Efficient load rebalancing for distributed file system in Clouds

Mr. Mohan S.Deshmukh et al. Int. Journal of Engineering Research and Application www.ijera.com
ISSN : 2248-9622, Vol. 6, Issue 5, ( Part -7) May 2016, pp.20-26
www.ijera.com 20 | P a g e
Efficient load rebalancing for distributed file system in Clouds
Mr. Mohan S. Deshmukh*, Prof. S.A.Itkar**
*(Department Of Computer Engineering. Pune University, ME II Year P.E.S’s Modern College Of Engineering)
** (Department Of Computer Engineering, Pune University, Assistant Professor P.E.S’s Modern College Of
ABSTRACT
Cloud computing is an upcoming era in software industry. It’s a very vast and developing technology.
Distributed file systems play an important role in cloud computing applications based on map reduce
techniques. While making use of distributed file systems for cloud computing, nodes serves computing and
storage functions at the same time. Given file is divided into small parts to use map reduce algorithms in
parallel. But the problem lies here since in cloud computing nodes may be added, deleted or modified any time
and also operations on files may be done dynamically. This causes the unequal load distribution of load among
the nodes which leads to load imbalance problem in distributed file system. Newly developed distributed file
system mostly depends upon central node for load distribution but this method is not helpful in large-scale and
where chances of failure are more. Use of central node for load distribution creates a problem of single point
dependency and chances of performance of bottleneck are more. As well as issues like movement cost and
network traffic caused due to migration of nodes and file chunks need to be resolved. So we are proposing
algorithm which will overcome all these problems and helps to achieve uniform load distribution efficiently. To
verify the feasibility and efficiency of our algorithm we will be using simulation setup and compare our
algorithm with existing techniques for the factors like load imbalance factor, movement cost and network traffic.
Keywords - Clouds, Distributed File system, Load balance, Movement cost, Network traffic
I. INTRODUCTION
Cloud computing is advanced technology in which
dynamic allocation of resources on-requirement
basis is carried out without any specific procedure.
Cloud computing is scalable since it uses key
technologies like MapReduce algorithms [4],
distributed file systems [2], [8], virtualization etc.
Distributed file systems are usually used for cloud
computing, which are based on MapReduce
technology. In this file system, files are divided
into number of small parts and these parts are
allocated to various distinct nodes. Nodes serve
both storage and computing functions. For
example, consider an application where counting a
number of distinct names of the persons in a given
country. Then application will find out distinct
names of the person and also the frequency of each
name. In this type of application, a cloud partitions
the file into fixed size parts and then assigns them
to different
nodes in the system. Then each node will
perform the counting task on part of file stored in
it. But as discussed the nodes may be upgraded,
deleted or added dynamically and also the file
chunks. This leads to load imbalance problem.
Uniform load distribution is challenging task in
cloud computing. In a load balanced environment
performs of the system will improve and we can
achieve high efficiency.
For load balancing mostly used approach is central
node technique [3]. In this technique dependency is
on central node for managing metadata information
of the file systems for balancing loads of storage
nodes. This approach is useful to simplify the
design and implementation of the distributed file
system [2]. But the main concern for this approach
is scalability of cloud computing system. As the
number of nodes, number of files and users
accessing the system increases central load fails
due to overload and performance bottleneck
problems. New developments were carried out to
solve these problems in central node but those
techniques do not serve the purpose successfully.
For example, in Hadoop DFS [8] federation
architecture with multiple namenodes for managing
metadata information is used. In this system
manual and static portioning is carried out [9]. But
since the load of namenodes may change over a
period of time and there is no provision for
migration of load for load balancing, any
namenode with excess load may become a
bottleneck and node without any load will be ideal
at the same time. So this technique also fails to give
uniformly load balance system.
In this paper, we emphasizes on studying
load rebalance problem in distributed file systems
which are large-scale, dynamic and data intensive.
This type of large-scale file systems has hundreds
or thousands of nodes. Our main objective is to
RESEARCH ARTICLE OPEN ACCESS

design a system which will have uniform load
distribution as far as possible. We also focus on
movement cost i.e. the network traffic caused by
nodes in interest of balancing the loads and also
improving the capacity of nodes which will
enhance the performance of the overall system. We
mainly suggest the offloading the load rebalancing
work to storage nodes by availing the storage nodes
to balance their loads spontaneously. For this we
structure the nodes in DHT network format.
Assignment of unique chunk handle for each file
chunk (part) is carried out. DHTs will allow nodes
to self-organize and repair by offering lookup
function in node.
In this paper will cover the following
points. 1st will go through the load rebalancing
problem, then our proposed algorithm for load
rebalancing and then evaluation of algorithm
through computer simulation and assessment of
simulation result. Then using cluster environment
performance measurement of our proposal.
II. WHAT IS LOAD REBALANCING
PROBLEM?
1. Consider a large-scale distributed file system
[8] which consists of set of chunk servers V in
a cloud and cardinality of V is n. n can be ten
thousand or more. Files will be stored on n
chunk servers. Let’s say set of files as F. each
file f belongs to F is partitioned into number of
parts and fixed-size chunks denoted by Cf
.Second load of chunk servers is directly
proportional to number of chunks hosted by
the server. Chunk servers may be replaced,
added or upgraded in the system and also the
files F may be appended, created or deleted at
any time [2]. This in turn affects the load
balancing of system and results in non-uniform
load distribution in the system. Figure 1
illustrates an example of load rebalancing
problem. Assumption here is that chunk
servers are homogenous and have same
capacity.
(1) (2)
(3) (4)
Fig. 1 example of load rebalancing problem, (1)
initial load distribution, (2) files f2 andf5 are
deleted (3) f6 is appended and (4) node 4 joins. The
nodes 1, 2, 3 are in load imbalanced state.
We will focus on designing a load
rebalancing algorithm to reallocate the chunks to
achieve the uniform load distribution to the system
as much as possible. We will also try to reduce the
movement cost i.e. the migration of chunks caused
for balancing loads of chunkservers [3]. Let A be
the ideal number of chunks that any chunkserver i
belongs to V is required to manage in load
balanced state,
Then we aim to minimize the load
imbalance factor in each chunkserver i as follows:
‖Wi- .A‖
(2) Where Wi is load of node and denotes the
absolute value function
The load rebalancing problem is NP- hard.
1st for simplicity we will assume a homogenous
environment, where migration of file chunk
between any two nodes requires a unit movement
cost and each chunk server has the identical storage
capacity. But for practical considerations we will
have to deal with nodes with heterogeneous
capacity and different movement costs.
III. SYSTEM ARCHITECTURE
Organization of chunk servers as DHT network i.e.
implementation of each chunk server using DHT
protocols Chord and Pastry. Partitioning of files
into number of fixed-size chunks, and each chunk
will have a unique chunk identifier known as chunk
handle (SHA1). This hash function returns a unique
chunk identifier for a given file’s path name string
and chunk index. For example, the identifiers of 1st
and 10th chunks of file “/user/sap/tmp/Y.log” are
respectively SHA1 (/user/sap/tmp/Y.log,0)and
SHA1 (/user/sap/tmp/Y.log, 10). Each chunk server
with unique ID will represented as 1, 2, 3…n.
Successor of chunk server will be i+1 and
successor of chunk server n as chunk server 1. To
discover a file chunk, the DHT lookup operation is
performed.

3.1 Adavantages Of Using Dhts
By making use of DHTs, it guarantees that
if any node leaves, then its locally hosted chunks
are reliably migrates to its successor and if node
joins, then it allocates the chunks who’s IDs are
immediately precede the joining node from its
successor [3]. Dependency of our proposal is on
node arrival and departure operations to manage
file chunks among the nodes.
Lookup latency delays can be reduced by
performing discovery of file chunks in parallel. To
further reduce latency delay we can use of DHTs
that offers one-hop lookup delay (Amazon’s
dynamo). We can integrate our proposal to existing
large scale distributed file systems as well. Our
proposal works perfect with both uniform and non-
uniform distribution of nodes & file chunks [10]
[11].
3.2 Load Rebalacning Algorithm:
A given system is said to be in load
balanced state if each chunk server hosts not more
than A chunks. Here, each chunk server 1st
estimates whether it is under loaded or overloaded
without global knowledge. If a given node i departs
and rejoins as successor of another node j, then we
represent node I as j+1, node j’ s original successor
as node j+2, the successor of node j’ s original
successor as node j+3, and so on. If any node in the
system is light node then it search for heavy node
and takes over at most A chunks from the heavy
node.
Table 1 : The Symbols Used In Algorithm
Symbol Description
| · | set cardinality
||·|| absolute value function
V set of chunkservers (storage nodes)
n |V|
m number of file chunks stored in V
O set of heavy (overloaded) nodes
U set of light (underloaded) nodes
A ideal number of file chunks hosted by a
node
Ā i
estimation of A by node i
L i
load (number of file chunks) stored in node
i ∈ V
V” vector containing randomly selected nodes
nv number of vectors collected and
maintained by a node
s |V|
Δ L and
ΔU
parameters identifying light and heavy
nodes
 i
 approximated by node i
Algorithm 1: SEEK(V,ΔL,ΔU): a light node i
seeks
An overloaded node j
Input: vector V = {s samples}, Δ L and ΔU
Output: an overloaded node, j
1. Ā i ← an 1 estimate for A based on { Ā j : j
∈ V};
2. if L i < (1 − Δ L ) Ā i then
3. V ←V ∪{i};
4. sort V according to L j (∀ j ∈ V) in ascending
order;
5. k ← i’s position in the ordered set V;
6. find a smallest subset P ⊂ V such that
(i) L j > (1 + ΔU ) Ā j , ∀ j ∈ P, and
(ii) Pj jL( Ā j ) ≥ k Ā i
7. j ← the least loaded node in P;
return j;
Algorithm 2: MIGRATE(i, j): a light node i
requests chunks from an overloaded node j
Input: a light node i and an overloaded node j
1. if L j > (1 + ΔU ) _ 1 Ā j and j is willing to
share its load with i then
2. i migrates its locally hosted chunks to i + 1;
3. i leaves the system;
4. i rejoins the system as j’s successor by having
5. i ← j + 1;
6. t ← Ā i ;
7. if t >(L j − (1 + ΔU ) Ā i ) then
t ← L j − (1 + ΔU ) _ Ā i ;
8. i allocates t chunks with consecutive IDs from
j;
9. j removes the chunks allocated to i and
renames its ID
In response to the remaining chunks it
manages;
Algorithm 3: MIGRATELOCALITYAWARE (i,
V”): a light
Node i joins as a successor of a heavy node j that is
physically closest to i
Input: a light node i and V = {V1, V2, . . . , V nv
}
1. C ← ∅;
2. for k = 1 to nv do
3. C ← C ∪ SEEK(V k );
4. j ← the node in C physically closest to i;
5. MIGRATE(i, j);
Algorithm 4: SEEKFORHETEROGENEITY (V,
Δ L and ΔU ): a light node i seeks an overloaded
node j in an heterogeneous environment where

nodes have different capacities (here,  i denotes
 approximated by node i)
In our proposal 1st will propose algorithm
in which nodes will have global knowledge of the
system & then in 2nd algorithm nodes without
global knowledge of the system. With the help of
global knowledge if node find that it is light node
then node leaves the system by migrating its locally
hosted chunks to its successor i+1 & then rejoins
quickly to as the successor of heavy node. For
relieving the load of heavy node, light node send
requests as min {Lj - A, A} chunks from heavy
node. That is light node request the exceeded load
from heavy node. After this if heavy node still
remains as heavy node then another light node
performs the same procedure. This process repeats
until the heaviest node doesn’t remains as heavy
node. This process is carried for all heavy nodes
and light nodes for load balancing.
This algorithm in this way helps to
achieve the load balanced state as quick as possible
and also helps to reduce the movement cost since
only the light nodes in the system migrates towards
heavy node.
We can reduce the time complexity of this
algorithm, if every light node knows to which
heavy node it needs to request beforehand and then
parallel load balancing can be done. For this 1st we
sort out the top light nodes and top heavy nodes in
the system, mapping of each light node to heavy
node is done. In this way all light nodes can
concurrently request chunks from heavy node
which will help to reduce the latency of sequential
algorithm to achieve the load-balanced state.
It is not possible to have global knowledge
system in a very large-scale distributed file system
so we propose algorithm which will work in
distributed manner without global knowledge.
Then we try to improve our proposal by taking
advantage of physical network locality to reduce
network traffic. We have to also consider the nodes
with heterogeneous capacity & also the high file
availability is asked from large-scale and dynamic
distributed storage systems where chances of
failure are more. To tackle this issue we try to
maintain the replica of each file chunk.
As discussed earlier, we try to propose a
algorithm where nodes will not have a global
knowledge but it’s a very challenging task, so in
our system we try to solve this problem by creating
a group of nodes by randomly selecting nodes. And
then each node contacts the number of nodes in
group & builds a vector denoted by V. vector
consists of entries, & each entry consists of ID,
network address & load statues of nodes in that
group. Using gossip-based protocol, each node
carries out exchange with its neighbors until its
vector has entries [24]. Then it calculates average
load of each node and consider it as estimation.
Then if node i is light node then it
searches for heavy node for requesting chunks.
Then node i performs sorting of nodes including
itself in its vector and finds its position in sorted
list. Node i find out the overloaded nodes such that
it exceeds the maximum load limit or equal to
maximum load limit. Node i then request the
chunks from heavy node.
But here there might be a case where
different nodes try to share the load of node j, for
this node j offloads its load to randomly selected
node. Also it is possible that number of heavy
nodes selects same node to share their loads. In this
case light node randomly picks up heavy nodes for
reallocation.
(1) (2)
(3) (4)
(5)
Fig. 2. An example illustrating our algorithm,
where (1) the initial loads of chunkservers N1;N2; .
. .;N10, Fig.2.2 N1 creates a sample of the loads of
N1, N3, N6, N7, and N9 for performing the load
rebalancing algorithm, Fig.2.3 N1 leaves and sheds
its loads to its successor N2, and then rejoins as
N9’s successor by allocating AeN1 chunks (the

ideal number of chunks N1 estimates to manage)
from N9, Fig.2.4 N4 collects its sample set
fN3;N4;N5;N6;N7g, and Fig.2.5 N4 departs and
shifts its load to N5, then allocates exceeded
chunks from N6 ad rejoins as successor to N6.
3.3 Use Of Physical Network Locality:
DHT network can be useful to find out
logical proximity. But it’s not useful to find out
physical location using logical proximity in
practical. That is a message travelling between two
neighbors in DHT network may travel a long
physical distance through various physical network
links. So in this case heavy migration will require
for sending message which will induce heavy
network traffic and consumption of network
resources.
To overcome this problem, instead of
creating one vector per algorithmic round, each
light node creates NV vectors using the same
procedure. Then from NV vectors, node i search
NV heavy nodes and then selects physically close
heavy node based on message round-trip delay.
For algorithm 3 demonstration consider the above
example. Let NV=2. Create two sample sets
v1={N1,N3,N6,N7,N9} and
v2={N1,N4,N5,N6,N8}. N1 identifies Node N9
and Node N8 in v1 and v2 respectively. Suppose
N8 is closer than N9 then node i will join as a
successor of N8 & also node i will offload its load
to its successor. Further to reduce network traffic
we can initialize DHT network such that every two
nodes with adjacent
IDs are geometrically close. For this we
use space filling curve, which visits each IP address
and assigns a unique ID to each address such that
geometrically close IP addresses are assigned with
numerically close IDs. For invoking the IDs, we
can use IP address as input to space filling curve.
3.4 Use Of Node Heterogeneity:
Nodes in the system may be having
different capacities in terms of number of file
chunks it can accommodate. Consider the
capacities of nodes as (C1, C2, C3……, n). We
modify our basic algorithm in as each node i
approximates the ideal hosting of file chunks in
load balanced state as follows:
Load per unit capacity is which a node should
manage in load balanced state and which a node
should manage in load balanced state and
m is number of file chunks.
As we know load of node is directionally
proportional to the number of file chunks the node
has stored [8]. So we have taken into consideration
this while designing. To find out average load per
unit capacity we will use gossip-based aggregation
protocol [20] [21]. Our basic algorithm is then
modified taking into consideration of node
heterogeneity.
IV. IMPLEMENTATION
For implementation of our algorithm we
will use computer simulations. For this we will use
chord and gossip-based aggregation protocols [20].
The number of nodes in the system will be n=1,000
and number of file chunks m=10000. Number of
chunks hosted by a node will make use of
geometric distribution. Figures will elaborate this
concept. We have use standalone load balancing
server which will acquire global knowledge of the
file chunks in the system from the namenode.
These namenodes manages the metadata of
complete file system. Now by making use of this
global knowledge, it partitions the nodes as
overloaded nodes and light nodes, then balancer
will randomly pick up the light node and heavy
node and balances their nodes according to our
algorithm. The reallocation will end when there is
no pair of light and heavy node found by balancer.
As we have modified our algorithm to reduce
network traffic, load balancer will try to reallocate
loads of nodes in same rack 1st and if no node in
same rack is found by balancer then it will access
other new rack for reallocation.
4.1 Implementation Result
Implementation results will prove that our
approach is performs well than centralized
approach, since load balancer collects the
information from namenode. This is because each
node randomly selects other node without global
knowledge of the system. Our proposal is
distributed and it is not require gaining global
knowledge of the system.
Movement cost of our proposal will be
very low since we will make use physical network
locality. And making use of this information and
arrangement of nodes accordingly will help to
reduce migration of nodes and file chunks. Also in
our proposal only light nodes will offload its load
to successor to achieve load balanced state will be
low. Because the movement cost require for
reallocation of light nodes will be less.
Messages generated by our algorithm are
within the limits and are less as compared to other
approaches. This will result in less message
overload[24]. In our proposal we are depending on
a chord protocol. Number of operations required
for rejoining and departure in our algorithm are

considerably more than centralized approach but
less than other approaches.
As we have modified our algorithm for
making use of physical network locality, we have
found that our algorithm will work well in load
rebalancing task. Since in our algorithm we will
create groups of nodes with physically close to
each other and while reallocation we will make use
of this information, load balancer will 1st reallocate
the nodes from the same rack then if still require
then only access the other rack for load balancing
[27]. As mentioned by making use of chord ring all
adjacent nodes in the ring are physically close also.
This helps in achieving fast convergence of system
to uniform load balanced state with low network
traffic and less consumption of network resources.
Following figure will elaborate the concept of how
our algorithm will work
Shows the time elapsed of HDFS load balancer and
our proposal
1) Distribution of chunks for HDFS
2) Expected distribution of chunks for our
proposal
V. EXPERIMENTAL SETUP
For implementation we are using HDFS
0.21.0 & for assessment of implementation we are
using load balancer in HDFS.
For demonstration we will use small cluster
environment, which consists of a single dedicated
name node and 25 datanodes,
Software requirement: Ubuntu 10.10
Hardware requirement: Intel core 2 duo
E7400 processor, 3gb RAM (RAM size depends
upon number of file chunks to be processed)
For implementation, a number of clients
are established and these clients will issue request
to the namenode. Requests like create & delete
directories, in our proposal we will set up 6 clients
for generating requests. Further we will limit the
processor cycles available for namenode by
periodically varying maximum processor
utilization. When there will be lower processor
availability then less number of cycles will be
available for namenode to allocate to handle the
clients requests.
We will use maximum 256 chunks
scattered in the file system for connecting all nodes
with 100mbps network. For each execution of
algorithm we will calculate the time required to
complete load balancing, also for load balancer in
HDFS and our proposal. We will perform 20 runs
for a given processor utilization and calculate the
average time required for algorithm execution.
Random sampling of 10 nodes will be carried out.
VI. EXPECTED OUTPUT
Uniform load distribution among all nodes
in cloud.
Reduced movement cost.
Network traffic is reduced.

VII. CONCLUSION AND SUMMARY
Our proposed algorithm will prove as
efficient approach to tackle load rebalancing
problem for large-scale, dynamic distributed file
systems in clouds. Our proposal helps to achieve
the load balanced state and also reduces the
movement cost at far extend, by making a perfect
use of node heterogeneity and physical network
distribution. We will compare our proposal with
existing systems for better assessment of our
proposal, for this we will deal with heavy loaded
nodes. Our proposal has number of enhancements
than the centralized approach or typical distributed
approach which will result in high performance
load balancing technique.
For future work we can consider issues
which require in depth knowledge of issues like
metadata management, file consistency models and
replication strategies.
REFERENCES
[1]. Chung, H. Yi, Shen, Haiying, Chao, Y. Chang, ”Load
rebalancing for distributed file system in clouds” , IEEE
Trans, Software Engineering., volume 24 no. 5, May
2013.
[2]. S. Ghemawat, H. Gobioff, and S.-T. Leung, “The
Google File System,” Proc. 19th ACM Symp. O. S.
Principles, pp. 29-43, October 2003.
[3]. K. McKusick and S. Quinlan, “GFS: Evolution on Fast-
Forward,” Communication. ACM, volume 53, no. 3, pp.
42-49, January 2010.
[4]. J. Dean and S. Ghemawat, “MapReduce: Simplified
Data Processing on Large Clusters,” Proc. Sixth Symp.
O. S. Design and Implementation, pp. 137-150,
December 2004.
[5]. VMware, http://www.vmware.com/, 2012.
[6]. Hadoop Distributed File System,
http://hadoop.apache.org/ hdfs/, 2012.
[7]. HDFS Federation,
http://hadoop.apache.org/common/docs/r0.23.0/hadoop-
yarn/hadoop-yarn-site/Federation.html, 2012.
[8]. I.Stoica, R. Morris, D. Liben-Nowell, D.R. Karger,
M.F. Kaashoek, F. Dabek, H. Balakrishnan, “Chord: A
Scalable Peer-to-Peer Lookup Protocol for Internet
Applications,” IEEE/ACM Transactions. Networking,
volume 11, no. 1, pp. 17-21, February 2003.
[9]. A. Rowstron and P. Druschel, “Pastry: Scalable,
Distributed Object Location and Routing for Large-
Scale Peer-to-Peer Systems,” Proc. IFIP/ACM Int’l
Conf. Distributed Systems Platforms Heidelberg, pp.
161-172, November 2001.
[10]. G. DeCandia, D. Hastorun, M. Jampani, G. Kakulapati,
A. Lakshman, A. Pilchin, S. Sivasubramanian, P.
Vosshall, and W. Vogels, “Dynamo: Amazon’s Highly
Available Key-Value Store,” Proc. 21st ACM Symp.
Operating Systems Principles, pp. 205-220, October
2007.
[11]. A. Rao, K. Lakshminarayanan, S. Surana, R. Karp, and
I. Stoica, “Load Balancing in Structured P2P Systems,”
Proc. Second Int’l Workshop Peer-to-Peer Systems, pp.
68-79, February 2003.
[12]. D. Karger and M. Ruhl, “Simple Efficient Load
Balancing Algorithms for P-to-P Systems,” Proc. 16th
ACM Symp. Parallel Algorithms and Architectures
(SPAA ’04), pp. 36-43, June 2004.
[13]. P. Ganesan, M. Bawa, and H. Garcia-Molina, “Online
Balancing of Range-Partitioned Data with Applications
to Peer-to-Peer Systems,” Proc. 13th Int’l Conf. Very
Large Data Bases, pp. 444-455, September 2004.
[14]. J.W. Byers, J. Considine, and M. Mitzenmacher,
“Simple Load Balancing for Distributed Hash Tables,”
Proc. First Int’l Workshop Peer-to-Peer Systems, pp. 80-
87, February 2003.
[15]. Y. Zhu and Y. Hu, “Efficient, Proximity-Aware Load
Balancing for DHT-Based P2P Systems,” IEEE Trans.
Parallel and Distributed Systems, volume 16, no. 4, pp.
349-361, April 2005.
[16]. H. Shen and C.-Z. Xu, “Locality-Aware and Churn-
Resilient Load Balancing Algorithms in Structured P2P
Networks,” IEEE Trans. Parallel and Distributed
Systems, volume 18, no. 6, pp. 849-862, June 2007.
[17]. H.-C. Hsiao, H. Liao, S.-S. Chen, and K.-C. Huang,
“Load Balance with Imperfect Information in Structured
Peer-to-Peer Systems,” IEEE Trans. Parallel Distributed
Systems, volume 22, no. 4, pp. 634-649, April 2011.
[18]. M.R. Garey and D.S. Johnson, Computers and
Intractability: A Guide to the Theory of NP-
Completeness. W.H. Freeman and Company, 1979.
[19]. D. Eastlake and P. Jones, “US Secure Hash Algorithm 1
(SHA1),” RFC 3174, September 2001.
[20]. M. Jelasity, A. Montresor, and O. Babaoglu, “Gossip-
Based Aggregation in Large Dynamic Networks,” ACM
Trans. Computer Systems, volume 23, no. 3, pp. 219-
252, August 2005.
[21]. M. Jelasity, S. Voulgaris, R. Guerraoui, A.-M.
Kermarrec, and M.V. Steen, “Gossip-Based Peer
Sampling,” ACM Trans. Computer Systems, volume 25,
no. 3, August. 2007.
[22]. C. Guo, G. Lu, D. Li, H. Wu, X. Zhang, Y. Shi, C. Tian,
Y. Zhang , S. Lu, “BCube: A High Performance, Server-
Centric Network Architecture for Modular Data
Centers,” Proc. ACM SIGCOMM ’09, pp. 63-74,
August 2009.
[23]. H. Abu-Libdeh, P. Costa, A. Rowstorn, G. O.Shea, and
A. Donnelly, “Symbiotic Routing in Future Data
Centers,” Proc. ACM SIGCOMM 10, pp. 51-62,
August 2010.
[24]. S. Surana, B. Godfrey, K. Lakshminarayanan, R. Karp,
and I. Stoica, “Load Balancing in Dynamic Structured
P2P Systems,” Performance Evaluation, volume 63, pp.
217-240, March 2006.

Efficient load rebalancing for distributed file system in Clouds

More Related Content

What's hot (17)

Similar to Efficient load rebalancing for distributed file system in Clouds (20)

Recently uploaded (20)

Efficient load rebalancing for distributed file system in Clouds