Survey on Load Rebalancing for Distributed File System in Cloud

International Journal of Innovative Research in Advanced Engineering (IJIRAE)
Volume 1 Issue 2 (April 2014)
___________________________________________________________________________________________________
ISSN: 2278-2311 IJIRAE | http://ijirae.com
© 2014, IJIRAE All Rights Reserved Page -78
Survey on Load Rebalancing for Distributed File System in
Cloud
Prof. Pranalini S. Ketkar Ankita Bhimrao Patkure
IT Department, DCOER, PG Scholar, Computer Department DCOER,
Pune University Pune university
pranalini.ketkar@gmail.com ankitapatkure@rediffmail.com
Abstract— Distributed file system is used as a key building block of cloud computing. In distributed file system, a
large file is divided into number of chunks and allocates each chunk to separate node to perform MapReduce function
parallel over each node. In cloud, if number of storage nodes, number of files and assesses to that file increases then
the central node (master in MapReduce) becomes bottleneck. The load rebalancing task is used to eliminate the load
on central node. Using load rebalancing algorithm the load of nodes is balanced as well as the movement cost is
reduced. In this survey paper the problem of load imbalancing is overcome.
Keywords— DFS, MapReduce, Load balancing, Distributed Hash Table, cloud
I. INTRODUCTION
In Cloud computing, the number of computers that are connected using communication network. The notation
of cloud indicates that internet is mandatory to perform the various cloud operations i.e. to create delete, append and
replace. It is used in IT-companies to share information and resources with the all users. There are various characteristics
of cloud i.e. Scalable, on demand service, User centric, Powerful, Versatile, Platform independent etc. In cloud three
technologies are included the MapReduce programming, Virtualization and distributed file systems for the data storage
purpose.
Distributed file system is classical model of file system that is used in the form of chunks for cloud computing.
Cloud computing application is based on the MapReduce programming used in distributed file system. MapReduce is the
master-slave architecture in Hadoop. Master act like Namenode and Slave act like Datanode. Master takes large problem,
divides it into sub problem and assigns it to worker node i.e. to multiple slaves to solve problem individually. In
distributed file system, a large file is divided into number of chunks and allocates each chunk to separate node to perform
MapReduce function parallel over each node. For example in word count application it identifies the occurrences of each
distinct word in large file. In this application a large file is divided into fixed-size chunks (parts) and assigns each chunk
to different cloud storage node. Then each storage node calculates the occurrences of each distinct word by scanning and
parsing its own chunk. Then give its result to master to calculate the final result. In distributed file system, the load of
each node is directly proportional to number of file chunks that node consists.
As the increase in storage and network, load balancing is the main issue in the large scale distributed systems.
Load should be balance over multiple nodes to improve system performance, resource utilization, response time and
stability. Load balancing is divided into two categories: static and dynamic. In static load balancing algorithm, it does not
consider the previous behaviour of a node while distribute the load. But in case of dynamic load balancing algorithm, it
checks the previous behaviour of node while distribute the load. In cloud, if number of storage nodes, number of files and
assesses to that file increases then the central node (master in MapReduce) becomes bottleneck. The load rebalancing
task is used to eliminate the load on central node. In load balancing algorithm, storage nodes are structured over network
based on the distributed hash table (DHT); each file chunk having rapid key lookup in DHTs, in that unique identifier is
assign to each file chunk [1]. DHTs enable nodes to self-recognize and repair while it constantly offers lookup
functionality in node. Here aim to reduce the movement cost which is caused by load rebalancing of nodes to maximize
the network bandwidth. Each Chunkserver first find whether it is light node or heavy node without global knowledge of
node. The numbers of file chunks are migrated from heavy node to light node to balance their load. This process repeats
until all heavy nodes becomes the light nodes. To overcome this load balancing problem each node perform load
rebalancing algorithm independently without global knowledge about load of all nodes.
II. LITERATURE SURVEY
MapReduce: Simplified Data Processing On Large Clusters [9]
MapReduce is the programming model used in implementation for processing and generating large scale
datasets. It is used at Google for many different purposes. Here map and reduce functions are used. Map function
generate set of intermediate key pairs and reduce function merges all intermediate key values associated with same
intermediate key. The map and reduce function allows to perform parallelize operation easily and re-execute the
mechanism for fault tolerance. At the run-time, system takes care of detail information of partitioning the input data,
schedule the program execution across number of available machines, handling failures and managing inter-
communication between machines.

___________________________________________________________________________________________________
© 2014, IJIRAE All Rights Reserved Page - 79
In distributed file system nodes simultaneously perform computing and storage operations. The large file in
partitioned into number of chunks and allocate it to distinct nodes to perform MapReduce task parallel over nodes.
Typically, MapReduce task processes on many terabytes of data on thousands of machines. This model is easy to use; it
hides the details of parallelization, optimization, fault-tolerance and load balancing. MapReduce is used for Google’s
production Web search service, machine learning, data mining, etc. Using this programming model, redundant execution
used to reduce the impact of slow machines, handle machine failure as well as data loss.
Load Balancing Algorithm for DHT based structured Peer to Peer System [10]
Peer to peer system have an emerging application in distributed environment. As compared to client-server
architecture, peer to peer system improved resource utilization by making use of unused resources over network. Peer to
peer system uses Distributed Hash Table (DHTs) as an allocation mechanism. It perform join, leave and update
operations. Here load balancing algorithm uses the concept of virtual server to temporary storage of data. Using the
heterogeneous indexing, peers balanced their loads proportional to their capacities. In this, decentralized load balance
algorithm construct network to manipulate global information and organized in tree shape fashion. Each peer can
independently compute probability distribution capacities of participating peers and reallocate their load in parallel.
Optimized Cloud Partitioning Technique to Simplify Load Balancing [5]
Cloud computing has some issue regarding resource management and load balancing. In this paper, Cloud
environment is divided into number of parts by making use of cloud cluster technique which helps for the process of load
balancing. The cloud consists of number of nodes and it partitioned into n cluster based on cloud cluster technique. In
this, it consists of main controller which maintains all information regarding all load balancer in cluster and index table.
Initial step is to choose the correct cluster and it follows algorithm as
1) In cloud environment, the nodes connected to central controller are initialized as 0 in index table.
2) When controller receives new request, it queries the load balancer of each cluster for job allocation.
3) Then controller pass index table to find next available node having less weight. If found then continue the
processing otherwise index table reinitialized to 0 and in an increment manner then again controller passes table
to find next available node.
4) After completing the process the load balancer update the status in allocation table.
Cloud partitioning method consist 2 steps:-
1) Visit each node randomly to match it with neighbor node. If it having same characteristics and shares similar
data with minimal cost then two nodes are combined into new node with share same details. Repeat until there is
no neighbor node having similar characteristics. Subsequently update the cost between neighbor two nodes and
current neighbor node.
2) After joining two nodes into new node having similar characteristics visited node send the information to new
node instead of sending it twice.
It gives the high performance, stability, minimum response time and optimal resource utilization.
Histogram-Based Global Load Balancing in Structured Peer to Peer System [11]
Peer to peer system having solution for sharing and locating resources over internet. In this paper there are two
key components. First is histogram manager that maintain histogram that reflect global view of distribution of load.
Histogram stores statistical information about average load of no overlapping groups of nodes. It is used to check
whether node is normally loaded, Light or heavily loaded. Second component is load balance manager that take the
action of load redistribution if node becomes light or heavy. Load balancing manager balance the load statically when
new node joins and dynamically when existing node become light or heavily loaded. The cost of constructing histogram
and maintaining it may be expensive in dynamic system. To reduce the maintaining cost two techniques are used.
Constructing and maintaining histogram is expensive if node join and leave system frequently. Every new node in peer to
peer system find its neighbour node and these neighbour nodes need to share its information with new node to setup
connection. Now the cost of histogram is totally based on histogram update message caused by changing the load of
nodes in the system .To reduce the cost approximate value of histogram is taken.
III. ALGORITHMS FOR LOAD REBALANCE [1]
It contains three different modules.
1) File Allocation

___________________________________________________________________________________________________
2) DHT Devision
3) Load Rebalancing
4) Advantage of Node heterogeneity
File Allocation
In this, large file is partitioned into number of chunks (C1, C2, C3…..Cn) and it allocates to sub servers
(Chunkserver). Here the files can be added, deleted or appended dynamically to sub server. It will help to avoid the data
loss. Fig. [1] Shows that, given large file is dived into number of parts and that parts are distributed over different
Chunkserver.
Chunk server Chunk server
C1 C2 C3
Fig 1: File allocation
DHT Devising
The storage nodes are structured over network based on the distributed hash table (DHT); each file chunk
having rapid key lookup in DHTs, in that unique identifier is assign to each file chunk. DHT guarantees that if any node
leaves then allocated chunks are migrated to its successor; if node joins then it allocate the chunks which are stored in
successor. DHT network is transparent. It specifies the location of chunks that can be integrated with existing distributed
file system where master node manages namespace of file system and mapping of chunks to storage nodes. Fig. 2 shows
the structure of Distributed Hash Table which is visible to all nodes.
put (key, data) get(key) data
........
Fig 2: Distributed Hash Table
Load Rebalancing
First calculate whether loads are light or heavy in each sub server.
A node is Light if, Number of chunk < (1-∆L) A
And a node is heavy if, Number of chunk > (1-∆U) A
Where ∆L and ∆U are the system parameters. All heavy nodes shared its load with light node. If node ‘i’ is the lightly
loaded then it contact to its successor ‘i+1’and migrate its load to its successor and then join instantly to heavy node.
Name node
Default File Default File Default File
Distributed application
Distributed Hash Table
Chunk server Chunk server Chunk server

___________________________________________________________________________________________________
Here node equalization technique is used to reduce latency, overload and resource utilization. The heavily loaded node
transfer requested chunks to lightly loaded node and it traverse through physical network link.
Light Node Heavy Node
Chunkserver1 Chunkserver2 Chunkserver3
Fig 3: Heavy node and light node
Consider example, the capacity of each Chunkserver is 256MB i.e. 3 chunks (each chunk is of 64 MB) then
according to above Fig. 3 and load rebalancing algorithm chunkserver1 is light node having load 192MB (no of chunk <
(1-∆L) A) and Chunkserver 3 is the heavy node having load 320MB (no of chunk > (1-∆U)A). Using the load
equalization technique, heavy node transfers its load to light node i.e. Chunkserver 3 transfer its some load to
Chunkserver 1.
Fig4 (a): Initial load of chunks Fig4 (b): Rebalancing the load
Consider example, in Fig. 4(a) capacity of each node is 5 chunks (A) but here N1 is the light node having 2
chunks and N5 is heavy node consist 10 chunks. Then N1 identifies that N5 is the heavy node then it contact to its
successor i.e. to N2 and transfer its own load to N2. And request to heavy node for some load. Min (Lj-A),A) that much
load is transfer from heavy node to light node. According to formulae 5 chunks are transfer from N5 to N1 (Fig 4(b)).
Advantage of Node Heterogeneity
Nodes on which file is distributed are heterogeneous in nature. According to nodes capacity there is one
bottleneck resource. Consider, capacity of nodes (Cp1 , Cp2, Cp3,...., Cpn). Each node consist approximate number of
file chunks. The load on that node needs to be balanced as follows:
Ai = γ Cpi
where γ is the load per unit capacity of node. And
γ = m/Σ k=1 Cpk
where m is the number of file chunks stored on system.
Central node
K1
K1
K1K2
K3
K2 K3
K4
K4
K2 K4
K3

___________________________________________________________________________________________________
IV. CONCLUSIONS
Cloud computing application is based on the MapReduce programming used in distributed file system. Load
imbalancing Problem mostly occur in dynamic, large-scale and distributed file system. Load should be balance over
multiple nodes to improve system performance, resource utilization, response time and stability.
In the load rebalancing algorithm the load of node is balanced as well as the movement cost also reduced. The load
balancing task performs independently without global knowledge of system. This load balancing algorithm has fast
concurrency rate. Load balancing reduce the load and movement cost while it uses physical network locality and node
heterogeneity.
REFERENCES
1. Hung-Chang Hsiao, Member, IEEE Computer Society, Hsueh-Yi Chung, HaiyingShen, Member, IEEE, and
Yu-Chang Chao, proposed a “Load Rebalancing for Distributed File Systems in Clouds” IEEE transaction on
parallel and distributed systems, vol. 24, no. 5, May 2013
2. J. Dean and S. Ghemawat, “MapReduce: Simplified Data Processing on Large Clusters,” Proc. Sixth Symp.
Operating System Designed Implementation (OSDI ’04), pp. 137-150, Dec. 2004.
3. S. Mohammad, S. Breß, and E. Schallehn, “Cloud Data Management: A Short Overview and Comparison of
Current Approaches,” Proc. 24th GI-Workshop Foundations of Databases (Grundlagen von Datenbanken), 2012.
4. Apache Hadoop, http://hadoop.apache.org/, 2013
5. P.Jamuna and R.Anand Kumar “Optimized Cloud Computing Technique To Simplify Load
Balancing”International Journal of Advanced Research in Computer Science and Software Engineering,
Volume 3, Issue 11,November 2013.
6. Kokilavani .K, Department Of Pervasive Computing Technology, Kings College Of Engineering, Punalkulam,
Tamil nadu “Enhance load balancing algorithm for distributed file system in cloud” International Journal of
Engineering and Innovative Technology (IJEIT) Volume 3, Issue 6, December 2013
7. Hadoop Distributed File System, http://hadoop.apache.org/hdfs/, 2012.
8. Hadoop Distributed File System “Rebalancing Blocks,” http://
developer.yahoo.com/hadoop/tutorial/module2.html#rebalancing,2012
9. J. Dean and S. Ghemawat, “MapReduce: Simplified Data Processing on Large Clusters,” Proc. Sixth Symp.
Operating System Design and Implementation (OSDI ’04), pp. 137-150, Dec. 2004.
10. ChahitaTanak, Rajesh Bharati “Load Balancing Algorithm for DHT Based Structured Peer to Peer
System”International Journal of Emerging Technology and Advanced Engineering (ISSN 2250-2459, ISO
9001:2008 Certified Journal, Volume 3, Issue 1, January 2013)
11. QuangHieu Vu, Member, IEEE, Beng Chin Ooi, Martin Rinard, and Kian-Lee Tan “Histogram-Based Global
Load Balancing in Structured Peer-to-Peer Systems” IEEE transaction on knowledge and data engineering,vol.
21, no. 4, April2009.

Survey on Load Rebalancing for Distributed File System in Cloud

Recommended

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Survey on Load Rebalancing for Distributed File System in Cloud (20)

More from AM Publications (20)

Recently uploaded (20)

Survey on Load Rebalancing for Distributed File System in Cloud