling [3]. The service provider must use a type of scheduling that maximizes resource usage while satisfying the demands of the clients. When tasks are completed simultaneously, the service provider distributes multiple tasks to a single or several VMs. While the cloud environment experiences an overall variance in demand, the static load balancing strategy is crucial [4]. Static algorithms cannot function in the cloud environment because workload changes over time. Hence, the workload balancing among the VMs needs dynamic methods. The scheduling process and NP-hard problem consider task scheduling based on the dynamic and heterogeneity properties of CC [31,32,33].

A new branch of networked computing called cloud computing disseminates data technical services across the Internet. More capable computing offers centralized storage, memory, computation, and bandwidth [5]. The backbone of CC is virtualization which enables virtual machines (VMs). In the cloud computing environment, the VMs execute the tasks presented by the clients. Because of the nature of the dynamic cloud, there is variation in physical machine (PM) load [6, 30]. The workload across many computing resources including disk drives, CPUs, network links, and computers is distributed via load balancing.

Dynamic and static are the two basic kinds of load balancing algorithms [28, 29]. At the runtime, the distribution of workload is changed by making the dynamic load balancing model. The dynamic load balancing methodology is used due to the unpredictability of workload[7]. The increasing demand and users urge the necessity to minimize the task completion time and enhance the load balancing capacity. The NP‐complete problem considers the heterogeneous environment-based task scheduling [19,20,21,22,23, 34, 35]. A faster response time enhances system performance overall, and a faster completion time is crucial. According to the rapid development of data centers, energy efficiency is a prominent contradiction in the cloud. Nevertheless, it may lead to a few problems such as performance reduction and service response delay [24, 25]. It is critical to achieving a better balance between energy and performance. The major steps involved in this study are delineated as follows:

  • A hybrid firebug and Tunicate Optimization (HFTO) algorithm is proposed for task scheduling and load balancing in the cloud.

  • The HFTO algorithm is a combination of the Firebug Swarm Optimization (FSO) algorithm and Tunicate Swarm Optimization (TWO) algorithm in which the HFTO algorithm method has several advantages, including enhanced search capability and faster convergence.

  • The HFTO algorithm minimizes the time taken for VM creation and optimizes different Quality of Service (QoS) parameters such as fault tolerance, response time, efficiency, and makespan.

  • The proposed HFTO algorithm improves the time complexity and computational complexity.

The remainder of this article is organized as follows: Sect. 2 provides a review of the current works. The Hybrid Firebug and Tunicate Optimization (HFTO) algorithm are formulated in Sects. 3, and 4 presents the challenges that are the focus of this study. Section 5 gives an overview of the proposed work, while Sect. 6 discusses the experimental findings. Section 7 concludes the article.

2 Related Works

Ebadifard et al. [8] developed an autonomous task scheduling approach for load balancing in cloud computing. An increasing amount of inter-VM communication overheads is the major challenge in dynamic load balancing. This method's performance is evaluated by applying the CloudSim tool and then compared with Naive Bayesian Load Balancing, Honey-Bee, Autonomous, and Round Robin techniques. The experimental results demonstrated maximum load balancing and minimal communication overheads, as well as shorter response times and higher resource usage. However, geographical clouds with dispersed Datacenters are not taken into account.

Hybrid Tabu–Harmony task scheduling (HTHTS) algorithm was proposed by Alazzam et al. [9]. The Harmony and Tabu search algorithms were merged to improve the quality of the findings. While compared to the round-robin, Harmony search, and Tabu search methods, this task scheduling method yielded better results in the case of the total cost, makespan, and throughput. While throughput and cost are increased, task scheduling performance is hindered by crowded VMs.

The Service level agreement-based Load Balancing (SLA-LB) algorithm was suggested by Lavanya et al. [10]. The overall makespan of the task is reduced to exaggerate the performance of task scheduling in clouds. The expected time to complete (ETC) generates the threshold data. The expected robust threshold value optimizes load balancing across machines with the minimum configuration machine allocation to the task, as well as system utilization. Compared to the existing algorithms, the SLA-LB algorithm provided better results using the measures such as VM utilization, gain cost, penalty, and makespan. The experimental results showed that improved task allocation was associated with poor resource utilization and complicated costs. The new energy-efficient load balancing global optimization model was introduced by Lu et al. [11]. The load balancing and energy consumption optimization is their major objective function. Based on the experimental investigation, the minimum energy consumption with better power management is obtained but this model is not suitable for large data centers.

Khorsand et al. [12] introduced an energy‐efficient task‐scheduling (EETS) algorithm in cloud computing. Based on the best–worst Method (BWM), an energy-efficient task-scheduling model is introduced to determine cloud scheduling solutions. For each criterion, the importance weights are assigned by applying the BWM process. The effectiveness of this strategy is evaluated using a variety of statistical testing standards, including ANOVA and the CloudSim toolkit. Nevertheless, it does not consider reliability and computational complexities. An Improved Firework Algorithm was proposed by Wang et al.[13] for task scheduling in cloud computing. The primary choice is lightweight computing because of its weak processing ability and fewer computing resources. The experimental results demonstrated better load balancing with minimum task processing time and it does not share the resources among data centers. Table 1 lists the summary of related works.

Table 1 List of related works
Full size table

3 Formulation of Hybrid Firebug and Tunicate Optimization (HFTO) Algorithm

The Firebug Swarm Optimization (FSO) and Tunicate Swarm Optimization (TWO) algorithms are combined to form the Hybrid Firebug and Tunicate Optimization (HFTO) algorithm. As a result, the following is a breakdown of the HFTO algorithm's steps:

3.1 Firebug Swarm Optimization (FSO) Algorithm

The two basic behaviors of the Firebug Swarm Optimization (FSO) algorithm are gregarious behaviors and solitary individuals forming aggregations. These aggregations help select compatible partners for reproduction and lessen predator activity [14]. The Firebug behaviors of FSO algorithm models are explained below:

  • Male firebugs use pheromones, which are chemical messages, to defend and attract the female bug colonies.

  • The fittest female in its colony only mates with every male bug.

  • The male bug based on the colony strongly attracts the female bugs.

  • Fit females' chemical signals attract male bugs who would never join their colony.

  • Even if a single aggregation moves together, the swarm members never disperse.

The following section explains the mathematical model of the FSO algorithm.

3.1.1 Female Colonies Formation

The fit bugs are connected with the lower cost values because the firebugs search for fit mates. In the search space, the FSO model initializes with the female bugs \(M_{N} M_{G}\) in which every male bug contains the female bugs colony \(M_{G}\). The uniform random vector variable treats each female bug's initial location in accordance with the search space.

3.1.2 Selection of Males

The placement of each male insect is first determined by the best female bug in the colony, but the male bug mates with the healthiest female bug. Every male bug position is updated by the best female in the colony. The location updates are implemented using operations on elements like Hadamard matrix multiplication.

3.1.3 Female Bug Chemotactic Movement

According to the male selection behavior, the female bug’s location is updated after initialization. The FSO strategy relies heavily on avoiding scalar operations in order to achieve a successful result. Store all of the female bug positions in one matrix corresponding to a specific male colony. Let \(n(n) \cdot G\) be the corresponding female bug position stored in a matrix \(M_{G}\). The effective Hadamard multiplication model with a matrix update equation updates the female bugs.

$$N_{y} \leftarrow repmat(n(n) \cdot y,1,M_{G} )$$
(1)
$$N_{z} \leftarrow repmat(n(b) \cdot y,1,M_{G} )$$
(2)

The random integer b falls in interval 1 and \(M_{G}\). With the function repmat (B, m, n), the n copies of B return a matrix in addition to the row dimension.

$$n(n) \cdot G \leftarrow n(n) \cdot G + C_{1} \Theta \left( {N_{y} - n(n) \cdot G} \right) + C_{2} \Theta \left( {N_{z} - n(n\_\cdot G} \right)$$
(3)

3.1.4 Male bug’s Attraction Toward Fittest Female Bugs

The fit females attract every male bug. The male bug moves to a comparable fittest bug based on the specific geographical position, and the entire population is limited to the specified geographical position, preventing the swarm from dispersing. Equation (4) explains the update rule for the male bug movement behavior to the fittest female bug.

$$n(n) \cdot G \leftarrow n(n) \cdot y + C_{3} \Theta (h - n(n)y)$$
(4)

3.1.5 Swarm Cohesion

Individual bugs are not scattered in a swarm when the entire swarm moves as a single entity [15]. The herd cohesion model is represented in Eq. (5).

$$n(n) \cdot y \leftarrow n(n) \cdot y + C_{4} \Theta (h - n(a)y)$$
(5)

Individual bugs must choose the best match for reproduction based on the reproduction swarm behavior of firebugs. Based on the triangular vector law, the \(y^{1}\) and \(y^{2}\) express the vector y is expressed as:

$$y = y^{1} + \beta (y^{2} - y^{1} )$$
(6)

The relocation of the male insect to the fittest female bug is accomplished by Eq. (7).

$$n(n) \cdot y \leftarrow n(n) \cdot y + \beta (h - n(n) \cdot y$$
(7)

The element-wise Hadamard multiplication that takes the place of the scalar vector is shown in Eq. (8).

$$n(n) \cdot y \leftarrow n(n) \cdot y + C_{4} \Theta (h - n(a) \cdot y$$
(8)

The below expression delineates the weak movement of female bugs to random males and the strong movement to the dominant.

$$n(n) \cdot G \leftarrow n(n) \cdot G + C_{1} \Theta (N_{y} - n(n) \cdot G + C_{2} \Theta \left( {N_{z} - n(n) \cdot G} \right)$$
(9)

The movement to the random male and the dominant male is denoted as \(C_{2} \Theta \left( {N_{z} - n(n) \cdot G} \right)\) and \(C_{1} \Theta (N_{y} - n(n) \cdot G\). Additionally, the degree of attraction to dominant and random male bugs is \(C_{1}\) and \(C_{2}\).

3.2 Tunicate Swarm Optimization (TSO) Algorithm

The conflict among the search agents avoidance is performed via a vector to calculate the position of the new search agent [15].

$$\overrightarrow {P} = \frac{{\overrightarrow {H} }}{{\overrightarrow {M} }}$$
(10)
$$\overrightarrow {H} = a_{2} + a_{3} - \overrightarrow {F}$$
(11)
$$\overrightarrow {F} = 2.a_{1}$$
(12)

The gravity force and the water flow advection are \(\overrightarrow {H}\) and \(\overrightarrow {F}\). The random interval 0 and 1 for the random variables such as \(a_{1}\),\(a_{2}\) and \(a_{3}\). Where \(\overrightarrow {M}\) express the social force between the search agents.

$$\overrightarrow {M} = [S_{\min } + a_{1} .S_{\max } - S_{\min } ]$$
(13)

From Eq. (13), \(S_{\max }\) and \(S_{\min }\) are the subordinate and initial speeds. The values of \(S_{\min }\) and \(S_{\max }\) are 1 and 4.

3.2.1 Follows the Optimal Agent Position

For reaching the optimal solution, following the current best agent is important. There is no fighting among the agents in the swarm. Equation (14) describes the best agent with its optimal position [16].

$$\overrightarrow {RD} = \left| {Y_{best} - r_{andom} \cdot \overrightarrow {{S_{p} (y}} )} \right|$$
(14)

From the above equation, the food source and search agent distance is \(\overrightarrow {RD}\). Where, \(Y_{best}\) and y is the food source position and the current iteration of tunicate. The 0 to 1 interval range for the random integer \((r_{andom} )\) and \(\overrightarrow {{S_{p} (y}} )\) is the position of the tunicate.

3.2.2 Keep Near to the Best Agent

Equation (15) computes the best position to ensure the search agent is still near to the best agent. The best food source position \(Y_{best}\) is and the tunicate position is \(\overrightarrow {{S_{p} (y}} )\).

$$\overrightarrow {{S_{p} (y)}} = \left\{ {\begin{array}{*{20}l} {Y_{best} + \overrightarrow {P} \, \cdot \,\overrightarrow {RD} } \hfill & {\,if\,\,\,\,\,\,r_{andom\,} \ge 0.5} \hfill \\ {Y_{best} - \overrightarrow {P} \, \cdot \,\overrightarrow {RD} } \hfill & {if\,\,\,\,\,\,r_{andom\,} < 0.5} \hfill \\ \end{array} } \right.$$
(15)

3.2.3 Swarm Behavior

Based on the positions of two agents, update the current agent position to model the swarm characteristics of tunicates.

$$\overrightarrow {{S_{p} (y + 1)}} = \frac{{\overrightarrow {{S_{p} (y)\,}} + \,\overrightarrow {{S_{p} (y + 1)}} }}{{2 + a_{1} }}$$
(16)
figure d

3.3 HFTO Algorithm

The Firebug Swarm Optimization (FSO) algorithm has several advantages such as reducing execution time, modern CPUs, and GPUs of fast Single Instruction Multiple Data (SIMD) feature in which it performs multiple arithmetic operations, lower cost, exploration, and Behaviour and etc. However, the FSO algorithm's exploitation swarm cohesiveness behavior required few changes because it degrades FSO's performance in terms of convergence speed, searchability, and computational complexity. As a result, we integrated the swarm behavior of tunicate swarm optimization to the FSO exploitation swarm cohesion, so improving the FSO algorithm's performance, and the newly developed model is known as the Hybrid Firebug and Tunicate Optimization (HFTO) method. As a result, the HFTO approach offers various advantages, such as improved search capacity and faster convergence. Algorithm 1 delineates the pseudocode of the HFTO algorithm.

4 Problem Formulation

The set of N number of virtual machines is described as \(VM = \left\{ {VM_{1} ,VM_{2} ,VM_{3} ,...,VM_{N} } \right\}\). Where \(TA = \left\{ {TA_{1} ,TA_{2} ,TA_{3} ,...,TA_{K} } \right\}\) describes the set of K tasks. We need to know how much the load has changed amongst the virtual machines to do load balancing [17]. The single virtual machine of CPU capacity (\(CA_{j}\)) is expressed in Eq. (17).

$$CA_{j} = NIRP_{j}$$
(17)

Equation (18) describes the capacity (C) of the physical machine.

$$C = \sum\limits_{j = 1}^{m} {CA_{j} }$$
(18)

The below equation calculates the load on single virtual machines.

$$L_{VMj} = \sum\limits_{k = 1}^{i} {SL_{k} }$$
(19)

Equation (19) calculates the load on single physical machines.

$$L = \sum\limits_{j = 1}^{n} {L_{VMj} }$$
(20)

The below equations explains the processing time of virtual machine and physical machine.

$$RS_{j} = \frac{{L_{VMj} }}{{CA_{j} }}$$
(21)
$$RS = \frac{L}{C}$$
(22)

The host is in a balanced state based on the value of the RS parameter. Equation (23) computes the workload standard deviation.

$$\sigma = \sqrt {\frac{1}{n}\sum\limits_{j = 1}^{n} {(RS_{j} - RS)} }$$
(23)

The priority for each task is applied to any virtual machines based on the physical machines [18]. Compared with the computed standard deviation \(\sigma\), the virtual machine is in a balanced state when the task is executed. the threshold is set from 0 to 1 and the standard deviation is within the threshold limit.

5 Proposed HFTO Algorithm for Resource Allocation and Load Balancing in the Cloud

The loads are balanced in the cloud data center and the optimum resource is assigned using the proposed HFTO algorithm. Virtual machines that are present on a single physical host are the subject of intrahost load balancing [19]. The mapping across load scenarios is balanced as a result of a thorough examination of the HFTO algorithm. Table 2 describes the relationship between the HFTO algorithm scheduling scenario and its behavior.

Table 2 The mapping among the HFTO algorithm scheduling scenario with its behavior
Full size table

5.1 Virtual Machine Classification

The proposed HFTO algorithm with its working process is delineated in Fig. 1. On the basis of the proposed model, load balancing tasks are removed from the workload underloaded VM, overloaded VM, and balanced VM. The VMs are mainly grouped based on their upper and lower boundaries. The upper boundary comprises 70% of the physical machine capacity whereas the lower boundary comprises 20% of the physical machine capacity. The overloaded VMs are represented using the \(\left( {L_{VMj} > 0.8*C} \right)\) and the underloaded VM are expressed using the \(\left( {L_{VMj} < 0.3*C} \right)\) criteria. The tasks are balanced by taking the tasks waiting in the overutilized VM and scheduling them in the underutilized VM. The priority and residual completion time are two essential factors that must be taken into account in order to reduce the task's makespan and response time.

Fig. 1
figure 1

Proposed HFTO algorithm for task scheduling and load balancing

Full size image

The proposed HFTO task scheduling paradigm seeks to maximize Quality of Service (QoS) metrics, including fault tolerance, response time, effectiveness, and makespan. The high-priority tasks are mainly taken both for pre-emption and exemption. The task is mainly exempted in a scenario where a high-priority task is executing in a particular VM to minimize the failure in terms of QoS on the service provider side. In this way, the Service Level Agreement (SLA) is not violated.

5.2 Task Scheduling

The optimal task scheduling is carried out using the HFTO algorithm. The main challenge with task scheduling is determining how to execute jobs that have been removed from virtual machines. The tasks withdrawn from the VM are mostly assigned to the new VM, which has a lower task load and tasks with varying priority. Equation (24) frames the objective function.

$$F(y) = \min \left( {MS_{t} + L_{{VM_{j} }} } \right)$$
(24)

The load on the VM is \(L_{{VM_{j} }}\) and the total number of tasks in the queue is \(MS_{t}\). The randomly dispersed solution in the search space is derived by Eq. (25).

$$y_{j,k} = y_{j,\min } + random[0,1] \cdot \left( {y_{j,\max } - y_{j,\min } } \right)$$
(25)

The upper and lower bounds with respect to \(y_{j,k}\) are \(y_{j,\max }\) and \(y_{j,\min }\). Based on the overloaded virtual machine groups, the female firebug updates the position by applying Eq. (26) after generating the initial population.

$$u_{j,k} = y_{j,k} + \lambda_{j,k} \cdot \left( {y_{j,k} - y_{i,k} } \right)$$
(26)

Equation (27) delineate the fitness function of a male firebug after updating the position.

$$R_{j} = \frac{{F(y_{j} )}}{{\sum\nolimits_{j = 1}^{m} {F(y_{j} )} }}$$
(27)

The population initializations of firebugs (FB) are described in Table 3. The matrix values are mainly the VM IDs present in the underutilized VM. The overall number of tasks largely indicates the tasks that were removed from the overused VMs. The selection probability of HFTO is higher when the fitness value is higher. The optimal scheduling is obtained after reaching the maximum iteration. However, the priority of the task running is considered when performing preemptive scheduling. The optimal scheduling plan mainly comprises the ID of the VM where the removed task is executed.

Table 3 Population initialization
Full size table

5.3 Optimal Scheduling with Preemptive Scheduling

Preemption happens when the removed task's priority is higher than the next task waiting in the queue. In certain cases, the running task's remaining estimated completion time is longer than the eliminated task burst time. Once the requirement is met, the removed task is carried out by the virtual machine. The preempted task state is preserved by the checkpoint mechanism. If the priority of the incoming task is greater than the priority of the task currently running in the VM, then it is executed first, or else it waits in the target VM queue. Next, the execution time of both tasks is checked, if the running task execution time is lower than the prioritized task, then it is executed. The higher-priority task is completed first if its running time exceeds that of the remaining task; otherwise, its current state is maintained. After the task allocation process is complete, the VM is assigned several parameters, such as the number of tasks provided, their priority, updated overloaded and underloaded VMs, etc. We mainly formulate the proposed model on the assumption that at least one physical machine may fail. The task scheduling module runs completely before the deadline. The makespan is computed by taking the Module Processing Time (MST) and Data Transmission Time (DTT). The MST is the variation between the ending time and task execution deadline. The DTT and MST values are normalized via different weight coefficients x and y. Since the MST and DTT both have equal importance, the coefficients x and y are allocated the same weights.

$$MP = x * DTT + y*MST$$

The checkpoint restore model is used here for fault tolerance.

6 Result and Discussion

A cloud simulation toolset, including CloudSim [26, 27], supports both behavior and system models of cloud system components, including resource provisioning rules, virtual machines, and data centers. CloudSim also shows how well the proposed model performs. At least four virtual machines are running on each of the 20 hosts in a single data center. The task length is in the 400–1 billion instruction range. When the virtual machines are started and scheduled to run, 300 cloudlets are created. With one CPU core, each node has minimum and maximum processing capabilities of 500 MIPS and 20,000 MIPS, respectively. 50 individuals comprise the population, and there is 100 iterations total. The parameters settings of the proposed model are described in Table 4.

Table 4 Proposed parameter settings
Full size table

6.1 Makespan Performance

Makespan mainly computes the overall time required to schedule the input tasks by selecting an optimal VM in the cloud and it is mathematically computed as follows:

$$N_{j} = \max \left\{ {CS_{j,k} /j \in S,\,\,\,j = 1,2,...,m\,\,\,and\,\,\,k \in virtual\,machines,\,\,k = 1,2,...,n} \right\}$$
(28)

In terms of makespan, Fig. 2 depicts the results achieved by comparing the proposed model to conventional techniques. Depending on the makespan in seconds, the number of tasks ranges from 100 to 500. The proposed method efficiency is validated using the significant performance measure called makespan. The makespan is calculated using Eq. (28). The time taken with the virtual machines to finish each task execution determines the makespan. The state-of-art comparison is carried out using the proposed method with existing Hybrid Tabu–Harmony task scheduling (HTHTS) [9], Service level agreement-based Load Balancing (SLA-LB) [10], Energy‐efficient task‐scheduling (EETS) [12], and Improved Firework Algorithm (IFA) [13]. Because of the constant number of virtual machines, the number of jobs increased by increasing the execution time. The proposed method demonstrated less makespan than other existing techniques such as HTHTS, SLA-LB, EETS, and IFA [28,29,30]. The proposed model minimizes the makespan by the heuristic-based load balancing algorithm created.

Fig. 2
figure 2

State-of-art comparison of makespan

Full size image

Figure 3 delineates the state-of-art result of makespan based on different task types. The four task types (Tj) are plotted on the horizontal axis such that \(j \in \left\{ {1,2,3,4} \right\}\). When the task types changed from 1 to 4, so did the makespan values of the state-of-the-art approaches, including HTHTS, SLA-LB, EETS, and IFA with the proposed model. Compared to the existing methods, the proposed technique provided better makespan results based on varying makespan. The shorter the makespan, the better the service quality and the better the scheduling. From the below equation, the task completion time in the cloud is \(\mu_{j}\).

$$N = \frac{{\min \sum\nolimits_{j = 1}^{M} {\mu_{j} } }}{M}$$
(29)
Fig. 3
figure 3

State-of-art comparison of response time based on task type

Full size image

6.2 Response Time Performance

A higher response time taken by a model specifies its weak capability of it in handling the overloaded VMs. Hence, a reliable model needs to have a minimal response time as possible to effectively manage the load in the network. Figure 4 displays a cutting-edge comparison of response time based on the number of tasks. Depending on the makespan in seconds, the number of tasks ranges from 100 to 500. For user satisfaction, the measurement criterion is response time. Equations (30) and (31) is used to calculate the response time. The low and medium-priority jobs are removed from overloaded virtual machines. When compared to the existing techniques, the amounts of task migrations are fewer. In comparison to existing approaches such as HTHTS, SLA-LB, EETS, and IFA, the proposed method obtains minimum makespan values, as shown in Fig. 4.

$$FS_{j,k} = \frac{{SL_{k} }}{{C_{j} }}$$
(30)
$$PS_{j,k} = CS_{(k - 1)j}$$
(31)
Fig. 4
figure 4

State-of-art comparison of response time based on the number of tasks

Full size image

The state-of-art comparison of response time with respect to arrival rate is described in Fig. 5. The existing Hybrid Tabu–Harmony task scheduling (HTHTS) [9], Service level agreement-based Load Balancing (SLA-LB) [10], Energy‐efficient task‐scheduling (EETS) [12] and Improved Firework Algorithm (IFA) [13] with the proposed method validate the performance of proposed work. With an average response time of 200 to 1000 s, this test uses a variable arrival rate ranging from 20 to 100. The proposed method takes less average response time than other conventional methods including HTHTS, SLA-LB, EETS, and IFA.

Fig. 5
figure 5

State-of-art comparison of response time with respect to the arrival rate

Full size image

6.3 Fault tolerance Performance

The two most critical Fault Tolerance tasks are fault recovery and fault detection. To perform the former in a decentralized manner, all hosts place a monitoring ring on top of a peer ring and observe their counter-clockwise neighbors. The backup host will have all failed node tuples, which will also validate the given status. Backup nodes are never given additional resources; they are only utilized for recovery. The backup node gets overburdened as a result of or during the takeover. The offload initiates the scale-out or adaptive load balancing procedure without interfering with routine operations. The failure rate of different techniques is identified and the results are presented in Fig. 6a and b in terms of network usage and delay. The network usage is computed in terms of kilobytes and the delay is computed in terms of milliseconds. For a failure rate of 0.1%, the proposed model offers better performance in both delay and network usage.

Fig. 6
figure 6

Impact of failure rate. a Network usage and b Delay

Full size image

6.4 Performance Efficiency

The efficiency of the different models taken for comparison is mainly computed using the prediction time and prediction accuracy. The prediction time primarily computes the time required to allocate the appropriate task to the intended usage within the user-specified time limit. The prediction accuracy mainly implies the total amount of tasks in the queue that are executed within the limited time specified by the user to the overall tasks present in the cloud. Hence an efficient methodology should consume minimal prediction time and higher prediction accuracy. The predicted task state is measured using a minimum and maximum task arrival state threshold value. The state-of-art comparison of overall efficiency is delineated in Fig. 7. This analysis is conducted in the context of user requests arrival rate and time. The proposed technique is compared to existing Hybrid Tabu-Harmony task scheduling (HTHTS) [9], Service level agreement-based Load Balancing (SLA-LB) [10], Energy-efficient task scheduling (EETS) [12], and Improved Firework Algorithm (IFA) [13]. The proposed method demonstrated superior efficiency results to other existing methods including HTHTS, SLA-LB, EETS, and IFA.

Fig. 7
figure 7

State-of-art comparison of efficiency

Full size image

6.5 Load Balancing Performance

The load balancing capability of any strategy is determined by its ability to maximize resource utilization while minimizing total completion time. A system with a higher load balancing capacity has a higher task configuration capacity by identifying the optimal VM for task allocation. In this way, the resource utilization capability is increased when the number of incoming tasks with different sizes is managed correctly by the VM assigned. The load balancing performance results are delineated in Fig. 8. The load balancing test checks the task scheduling performances provided. Figures 8a–c depict the load balancing analysis in terms of CPU consumption, I/O utilization, and memory utilization. The proposed method is used to compare load balancing performance with existing Hybrid Tabu-Harmony task scheduling (HTHTS) [9], Service level agreement-based Load Balancing (SLA-LB) [10], Energy-efficient task scheduling (EETS) [12], and Improved Firework Algorithm (IFA) [13]. The HTHTS, on the other hand, selected idle resources over SLA-LB, EETS, and IFA. However, the proposed work allocates resources more efficiently than existing methods. Based on the investigation of Fig. 8a–c, the proposed techniques take less CPU utilization, Input–Output utilization (IOU), and memory utilization. We mainly took the terms such as CPU utilization, IOU, and memory into consideration because they are mainly related to resource wastage. An effective load balancing scheme should resist resource wastage as maximum as possible.

Fig. 8
figure 8

Load balancing performances, a CPU utilization, b Input Output utilization, and c Memory utilization

Full size image

7 Conclusion

In this article, a task scheduling framework using Hybrid Firebug and Tunicate Optimization (HFTO) technique is presented. CloudSim simulates the proposed model's performance. The state-of-art comparison is carried out using the proposed method with existing Hybrid Tabu–Harmony task scheduling (HTHTS), Service level agreement-based Load Balancing (SLA-LB), Energy‐efficient task‐scheduling (EETS), and Improved Firework Algorithm (IFA). Computationally intensive tasks are assigned to resources with low CPU utilisation, whereas lightweight tasks are assigned to resources with high CPU utilisation. The HFTO algorithm improves the Quality of Service (QoS) factors fault tolerance, response time, efficiency, and makespan. In terms of cloud task scheduling performance and load balancing efficiency, the proposed strategy outperformed the previous techniques.