1 Introduction

Big data analysis is a process of handling complex data set that is generated by online social networks and scientific research [1, 2], using advanced tools to produce meaningful information in a cloud environment [3]. The huge volume of data are processed using a MapReduce programming model [4] that executes large number of parallel task on the cluster of physical machines (PM). MapReduce divides the job into multiple tasks that are executed in parallel manner on data elements placed in various PM in a cluster. Apache Hadoop is an open source software framework that exploits the MapReduce programming model for processing big data in a cloud environment such as Amazon, Oracle, Microsoft, and IBM [5].

A scheduling algorithm influences the performance of MapReduce framework in a computing environment as it manages the large cluster of hardware machines, controls the allocation of data and task on these hardware machines [6]. The scheduler is expected to satisfy various quality metrics such as data locality, fairness, throughput, response time, availability, Energy efficiency, CPU utilization, memory utilization, disk utilization, and security metrics [6]. Here, data locality is one of the important quality metrics and a critical factor that affects the performance of MapReduce clusters [7]. It reduces the cross-rack data movement and improves the throughput as a task is executed in a data-local manner. Fairness is providing equal opportunity to users by sharing the resources among them in a fair manner. For each user, the scheduler must ensure fairness by assigning the required number of computational resources. Throughput involves the number of request (jobs/tasks) completed over a specific period of time and it is achieved by improving the data locality and effective utilization of resources [8]. Effective resource utilization involves improving the resource usage and minimization of the system states being idle in a cluster.

The MapReduce scheduler plays a key role for meeting the quality metrics to confer a consistent performance in the heterogeneous environment. However, designing a scheduler to satisfy all the quality metrics is a challenging and complex process. For example, if a scheduling policy enforces strict data locality constraints then it has to compromise fairness since, a node having the data of jobs may not exist at the head of the queue based on fairness constraints [5]. In a MapReduce model, scheduling is a multi-objective optimization problem that is NP-hard in nature [9]. Further, a scheduler is also expected to be adaptive for allocating the resources based on availability of the system and capacity of PM in the cluster. Hence, an effective locality and replica aware scheduling (ELRAS) strategy is proposed based on the requirements with the following features:

  • ELRAS strategy adapts itself for the heterogeneous clusters and achieves higher data locality rates during data intensive computations.

  • A data placement strategy is formulated to compute the available space in the nodes dynamically and for positioning the data. Then, the task scheduling algorithm is designed to satisfy the data locality constraints and placing the task in the node that holds the data.

  • An autonomous replication scheme (ARS) is proposed and it is used in combination with ELRAS to perform effective replication decision.

  • ELRAS strategy satisfies the quality metrics such as execution time, data locality, resource utilization and throughput.

2 Background Research

2.1 Hadoop MapReduce Model

The Hadoop system architecture is illustrated in Fig. 1 and it is based on two important components: Firstly, the MapReduce programming model is utilized for parallel computations and then Hadoop Distributed File System for storing the data [10]. The MapReduce framework offers the execution environment for MapReduce jobs and a cluster of task trackers are employed for computations that is controlled by one centralized job tracker. All jobs are submitted to the job tracker and subsequently it is divided into multiple map and reduce tasks. The scheduler in the job tracker allocates the tasks to various task trackers for execution. Hadoop also implements the shuffle service using two phases such as shuffle and reduce tasks. Shuffle phase fetch the map outputs related with a reduce task from several nodes and combine them into one reduce input. An external merge sort algorithm is utilized as the intermediate data is too large to fit in memory.

Fig. 1
figure 1

The overview of Hadoop architecture

Full size image

Hadoop distributed file system (HDFS) has one name node to save the metadata and multiple data nodes and that are used for storing the data blocks. Hadoop is deployed on physical clusters in which typical deployment includes two master nodes (job tracker and name node) and several slave nodes to run the task tracker. Here, a slave node is divided into a fixed number of map and reduce slots, also a slot-based task scheduler is applied for scheduling. During the execution of a MapReduce job, the slave nodes periodically check the availability of task slots and requesting the master node for new tasks.

2.2 Related Work

MapReduce Scheduling is a multi-dimensional problem as it needs to optimize a set of quality metrics in a heterogeneous environment and workload constraints. The data placement, task scheduling and replication mechanism also influence the performance of the schedulers. Related work is presented by focusing the various dimensions of the scheduling strategies.

2.2.1 Scheduling Policies Based on Quality Metrics

Hadoop uses three basic scheduling policies namely FIFO, fair-share and capacity scheduler. The FIFO algorithm [11] gives preference to the workload that arrives earlier and here starvation problem occurs because of small jobs has to wait for longer jobs to complete execution. The fair-share (FS) [12] and capacity scheduler (CS) [13] equally distributes the computing resources among the users to ensure strict fairness. FS allocates one pool to each user in the system, and the resources are shared equally among all the pools. CS employs queues as an alternative method instead of pools, where each queue is allotted to a user and resources. In the recent literatures, several scheduling strategies are proposed by optimizing the basic schedulers or new scheduling strategies based on the combination of various quality metrics as an objective. Data locality is essential for data intensive computations to reduce the data transfer and for improving the throughput [14]. Further, the overall throughput can be improved by reducing the response time during execution. It is a complicated process to ensure a high throughput while adapting to varying workloads and fluctuating resource characteristics.

Earlier, delay scheduling [15] was proposed to deal with the conflict between locality and fairness needed for sharing the resources. According to the fairness constraints, a scheduled job couldn’t initiate a local task and further it delays the execution and launches the other task. Recently, Purlieus architecture [16] is used for allocating resources while a MapReduce application is executed in a cloud environment. It also suggests that data placement and virtual machine placement greatly influence the execution time of MapReduce application. The classification and optimization scheduler for Hadoop (COSHH) [17] classifies jobs using a k-means clustering algorithm concerning to the recorded execution times. Then, it determines the allocation matrix for jobs using a linear programming approach by maximizing the resource utilization. A hybrid algorithm [18] maintains the throughput of dynamic MapReduce by various scheduling algorithms for different combinations of workload volumes. During low workload periods, it uses the FIFO policy to achieve lower completion times and FS as the system load is balanced. As the workload volume is high, it uses the COSHH approach to achieve significant fairness level and better average completion time (ACT).

The Quincy scheduling algorithm [19] maintains a balance between data locality and fairness for the incoming workload. Here, a bipartite graph is used to represent nodes, jobs, and tasks while a weight is assigned based on the amount of data transferred for the edge that connects the node and task. Then, a scheduling plan is devised by computing the minimum flow cost in the graph under the fairness constraint. In a baseline scheduler [20], the scheduling decisions are made without any global cluster information and the duplications are made to the idle nodes for load balancing and ensuring locality constraint.

2.2.2 Scheduling Policies Adapting to Heterogeneous Environment

The presence of large number of heterogeneous nodes causes various difficulties for effectively utilizing the resources such as CPU, network I/O, and disk I/O [21]. Reducing the task execution time also significantly improves the resource utilization [22]. The scheduling algorithm focusing resource utilization verifies the available bandwidth of different types of resources on a machine against the requirements of a job before scheduling to minimize busy waiting and resource utilization skewness.

Late scheduler [23] enhances the performance of MapReduce framework by reducing the overhead of speculative task in a heterogeneous environment and also it computes the completion time of an application speculatively. The triple-queue scheduler [24], first allocates the map tasks of a fresh job to an available slot then, using the execution data, the scheduler moves the job to a CPU-bound or a disk IO-bound queue. The scheduler continuously monitors the execution times of tasks to move jobs dynamically from one type of queue to another. This scheme is proven to achieve an overall improvement in throughput.

2.2.3 Replication Mechanism

Data replication is an important optimization methodology to handle large volumes of data thereby replicating the data at different locations and achieves higher data availability. Hierarchical cluster scheduling (HCS) algorithm is proposed for data intensive computations in a grid environment [25]. HCS algorithm is also used in combination with hierarchical replication scheme (HRS) to achieve improvement in the execution time when used for data intensive computations. Here, the author(s) had suggested that an effective replication scheme and a scheduling strategy reduce the overhead and job execution time irrespective of the environment (homogeneous or heterogeneous).

HRS [25] decides the number of replica based on the previously executed data set and it is used with HCS to attain optimal results. bandwidth hierarchy based replication (BHR) [26], [27] tries to increase the necessary data in the same region to fetch the replica easily since, the bandwidth within the region is high. Moreover, the regional popularity of files is computed and the best replica for a job is identified through BHR. Least recently used (LRU) replication mechanism replicates all the files and it removes the file that has been used recently [28]. Some of the related works from the recent literatures are listed in Table 1.

Table 1 Comparison among the existing scheduling policies in MapReduce Framework
Full size table

3 Problem Analysis and System Design

Data locality improves the performance of MapReduce framework in heterogeneous clusters by placing the computations in nodes that hold the data blocks. Heterogeneity has an impact on Hadoop scheduler because in a cluster, the node parameter always fluctuates. Hence, the solutions proposed for improving the data placement, task scheduling and replication mechanism is summarised as follows:

  • An adaptive scheduling algorithm named ELRAS that identifies the capacity of node in a heterogeneous cluster environment is proposed. A three layer mapping relationship is used to acquire cluster configuration and node statistics information dynamically. Further, a method for tracking the data objects is integrated with the scheduler.

  • An ARS is proposed to determine the data objects to be replicated and the location for placing the replica.

  • The environment and task are heterogeneous because hardware parameters and workloads always vary with respect to time. Hence, the cluster setup for evaluation is configured with heterogeneous nodes and the workload characteristics.

3.1 Proposed Work

3.1.1 Overview of Proposed Work

The architecture of the ELRAS strategy is presented in Fig. 2 and sequence of operations is described in Fig. 3.

Fig. 2
figure 2

Architecture of ELRAS strategy

Full size image
Fig. 3
figure 3

Sequence of flow of operations in the ELRAS architecture

Full size image

  • In Fig. 2, the JobTracker holds the node statistics table (NST) that includes entries such as rack id, physical machine id, data object id, replica object id, task id of running task, CPU utilization rate, unused storage space (Si) and status flag.

  • Scheduler is located in JobTracker to make data placement decisions, task scheduling decision, replica node decision and to collect the results from the slave nodes.

  • The job queue maintains the list of jobs and task that is ready for execution. TaskTracker (slave node/data node) executes the task and returns the results. Further, each slave node executes the ARS for deciding the object for replication and issue replication request to the JobTracker.

3.1.2 Efficient Locality and Replica Aware Scheduling Strategy (ELRAS)

  1. (1)

    Creation and Modification of Node Statistics Table (NST)

The nodes in a cluster are organized according to a three layer mapping relationship (rack layer—physical machine layer—virtual machine layer) as illustrated in Fig. 4. When new PM are added to the network, the parameters such as Rack Id, Node_Id of the physical Machine, identity of currently running task, data objects stored, CPU utilization rate, status of the physical machine (1-for overloaded), and available free space (Si) are added to the NST. The NST is located in the JobTracker and it frequently updates the information as a new physical machine (PMi) is added or whenever an update request (URi) arises from the existing PM. Algorithm 1 describes the dynamic creation and modification of NST and Table 2 is an illustrative example for NST.

figure c
Fig. 4
figure 4

Mapping relationship between different layers

Full size image
Table 2 Example for node statistics table
Full size table
  1. (2)

    Initial Data Placement and Replica Placement

Once a file is submitted by a user for processing, it is divided into multiple blocks based on the size specified by the user or fixed default size. The number of data objects depends on the file size and the specified block size. Each data objects are placed in the available free space of the PM, starting from the first rack and the available capacity is identified through the NST. For each data placement decisions, the NST is updated and when replication request comes for a data object, the replica is placed in a node wherever the primary copy of the data object does not exist.

The proposed algorithm solves two key issues associated with data locality, namely (1) identifying the PM compatible for a data set (data placement), (2) placing the virtual machines in the identified node. The tasks are placed in the form of virtual machines (VM) in the computing nodes. For example, the data object placement strategy in a compatible free slot is displayed in Fig. 5. At this point, PM5 consists of free space and the data objects A3, B3 and C3 are placed in PM5. Algorithm 2 describes the process involved in the data placement. Here, γD ij is the computing and storage space required for placing data and task. Where,γD ij is identified from the ACT divided by the number of task slot.

figure d
Fig. 5
figure 5

Data placement strategy

Full size image
  1. (3)

    Task Scheduling in Physical Machines

In the task scheduling process, the user submits a job for execution and each job is placed in a job queue. It is further divided into sub-tasks and placed in the task queue and a Task_Id is allocated. Then, the tasks are scheduled in the PM that are not overloaded and hold the data objects or replica. Overloaded status is identified from the NST where the status flag of the corresponding overloaded node is set to one. Algorithm 3 describes the process involved in the task scheduling to the corresponding PM in clusters.

figure e
  1. (4)

    Replica Node Decision

The algorithm 4 describes the mechanism for identifying a PM to place the replica object. If NST receives a replication request of a data object, then a node is selected based on the availability with a constraint. Initially, the replica object and data objects will be scheduled in PM at different racks, and if it is not possible then an alternate PM in the same rack is identified. Subsequently, if a node is not available for placing the replica, then the task waiting for execution is set to wait status until the PM holding the data object is freed from overloaded status.

figure f
  1. (5)

    Autonomous Replication Scheme (ARS)

Assumption 1

Replicating the data objects only when it is necessary, consistently reduces the overhead.

Assumption 2

It is possible to replicate the data objects possess high hit rates by computing its popularity.

The algorithm 5 (ARS) is executed at each node and it decides to issue the replication request for a data object based on the load processing capacity of the PM. When the task arrival for the data object is too high, it decides further to replicate. The popularity of the data object λ(D ij ) is computed by adding the request arrival rate. When it exceeds the ideal request arrival rate of the physical machines (λ(PM i )), then the replication decision is made. As the decision for replication is made, a request is issued to NST and the task in the queue with the wait status is triggered for scheduling. The ideal request rate of a PMi is expressed as

$$\lambda (PM_{i} ) = \frac{{C_{i} }}{{\sum\nolimits_{k = 1}^{M} {C_{k} } }}\lambda$$
(1)

where, C is the computing capacity of the node and λ is the average arrival rate. The popularity of the data object Dij is computed as follows using Eq. (2),

$$\lambda \,(D_{ij} )\, = \,\sum\limits_{j = 1}^{N} {\lambda_{ij} }$$
(2)

In a PM, once a replica is placed the time stamp is triggered automatically and it checks for the arrival of task within the limit of time stamp. As soon as the time stamp exceeds the limit, the replica is removed automatically.

figure g

4 Testbed and Evaluation

4.1 Experimental Setup

In the experimental work, a heterogeneous cluster is built in the Amazon EC2 environment and the corresponding node configuration is presented in Table 3. The Hadoop 1.0.0 is used for all the experiments and the block size of file system is configured as 64 MB. The ELRAS algorithm is implemented and its performance is evaluated in a heterogeneous environment with 28 nodes and different configurations. Here, the scheduler in the Hadoop package is modified with ELRAS for the evaluation of proposed approach.

Table 3 Amazon EC2 configuration
Full size table

4.2 Benchmark used for Analysis

HiBench [36] is a benchmarking suite and its workloads such as WordCount, TeraSort, Grep and K-Means clustering are used to test the performance of proposed framework in the Hadoop environment. The WordCount benchmark reads the input text files and computes the occurrences of each word from the input file. The grep command is employed for matching the patterns in plain-text datasets [36]. The TeraSort benchmark is used to sort the huge dataset as quick as possible and it involves the following processes: (1) Generating an input data set by TeraGen module, (2) Executing the task on the input dataset, and (3) Validating the result using TeraValidate module. The k-Means clustering is an iterative approach that is implemented as a series of MapReduce rounds. It partitions ‘n’ objects into k-clusters where each object belongs to the cluster with the nearest distance. The input to the algorithm is an initial set of cluster centres and a set of objects represented as d-dimensional vectors.

4.3 Results and Discussion

4.3.1 Heterogeneous Cluster

To evaluate the performance of ELRAS approach using Hadoop, a default scheduler is modified with the proposed method in the Hadoop package. The proposed method is evaluated by using workloads presented in Table 4. Firstly, to analyse the efficiency of the schedulers in a lightly-loaded system, a sample workload with ten jobs is evaluated. Then, multiple experiments are conducted using various workloads by increasing the total number of jobs in the workload to investigate the performance variation.

Table 4 Application characteristics of workload in terms of data size
Full size table

The ELRAS algorithm is compared with various scheduling policies for different performance metrics. The existing competent algorithms such as Maestro, HCS, Apollo, Baseline, and delay schedulers are compared with ELRAS for validating the results and it is discussed in the subsequent sections. In addition, the existing replication schemes such as LRU, BHR, and HRS are compared with the proposed ARS for comparing the performance variations. The heterogeneity in workloads is proved by testing the algorithms for mixed workloads (i.e.), combination IO bound and CPU bound workloads. It is observed that the algorithm performs smoothly for mixed workloads and the results are presented with respect to the total job execution time, cross rack communication, computational resource usage, data locality, and throughput.

4.3.2 Total Job Execution Time

The total job execution time is the combination of queuing time, access time and execution time. It also includes the time required for moving a file to a desired location until the execution process completes. Figure 6 presents the total execution time of a job for an input data size 100 GB and 120 jobs. The replication schemes such as LRU, BHR, HRS and the proposed ARS are used in combination with ELRAS and hierarchical cluster scheduling (HCS) strategy. ELRAS scheduling strategy with the ARS provides excellent results as compared with other strategies.

Fig. 6
figure 6

Total job execution time of ELRAS and HCS with different replication schemes for workload. a Word count, b TeraSort, c Grep, d K-means

Full size image

In the proposed ELRAS scheduling strategy, while it is accompanied with BHR and HRS replication mechanisms a small deviation in the total execution time is observed. If the HCS + ARS combination is compared with HCS + HRS strategy, then it is noticed that no improvement in the total job execution time. Moreover, the file transmission time is the major factor to influence the total job execution time. Here, accurate scheduling and data placement approach in ELRAS effectively reduces the file transmission time. Thus, the presented results indicate the vital role of scheduling strategy and replication scheme in MapReduce model to reduce the total job execution time.

Figure 7a presents the total job execution time of the ELRAS scheduler as it is used with LRU, BHR, HRS, and ARS for multiple numbers of jobs (ranging from 10 to 120). It is observed that, ELRAS + ARS combination has less job execution time because of the replication scheme employed. The results indicate that replication is also one of the major factors affecting the scheduler in a fluctuating environment. Figure 7b presents the various existing scheduling schemes such as HRS + HCS, Apollo, Maestro, and Baseline compared with the ELRAS + ARS scheduler in terms of total job execution time. Here, ELRAS + ARS gives best performance as it is compared with existing scheduling policies. Figure 8 illustrates the performance of ELRAS scheduler in the heterogeneous and the homogenous environment and the observation shows that the proposed method is adaptable to fluctuating environment. For the heterogeneous environment, the configuration shown in Table 3 is used and for the homogenous environment Type 1 nodes shown in Table 3 (number of nodes—28) is selected.

Fig. 7
figure 7

Total job execution time for mixed workload and 100 GB dataset. a Replication Schemes + ELRAS strategy, b different scheduling schemes

Full size image
Fig. 8
figure 8

Total job execution time for mixed workload and 100 GB dataset of ELRAS strategy in heterogeneous and homogenous environment

Full size image

4.3.3 Computing Resource Usage

The computing resource usage is the percentage usage of all computing elements exists in the cluster. Scheduling the task to a node at which no required data exists that leads to increase in the access latency because of redundant data transfer. Figure 9a, b illustrate the computing resource usage for the mixed workload with 100 and 50 GB data size. ELRAS with ARS gives the maximum computational usage about 92% and while ELRAS is combined with HRS the computational usage is only about 82%. The maximum computational usage increases about 1% than the previous case as HCS strategy is combined with the ARS scheme. Figure 10 displays the computational resource usage for (100 GB dataset and 120 jobs) various existing schedulers that are compared with the ELRAS strategy. The results indicate that for a mixed workload category, ELRAS + ARS can be exploited to achieve a better performance gain in terms of computational resource usage.

Fig. 9
figure 9

Computational resource usage of ELRAS compared with HCS for different replication schemes. a 100 GB data and 120 jobs, b 50 GB data and 50 jobs

Full size image
Fig. 10
figure 10

Computational resource usage for different scheduling strategy for mixed workload (data size—100 GB and #jobs—120)

Full size image

4.3.4 Cross-Rack Communication

The number of cross-rack communications for each scheduling and replication mechanism is illustrated in Fig. 11. As the ELRAS + ARS scheme is used in combination, the number of cross-rack communications is reduced because of excellent data locality. The data object is replicated as the node holding the data object is in overloaded state and there is no space for scheduling the task. ELRAS approach balances the load properly and reduces communication overhead to the maximum thereby scheduling the task as well as data effectively.

Fig. 11
figure 11

Number of cross-rack communications of ELRAS compared with HCS. a Word count, b TeraSort, c Grep, d K-means

Full size image

4.3.5 Data Locality

Data locality is placing the task in the nodes where the input data resides. Locality is crucial for performance in large clusters because of network bisection bandwidth becomes a bottleneck [15]. Running the task in the same node where the data exists (data locality/node locality) is efficient and if it is not possible then running the task in the same rack (rack locality) is adopted. Figure 12 illustrates the data locality and rack locality of multiple map task as the experimentation are done with the mixed workload. Locality is achieved by replicating the hot files obviously that leads to queuing delay. Hence, for each Job separate task queue is maintained and the tasks are handled in parallel aspect.

Fig. 12
figure 12

Data-locality of different schedulers

Full size image

Networks in large clusters are organized in a hierarchical fashion as the nodes are grouped under series of racks. Each rack or switch can hold a maximum of 80 nodes at the lowest level [19] and the bandwidth inside a rack is higher than bandwidth between the racks. Table 5 shows the data locality and the rack locality of various Map task while executing ELRAS compared to delay scheduling. The delay scheduling yields 99% data locality and 100% rack locality for 10 Map task. Then, as the number of Map task is increased to 100, the locality rate decreases. When ELRAS is investigated for data locality, it achieves 99% data locality and 100% rack locality for 10 to 100 Map task. As the number of Map task is increased above 100, the data locality and rack locality decreases and it is negligible. Figure 12 illustrates the node and rack locality from 3 to 1000 Map task for various scheduling schemes. Different strategies such as FIFO, delay scheduling, and ELRAS are compared and the overall performance statistics are presented in Fig. 13.

Table 5 Performance comparison of delay scheduling and ELRAS for node and rack locality
Full size table
Fig. 13
figure 13

Data locality for each bin (Table 5) under mixed workload

Full size image

Figure 13 displays the data locality rate for a mixed workload case represented in Table 6. When the number of jobs is less, the data locality rate is below 30% for FIFO and FS schedulers. Delay scheduling performs better than FIFO and FS by achieving a locality rate of 72%. Nevertheless, the ELRAS outperforms delay scheduling by yielding an optimal locality rate about 82% for a minimum workload.

Table 6 Number of Jobs in mixed workload used for data locality measurement
Full size table

4.3.6 Throughput

The throughput of existing schedulers HCS + HRS is less because of data locality problem and interdependence between Map and reduce tasks. As the strict fairness constraint is enforced the data-locality becomes degraded [8]. In ELRAS + ARS, the load across the cluster is completely balanced that ensures high throughput since, there is no task waiting in the queue. Figure 14 shows the throughput for the time period about 10 min while running a single workload and mixed workload cases. ELRAS achieves highest throughput rate of handling 120 jobs within 10 min in a 100 GB dataset. In addition, ELRAS achieves competitive performance in throughput as compared with Delay scheduling.

Fig. 14
figure 14

Throughput comparison for K-means and mixed workload

Full size image

From the experiments conducted and based on the observations the highlights of ELRAS are summarized as follows

  • The proposed work (ELRAS) suggests that enhancing data locality and reducing the cross-rack communication significantly decrease the total job execution time of data intensive computations.

  • A tracking method integrated identifies the data objects that enable efficient scheduling and reduction in cross-rack communication. Further, ARS integrated with ELRAS decides the data object for replication and adaptively removes the replica when it is idle.

  • The computing resource usage rate is high for ELRAS strategy while compared with the existing scheduling and replication approaches.

  • The algorithm is designed for heterogeneous cluster environment as the node characteristics are varying and dynamic in nature. The presented results show that the total job execution time for varying workload characteristics is reduced as compared to other scheduling schemes.

5 Conclusions

The ELRAS algorithm presented in this article is simply adapted for MapReduce applications in Heterogeneous cluster environment. It focuses on locality and replica aware scheduling and demonstrates that the data locality improves the throughput and reduces the cross-rack communications. This algorithm is flexible to adapt for the dynamic environment while the new nodes are added or removed. An ARS is used to make decision on replication and replicas are tracked effectively using the NST. The various performance metrics such as total job execution time, computing resource usage, number of cross-rack communication and throughput are studied individually for the combination of ELRAS + ARS in a heterogeneous cluster environment. It is also concluded that the data placement method, scheduling strategy and replication scheme play a vital role for improving the performance metrics. The results proved the efficiency of the algorithm for heterogeneous clusters and workloads. As a future work, the algorithm can be integrated with Auto-Scaling applications that are used in the commercial cloud environments.