14
Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1 Kiran Kumar Pulamolu and 2 D. Venkata Subramanian 1 Hindustan Institute of Science and Technology, Chennai, India. [email protected] 2 Hindustan Institute of Science and Technology, Chennai, India. Abstract Hadoop is designed to process a storm of big data. Now a day, Hadoop is used not only in web indexing but also extended to process large volume of streaming data with heterogeneous data types. It efficiently handles any form of data such as structured, semi-structured and unstructured formats. In this Hadoop technology there are several challenges and one of them concerns the resource sharing with fairness. Fair resource allocation is one of the key building blocks of Heterogeneous Hadoop Cluster. Many researchers proposed different models, techniques and algorithms to progress the performance of Hadoop cluster by sharing the resources effectively. The existing frame works such as YARN, Mesos and Dryad uses Memory less resource fairness, Long Term Resource Fairness (LTRF) algorithms which are not suitable for cloud based Hadoop clusters. This paper proposes a novel resource sharing algorithm called HRSYARN (Heuristics based resource sharing with fairness) to share resources with fairness in YARN. Based on HRSYARN resource sharing is done among the Tenants of the same cluster. The experimental results show that the performance of Hadoop cluster is improved by 30% effectively with HRSYARN over other models. Key Words:Hadoop, big data, HRSYARN, resource fairness, resource sharing. International Journal of Pure and Applied Mathematics Volume 116 No. 22 2017, 491-503 ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version) url: http://www.ijpam.eu Special Issue ijpam.eu 491

Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Heuristics Based Resource Sharing with Fairness in

YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian

1Hindustan Institute of Science and Technology, Chennai, India. [email protected]

2Hindustan Institute of Science and Technology, Chennai, India.

Abstract Hadoop is designed to process a storm of big data. Now a day, Hadoop

is used not only in web indexing but also extended to process large volume of streaming data with heterogeneous data types. It efficiently handles any form of data such as structured, semi-structured and unstructured formats. In this Hadoop technology there are several challenges and one of them concerns the resource sharing with fairness. Fair resource allocation is one of the key building blocks of Heterogeneous Hadoop Cluster. Many researchers proposed different models, techniques and algorithms to progress the performance of Hadoop cluster by sharing the resources effectively. The existing frame works such as YARN, Mesos and Dryad uses Memory less resource fairness, Long Term Resource Fairness (LTRF) algorithms which are not suitable for cloud based Hadoop clusters. This paper proposes a novel resource sharing algorithm called HRSYARN (Heuristics based resource sharing with fairness) to share resources with fairness in YARN. Based on HRSYARN resource sharing is done among the Tenants of the same cluster. The experimental results show that the performance of Hadoop cluster is improved by 30% effectively with HRSYARN over other models. Key Words:Hadoop, big data, HRSYARN, resource fairness, resource sharing.

International Journal of Pure and Applied MathematicsVolume 116 No. 22 2017, 491-503ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu

491

Page 2: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

1. Introduction The use of the Hadoop framework to process data stored on high-performance computing clusters is continuously increasing. Apache Hadoop, an open-source software framework uses Map Reduce programming model for distributed storage and processing of big data sets. In Hadoop, the modules are designed in such a way that the framework can handle the hardware failures automatically. Hadoop runs on the top of a HDFS (Hadoop Distributed File System), which divides an input data into fixed sized (typically 64MB) multiple blocks and duplicates each data block several times across computing nodes. Resources can be efficiently used by sharing them among multiple tenants [1][5]. The chances for resource sharing are being increased with deploying diversifying and heterogeneous Big Data applications in huge number. Latest research shows that dynamic resource allocation techniques like resource multiplexing and over committing can expansively improve the resource utilization over T-shirt model [6][10]. It is also known as scale-up mode as it adds/ removes resources directly on the existing VMs. Map Reduce programming model is implemented to use a parallel, distributed algorithm to process big data sets on a cluster. It fits well for huge variety of data and computes intensive applications whose data need to be parallelized into multiple map tasks followed by a set of reduce tasks. Map Reduce possesses Map () method to filter and sort and a Reduce () method that performs a summary operation. The Map Reduce Framework organizes the process by running various tasks in parallel, assembling the servers, handling data transmissions among the different parts of the system and provides fault tolerance and redundancy [11][12].

Figure 1: Hadoop Architectural Diagram

Storage resource management, SRM is a hands-on technique, utilizing which the effectiveness and speed of the available drive space in a storage area network can be optimized [15]. It is a high-speed network that connects shared pools of storage devices to multiple servers. It enables a storage administrator

International Journal of Pure and Applied Mathematics Special Issue

492

Page 3: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

with configuration management and to automate data backup and recovery. It can also be used to analyze the performance of Storage Area Network (SAN), to predict storage needs for future and to understand how and where to utilize thin provisioning, storage pools and tiered storage [13][14].

Figure 2: Design of Hadoop Distributed File System

Background

The first generation of Hadoop YARN uses a fine grained resource organization, where the jobs submitted can configure their required resources like Memory and CPU basing on their independent workloads. In the case of insufficient resources in the cluster, the Resource Manager will reserve resources on the selected node for the task. When the other tasks on this node completes, the resources released will not be allocated to other applications until the reservation is satisfied. In this approach, the newly released resources are assigned to only those jobs which are having prior reservation which offers a good resource isolation capability and also leads to ineffective resource utilization.

Let’s consider five Tenants, Tenant A, Tenant B, Tenant C, Tenant D and Tenant E sharing a common cloud cluster. Assume that 50GB of RAM as 50 shares will be allocated to all the tenants. If a tenant doesn’t require all the resources, it can lend its unused resources to the resource that is in need.

In the Table 1 given below:

New Demand: Represents the number of current shares a Tenant demanded

Total Demand: The total number of resources/shares a tenant demanded.

Current Allocation: The Number current shares allocated to a tenant.

Total Allocation: The total number of resources consumed by a tenant.

Current preemption: The current resources which are shared or borrowed to or from another tenant.

International Journal of Pure and Applied Mathematics Special Issue

493

Page 4: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Total pre-emption: The total number of resources which are shared or borrowed to or from another tenant.

For example, at time t1, Tenant A requires only 20 resources among the 50 allocated to it whereas, Tenant B requires 100 resources. So, it requires 50 resources more and hence it borrowed the unused resources of Tenant A. Similarly, Tenant C, Tenant D and Tenant E require 40, 80 and 30 resources respectively. So, the unused resources of Tenant C and E are borrowed by Tenant D.

Table 2 describes the scenario At time t2, Tenant A requires only 40 resources among the 50 allocated to it. Tenant B requests 60 resources out of which only 50 shares are allocated to it. Tenant C demands only 20 shares. So, the unused resources of Tenant C are allocated to Tenant D which is in need. Tenant E demanded 10 shares out of 50 available shares.

Table 3 illustrates the scenario at time t3, Tenant A requires 70 resources. As Tenant A is not having ample resources so it will borrow the remaining 20 resources from Tenant B which it provided previous. Tenant B requests 60 resources out of which only 30 shares are allocated to it as it returned 20 resources to Tenant A. Tenant C requests 40 shares. Though Tenant D demanded 60 resources, only 40 resources are allocated as the remaining 10 are returned to Tenant E from which it borrowed earlier. Tenant E is in need of 70 resources. So, it borrowed 10 from Tenant C which left unused and 10 from Tenant D which it lends prior.

Table 4 defines the state at time t4, Tenant A requires 80 resources. As Tenant A is not having sufficient resources so it borrowed the remaining 30 resources from Tenant B which it provided before. Tenant B demanded only 20 resources, so the remaining unused resources are allocated to Tenant A. Tenant C requests 80 shares. Tenant D demanded 10 resources; the unused 40 resources are allocated to Tenant C and Tenant E which are in need. Tenant E borrowed the remaining 10 resources from Tenant D to meet its requirement.

Table 5 describes the situation at time t5, Where Tenant A and Tenant D requests 30 and 40 resources respectively. The remaining unused resources are assigned to Tenant E and Tenant C correspondingly. Tenant B exactly demanded 50 resources which are allocated to it.

Table 1: Resource Sharing among Tenants at Time t1 Time t1 Demand Allocation Preemption

New Total Current Total Current Total Tenant A 20 20 20 20 30 30 Tenant B 100 100 80 80 -30 -30 Tenant C 40 40 40 40 10 10 Tenant D 80 80 80 80 -30 -30 Tenant E 30 30 30 30 20 20

International Journal of Pure and Applied Mathematics Special Issue

494

Page 5: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Table 2: Resource Sharing among Tenants at Time t2 Time Demand Allocation Preemption

t2 New Total Current Total Current Total Tenant A 40 60 40 60 10 40 Tenant B 60 140 60 130 0 -40 Tenant C 20 60 20 60 30 40 Tenant D 70 150 70 150 -20 -50 Tenant E 10 40 10 40 40 60

Table 3: Resource Sharing among Tenants at Time t3 Time Demand Allocation Preemption t3 New Total Current Total Current Total Tenant A 70 130 70 130 -20 20 Tenant B 60 190 30 170 20 -20 Tenant C 40 100 40 100 10 50 Tenant D 60 210 40 190 10 -40 Tenant E 70 110 70 110 -20 40

Table 4: Resource Sharing among Tenants at Time t4 Time t4

Demand Allocation Preemption New Total Current Total Current Total

Tenant A 80 210 80 210 -30 -10 Tenant B 20 190 20 190 30 10 Tenant C 80 180 80 180 -30 20 Tenant D 10 200 10 200 40 0 Tenant E 60 170 60 170 -10 30

Table 5: Resource Sharing among Tenants at Time t5 Time t5

Demand Allocation Preemption New Total Current Total Current Total

Tenant A 30 240 30 240 20 10 Tenant B 50 240 50 240 0 10 Tenant C 60 240 60 240 -10 10 Tenant D 40 240 40 240 10 10 Tenant E 70 240 70 240 -20 10

Table 6: Heuristic Table T1 T2 T3 T4 T5 T1 0 10 0 10 0 T2 -10 0 -10 0 0 T3 0 10 0 20 0 T4 -10 0 -20 0 -10 T5 0 0 0 10 0

2. Related Work Shanjiang Tang et.al stated a model for Data-Intensive Computing in the Cloud to bring fairness in resource sharing. A number of computing frameworks have become popular by addressing the challenges of big data [1]. Chen Wang et.al proposed a technique namely Dynamic programming (DP) to deal with scientific applications [2]. This paper presented how parallel programming can be simplified with DPX10; a DAG-based distributed X10 framework. This enables the user to write efficient parallel DP programs with less effort [3]. Xu

International Journal of Pure and Applied Mathematics Special Issue

495

Page 6: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Zhen et.al projected a rapid development tool to handle parallel applications. ParDT is implemented using Java. It allows programmers to make use of graphical editors and automatic code skeleton to build parallel applications with less effort. The ParDT is implemented on the top of open source technologies. Sculley et. al introduced a model called CRR (combined regression and ranking). The author demonstrated the effectiveness of CRR on a wide range of tasks together with click prediction for online advertisements [4]. Wei Wang et.al proposed a model on DCRR (Dynamic Cloud Resource Reservation) through Cloud Brokerage Infrastructure-as-a-Service. This technique helps in offering varied cost options for on-demand and reserved instances with a variety of concessions to draw the attention of different cloud users. The paper stated a new cloud mediating service in which instances are collected from cloud providers in large number and serves users at low costs [5]. Wei Wang et.al implemented their work in Multi-Resource Fair Allocation in Heterogeneous Cloud Computing Systems. They studied the multi-resource allocation problem in Cloud Computing systems where the resource pool is constructed from a large number of heterogeneous servers, representing different points in the configuration space of resources such as processing, memory, and storage. They designed a multi-resource allocation mechanism, called DRFH that generalizes the notion of Dominant Resource Fairness (DRF) from a single server to multiple heterogeneous servers [6]. DRFH provides a number of highly desirable properties. Moshe Babaioff et.al proposed a Framework for Economic Resource Allocation for the Cloud. Cloud computing reached a significant maturity from system perspective and allocation of costly hardware resources. In ERA we aimed to increase efficiency of cloud resources according to economic principles and it enables the development of scheduling and pricing algorithms to concert lower level cloud infrastructure [7]. M. Lavanya et.al proposed Adaptive Core Allocation for Map-reduce Clusters in Cloud Computing in which Map Reduce is vital for hadoop and for data intensive application by using cluster of machines. As, fixed slot-based resource management model suffering from performance degradation, Author proposed DRM (Dynamic Map Reduce) to overcome the drawbacks with traditional techniques [8]. Zhihong Liu et.al proposed DRM for Map Reduce with Partitioning Skew. Since, Map Reducer scheduler is suffering from an issue called partitioning skew where, map task is unevenly distributed among reduce tasks [9]. The author presented DREAM, a framework for runtime partitioning skew to repartition among reduce tasks. Wei Wang et.al proposed a model, OOMIA (Optimal Online Multi Instance Acquisition) in Iaas (Infrastructure as a service) clouds. It offers varied purchasing options. A user can either pay for the resources they reserved or run resources on demand. The major issue with this technique is to combine these two methods dynamically to serve users time varying demands at feasible cost. In this, the author stated two algorithms, one deterministic and another randomized, that dynamically combine the two instances options online.

International Journal of Pure and Applied Mathematics Special Issue

496

Page 7: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

3. Proposed Method HRSYARN scheduler enables the tenants to share resources with fairness in Multi node cloud clusters. The primary advantage of the newly proposed system is to share resources between multi tenants based on the resources available in the cluster by updating heuristic table to bring fairness. The heuristic table maintains the information regarding the resources that a tenant lend and borrowed to and from other tenants. This interest free loan lending model helps in attracting tenants towards sharing resources. Weighted arithmetic mean is used to calculate average resource requirement of all the individual tenants belonging to a cluster before sharing resources among them. Weighted Arithmetic Mean

• The weighted arithmetic mean is a measure of fundamental affinity of a set of quantitative observations with different importance for these observations.

• Each observation is weighted depending on its importance relative to other observations.

• The weighted arithmetic mean is calculated by dividing the summation of the products of observations and their weights with total weight.

Mathematical Definition

Formally, the weighted mean of a non-empty set of data which means:

HRSYARN Pseudo Code

The below algorithm demonstrates the Pseudo code of HRSYARN. In this, RA represents the set of resources available with the cluster and RAi denotes the resources allocated to tenant i. Let RD be the resources demanded by the tenants at time t whereas, RDi represents the resources demanded by Tenant i. If W represents the set of Weights of all the workloads, Wi stand for the weight workload of Tenant i. The main aim of HRSYARN is to share resources with fairness among multiple tenants of a cluster. So, this technique uses Weighted Arithmetic Mean (WAM) to bring fairness in sharing resources. To exhibit the working of the algorithm, let’s consider two Tenants A and B. If both the tenants were allocated 50 RAM shares each and at time t0, if A requires only 20shares whereas B requires 70. Then B will borrow the unused shares from A. Later if A requires excess resources than its share, then it will ask B to return the resources. At this moment, if B has unused resources it will return them to A. Otherwise, it calculates the number of resources to be returned to A based on Weighted Arithmetic Mean (WAM) and UP value (UP: Resources preempted by a resource or tenant). Weighed arithmetic mean (WAM) is calculated by dividing the summation of product of each individual tenant’s weights and resources demanded with summation of weights of all the tenants.

International Journal of Pure and Applied Mathematics Special Issue

497

Page 8: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Total resources demanded by A will be allocated to it if WAM calculated is less than available resources and UP is greater than zero. 50% of the resources demanded will be assigned to A if WAM is nearly equal to available resources and UP is greater than zero. 25% of the resources demanded will be allocated to A if WAM is greater than available resources and UP is greater than zero. If Zero resources are available, then the tenant has to wait till some tenant preempt its resources. Heuristic tale should be updated with every transition done.

1: RA: Available Resources in Cluster. 2: RA’= (RA1 ...RAn) Resources Allocation. RAi denotes resource allocation for client i. 3: RD =(RD1...RDn) resources demanded by tenants. rdi denotes resources demanded by client i. 4:W=(W1 ...Wn): Weight of workloads, Wi denotes the weight for client i. 5:if RDi less than RAi then 6:UAi <- RDi #Allocate demanded resources 7:UPi <- RAi – RDi #Resources Preempted 8:WAM <- ∑ Wi*RDi /∑ Wi 9:while : execute pending tasks 10: if WAM < RA && UP >0 then 11: UAi <- RDi 12: if WAM ≈ RA && UP >0 then 13: UAi <- RDi%50 14: if WAM > RA && UP >0 then 15: UAi <- RDi%25 16: else 17: Wait until there is a released resource ri from client i 18: Update heuristic table for client i.

4. Evaluation We presented the assessment of HRSYARN on 10Node cloud cluster. It consists of 2 Dell R720 rack servers. The specification of each server is Intel dual quad-core Xeon 5340 processor with 48 GB RAM. And built-in quad port Ethernet cards with 10Gbps bandwidth. VMware Esxi 5.1 is installed to create virtual machines on the top these servers. We used 10 node Hadoop Cluster for running workload performance with different schedulers. These VMs are created with 2 core with 8GB RAM, 200GB Storage [9]. We installed Stable Hadoop version 2.7.2 and each VM installed with CentOS 6.8. Out of which one node configure as Resource Manager with NameNode [7]. The rest of the 9 nodes are running with Node Manager and DataNode. These nodes are treated as slave nodes to store data over HDFS and run MapReduce applications. The chunk size of the HDFS is set to 64MB by default. The resource sharing in Hadoop is dynamically allocated based on heuristics table. To check the performance of HRSYARN, compared with LTYARN [10]. It improves job completion time by 30% compare to other schedulers. Map Reduce Benchmarks

The performance of the canonical MapReduce benchmarks taken from the MR literature was evaluated. These benchmarks will represent a range of different

International Journal of Pure and Applied Mathematics Special Issue

498

Page 9: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

MapReduce resource usage profiles. In Table 1 the summary of each benchmark is provided. The workloads of different benchmarks like Word Count, TeraGen, TeraSort and PI. Input sizes of the workloads are chosen based on the cluster resource availability. Here, Hadoop YARN Capacity, FIFO, LTRF and HRSYARN Scheduler to evaluate the performance of different workloads.

Table 7: Benchmark Details

Table 8: The Job Information for MapReduce Benchmarks – LTYARN

Figure 3: Map Reduce Average Execution Time in ms- LTYARN

Figure 4: Map Reduce Average Memory Usage in MB – LTYARN

International Journal of Pure and Applied Mathematics Special Issue

499

Page 10: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

Table 9: The Job Information for MapReduce Benchmarks – HRSYARN

Figure 5: Map Reduce Average Execution Time in ms- HRSYARN

Figure 6: Map Reduce Average Memory Usage in MB – HRSYARN

Figure 7: Map Reduce Slots–Workloads

The experimental results derived by applying HRSYARN on different application like PI, Grep Search, User Visits Aggregation, Teragen, WordCount and Terasort are exhibited in Table 8 and Table 9. These results prove that HRSYARN succeeded in utilizing the resources efficiently and in attaining fairness in resource sharing. It produced 30% better results over LTYARN in executing map reduce programs in terms of average time and memory usage

International Journal of Pure and Applied Mathematics Special Issue

500

Page 11: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

5. Conclusion This paper studies the Inter-Tenant resource fairness for YARN in heterogeneous environment. In existing Long term resource fairness policy used in YARN, only 50% of the resources are returned in cloud computing system. Because of its 50% resource return policy there is a serious problems, i.e., Tenants may not show interest to share resources to the other tenants though resources available freely. To address this problem in YARN, this paper proposed weighted arithmetic mean based resource allocation. The experimental results showed that this scheduler is best fit in hadoop cloud cluster environments. HRSYARN is a pluggable scheduler for YARN to bring fairness in resource sharing among tenants. The experimental results validated the effectiveness of HRSYARN which yielded 30% better results over LTYARN in executing map reduce programs in terms of average time and memory usage.

References [1] Yang L., Yukun Zeng, Xuefeng Piao, High-Responsive

Scheduling with Map Reduce Performance Prediction on Hadoop YARN, IEEE 22nd International Conference on Embedded and Real-Time Computing Systems and Applications (2016).

[2] Niu Z., Tang S., He B., Gemini: An adaptive performance fairness scheduler for data-intensive cluster computing, 7th International Conference on Cloud Computing Technology and Science (CloudCom) (2015), 66-73.

[3] Tang S., Lee B.S., He B., Fair resource allocation for data-intensive computing in the cloud, IEEE Transactions on Services Computing (2016).

[4] Interlandi M., Tetali S.D., Gulzar M.A., Noor J., Condie T., Kim M., Millstein T., Optimizing interactive development of data-intensive applications, Proceedings of the Seventh ACM Symposium on Cloud Computing (2016), 510-522.

[5] Cassales G.W., Charão A.S., Kirsch-Pinheiro M., Souveyet C., Steffenel L.A., Improving the performance of Apache Hadoop on pervasive environments through context-aware scheduling, Journal of Ambient Intelligence and Humanized Computing 7(3) (2016), 333-345.

[6] Tang S., Lee B.S., He B., Liu H., Long-term resource fairness: Towards economic fairness on pay-as-you-use computing systems, Proceedings of the 28th ACM international conference on Supercomputing (2014), 251-260.

International Journal of Pure and Applied Mathematics Special Issue

501

Page 12: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

[7] Cheng D., Rao J., Guo Y., Jiang C., Zhou X., Improving performance of heterogeneous map reduce clusters with adaptive task tuning, IEEE Transactions on Parallel and Distributed Systems 28(3) (2017), 774-786.

[8] Niu Z., Tang S., He B., An Adaptive Efficiency-Fairness Meta-scheduler for Data-Intensive Computing, IEEE Transactions on Services Computing (2016).

[9] Tang S., Lee B.S., He B., Fair resource allocation for data-intensive computing in the cloud, IEEE Transactions on Services Computing (2016).

[10] Du J., Yu C., Sun J., Sun C., Tang S., Yin Y. Easy HPS: A multilevel hybrid parallel system for dynamic programming. In 27th International Parallel and Distributed Processing Symposium Workshops & PhD Forum (IPDPSW) (2013), 630-639.

[11] Jian, D., Peng, Y., Research of Performance of Distributed Platforms Based on Clustering Algorithm, JCP 11(3) (2016) 195-200.

[12] Benifa J.B. Performance Improvement of MapReduce for Heterogeneous Clusters Based on Efficient Locality and Replica Aware Scheduling (ELRAS) Strategy, Wireless Personal Communications (2017), 1-25.

[13] Liu Y., Zeng Y., Piao X., High-Responsive Scheduling with MapReduce Performance Prediction on Hadoop YARN, 22nd International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) (2016), 238-247.

[14] Liu Z., Zhang Q., Ahmed R., Boutaba R., Liu Y., Gong Z., Dynamic Resource Allocation for Map Reduce with Partitioning Skew, IEEE Transactions on Computers 65(11) (2016) 3304-3317.

[15] Mathiya B.J., Desai, V.L. Apache Hadoop Yarn Map Reduce Job Classification Based on CPU Utilization and Performance Evaluation on Multi-cluster Heterogeneous Environment, Proceedings of International Conference on ICT for Sustainable Development, Springer Singapore (2016), 35-44.

[16] Moon T., Smola A., Chang Y., Zheng Z., Interval Rank: isotonic regression with list wise and pairwise constraints, International conference on Proceedings of the third ACM Web search and data mining (2010), 151-160.

International Journal of Pure and Applied Mathematics Special Issue

502

Page 13: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

503

Page 14: Heuristics Based Resource Sharing with Fairness in YARN ... · Heuristics Based Resource Sharing with Fairness in YARN: HRSYARN 1Kiran Kumar Pulamolu and 2D. Venkata Subramanian 1Hindustan

504