Challenges and New Avenues in Existing Replication Techniques
Furat F. Altukhaim and Almetwally M. Mostafa
Information Systems, King Saud University, Riyadh, Saudi Arabia
Keywords: Load Balance, Replication, Replication Challenges, Active Replication, Primary-Backup Replication, Chain
Replication, Mencius, Egalitarian Paxos, Object Ownership Distribution.
Abstract: Over recent years, the curve of the importance of data replication has risen steeply owing to the fact that
databases are increasingly deployed over clusters of different workstations over time. A variety of
replication techniques have been introduced to the distributed systems field which, in this paper, are
classified based on whether they have an unbalanced load between servers or not (classic and modern).
Replication techniques from both categories can be enhanced by avoiding some of the challenges that are
illustrated in detail in this paper. Moreover, this paper analyses replication techniques in each category by
exploring their strengths and weaknesses as well as providing possible novel solutions that can diminish or
eliminate these challenges and introduces a brief description of the Dynamic Object Ownership Distribution
Protocol that aims at increasing throughput by increasing the rate of performing transactions locally in
addition to viewing a promising preliminary results of its performance.
1 INTRODUCTION
We have witnessed a noticeable increase in attention
paid to the variety of replication techniques since the
world has now become a small village which makes
the need for data to be deployed over the whole
globe. Systems use redundant data through the
utilization of replication techniques. Replication can
be defined as creating two or more copies of a data
object with the intention of providing high
availability, consistency, and fault-tolerance. These
replicated data objects are stored at various servers
to allow accessibility by clients in cases where a
server is pulled away either by obligation or as an
option (Özsu and Valduriez, 2011, Charron-Bost et
al., 2010, Mostafa and Youssef, 2014b).
Replication techniques can be categorized into 1)
classic replication techniques which include Active
Replication when it is used by Paxos, the Primary-
Backup Replication which uses Passive Replication,
and Chain Replication which is derived from the
Primary-Backup Replication; and 2) modern
replication techniques which include Mencius,
Egalitarian Paxos, Object Ownership Distribution
Replication, and others. In this paper, when we
mention Paxos, we mean the replication process in
Paxos that uses Active Replication. Active
Replication (State Machine Replication), is a
technique that has been massively used to implement
critical systems such as data stores and coordination
services and in Internet scale infrastructures such as
Yahoo, Google, and MSN. It works by making all
the servers in the system execute the same set of
operations in the same order which requires that the
process hosted by the servers to be deterministic
(Charron-Bost et al., 2010, Schneider and Zhou,
2005, Sousa and Bessani, 2012, Dettoni et al., 2013).
In passive replication, there is one server, the
primary server (sequencer), which acts as a single
organizer for other servers in the system because it
executes operations and propagates the new state to
them (Rao, 2008, Cecchet et al., 2008, Lang et al.,
2010, Effatparvar et al., 2010, Budhiraja et al., 1993,
Mostafa and Youssef, 2014a).
Classic replication techniques have similar issues
in that they have a server that is a bottleneck, which
means that management responsibilities are
cohesively done by a single server and not
distributed between servers in a cell as effectively as
they could be. This leads to an imbalance in the
communication pattern, which in turn leads to
limiting the available network bandwidth. In
addition, in the case of a leader or primary server
failure, re-electing a new one can cause performance
degradation. This is because in situations where a
leader or primary server needs to be pulled out of the
system for any reason, it is not possible to perform
this operation until the server is shut down or
Altukhaim, F. and Mostafa, A.
Challenges and New Avenues in Existing Replication Techniques.
In Proceedings of the 6th International Conference on Cloud Computing and Services Science (CLOSER 2016) - Volume 1, pages 147-154
ISBN: 978-989-758-182-3
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
147
another protocol that is designed to do this particular
job is used, such as the Primary Replacement
Protocol (Mostafa and Youssef, 2014b) and the
Primary Shift Protoco (Mostafa and Youssef, 2013).
Thus, without a structured solution programmed to
be a part of the replication protocol itself, the
reconfiguration process to the leader or primary
server may not be easy or effective. On the other
hand, modern replication techniques are focusing on
distributing responsibilities of the single point of
failure that the classic ones struggle with.
There are many issues that can be found in a
replication cell which affect the cell’s performance.
For instance, having an unbalanced load distribution
between servers in a single cell due to the lack of an
inherent load balancing technique; the lack of
latency awareness between these servers and their
clients and between the servers themselves; poor
reconfiguration planning; the absence of techniques
that allow the system to be aware of the resources
capabilities ( such as CPU or memory) of its servers;
and the inability of the system to be aware of
external environmental factors such as network
changes and clients’ behavior.
The rest of the paper is structured as follows:
Section 2 discusses, reviews, and analyzes classic
replication techniques and shows a detailed
discussion regarding challenges that these
techniques struggle with in addition to comparing
their strengths and weaknesses. Section 3 discusses
the same aspects that have been discussed in section
2 but with modern replication techniques. Section 4
proposes suggestions that could be utilized to
improve replication techniques. Section 5 proposes
the Dynamic Object Ownership Distribution
Protocol and shows some of its preliminary results.
Section 6 concludes the paper.
2 DESCRIPTION AND
CHALLENGES OF CLASSIC
REPLICATION TECHNIQUES
This section reviews classic replication techniques
including Paxos, the Primary-Backup Replication,
and Chain Replication, and some of the challenges
associated with them.
2.1 Paxos
Paxos is a consensus protocol that uses state
machines in its replication process (Bolosky et al.,
2011, Lamport, 1998, Lamport, 2001, Lampson,
2001). It is the first example that comes to mind
when thinking of a consensus protocol. It results in
an agreement on the order of inputs between several
servers, even if some of the servers crash and restart
or the minority of them fail permanently. More
details about Paxos in the paper (Tan et al., 2014).
The Chubby lock service for loosely coupled
distributed systems (Burrows, 2006) and Spanner as
Google’s globally distributed database (Corbett et
al., 2013) are services that utilize Paxos as their core
replication protocol.
The challenge that Paxos is facing is that the
server leading the backups (followers) does all the
management work without including any of the
other servers to do that work. If the leader fails, then
one of the backups is chosen, depending on a
consensus decision to be the next leader. During the
election process, any operations that are trying to
reach this data partition need to stop working for a
while until a new leader is elected (throughput drops
to zero).
2.2 The Primary-Backup Replication
The Primary-Backup Replication (Cecchet et al.,
2008, Lang et al., 2010, Effatparvar et al., 2010,
Budhiraja et al., 1993, Mostafa and Youssef, 2014a,
Schneider and Zhou, 2005) has a single server which
is the primary one that is responsible for organizing
and managing the locking/unlocking operations of
objects. It is exclusively designated to do this job to
make sure that consistency and serialization are
applied all the time. Any other server in the system
is a backup and does not have any of the
management responsibilities. One of the services
that sometimes utilizes the Primary-Backup
Replication is Zookeeper (Hunt et al., 2010).
The challenge that the Primary-Backup
Replication has is similar to the one that we have
discussed with Paxos. It includes an unfair
distribution of the management responsibilities
between servers in a cell in addition to having
difficulties when applying the election process
which includes dropping the throughput to zero.
2.3 Chain Replication
It is one of the Primary-Backup Replication
approach forms that improves throughput and
availability (Van Renesse and Schneider, 2004). In a
paper entitled "Chain Replication for Supporting
High Throughput and Availability", the two
researchers, Can Renesse and Schneider, described
the Chain Replication process to be a group of
CLOSER 2016 - 6th International Conference on Cloud Computing and Services Science
148
various master/slave replications where servers that
store replicas of an object are linearly ordered to
form a chain. All decisions about object updates are
made by the head of the chain in a strict order and
these decisions are propagated down the chain. All
read-only queries are processed by the tail of the
chain which is the server that is positioned last. One
of the services that utilizes Chain Replication is
Hibari (Fritchie, 2010). Figure 1 portrays how Chain
Replication works.
The challenge that Chain Replication is facing is
similar to the one that we have discussed in Paxos
and the Primary-Backup Replication before. It
includes not having a fair distribution of the
management responsibilities between servers in its
cell, since the head server is responsible for the
update operations and the tail server for the read
operations. In addition, there is a process that is
employed in Chain Replication called the master
process which deals with a server failure depending
on the position of the failing server. So, it is difficult
to apply reconfiguration to the head or tail server
and having to apply the master process without
preventing clients from accessing the data partition
for a while.
The reason for these issues in these three
replication protocols is the lack of an inherent load
balancing technique that is built into the replication
protocol itself. Table 1 shows a comparison between
the classic replication techniques.
Figure 1: “Update request” and “Read request” in Chain
replication (Fritchie, 2010).
3 DESCRIPTION AND
CHALLENGES OF MODERN
REPLICATION TECHNIQUES
This section reviews modern replication techniques
including Mencius, Egalitarian Paxos, and Object
Ownership Distribution Replication and some of the
challenges associated with them. Their main
objective is to improve a cell's throughput by
distributing management responsibilities among
servers.
Table 1: Comparison between Paxos, Primary-Backup Replication, and Chain Replication.
Type of
Replication
Important Strengths Important Weaknesses
Paxos
- It is preferable when dealing with
byzantine faults.
- Disseminating updates is a
parallel process.
- The need for a leader election protocol.
- No consensus can be made during the leader election process.
- Most of the real-world servers are non-deterministic.
- Determinism requires ordering which makes the process harder.
- Cell’s throughput equals the throughput of the leading server
(bottleneck problem).
Primary-
Backup
Replication
- Cheap.
- Disseminating updates is a
parallel process.
- The need for a leader election protocol.
- Cell’s throughput equals the throughput of the primary server
(bottleneck problem).
- No transactions can be processed during the primary election
process.
Chain
Replication
- The primary role is split between
two servers (head and tail).
- It guarantees a strong consistency
and high throughput.
- For query requests, it provides
low latency.
- The need for the master process.
- For update requests, it provides high latency.
- Cell’s throughput equals the head’s throughput (write) and the
tail’s throughput (read).
- Disseminating updates is a serial process.
- During the master process run, specific transactions cannot be
processed after the failure of the head or tail server.
Challenges and New Avenues in Existing Replication Techniques
149
3.1 Mencius
Numerous protocols extend Paxos for the purpose of
improving consensus performance, two of its
variants being Fast Paxos (Lamport, 2006), which
reduces latency, and CoRefP (Dobre et al., 2006),
that simultaneously runs Paxos and Fast Paxos to
improve the problem of collisions. Neither of these
two variants are the most suitable choices when
thinking of wide-area applications. A better solution
can be found in Mencius (Mao et al., 2008) which is
a multi-leader state machine replication protocol that
comes originally from Paxos.
The idea of Mencius is to distribute the sequence
of the instances of the consensus protocol between
servers in a cell. For instance, if we have five servers
in a cell, the first server is responsible for instance 0,
5, 10 and so on, the second server is responsible for
instance 1, 6, 11, and so on, the third server is
responsible for instance 2, 7, 12 and so on, and in a
similar fashion with the rest of the servers. Figure 2
clarifies that.
Figure 2: Rotating the leadership between five servers in
Mencius (Wei et al., 2013).
Mencius is designed to eliminate some of the
challenges that Paxos has. First, it decentralizes the
responsibilities between servers by having an inherent
load balancing technique that rotates leadership
between all servers in a cell, thus eliminating the need
for leader election. Second, it lowers latency since
their clients use a local replica, which leads to
increasing latency-awareness between clients and
servers. Third, it smoothes the reconfiguration process
due to the fact that when pulling away a server from a
cell, another server arises to be the new leader. This
can be achieved by having a specific time period for
each server to hold the responsibility of leadership.
However, the challenge is that, at any point in time,
there is still one leader, which can have a negative
effect on performance.
3.2 Egalitarian Paxos
Egalitarian Paxos (Moraru et al., 2013), based on
Paxos, is a new consensus protocol that aims at a)
reaching an optimal latency to commit a command
in wide-area applications, b) having the best possible
load balance between all servers so the throughput
can be increased, c) having a graceful degradation in
performance when some of the servers fail for some
reason. One of the main properties of this protocol is
that there is no server that is designated to be a
leader and clients are able to select which server to
send a command to.
To clarify the approach of Egalitarian Paxos,
Figure 3 portrays an example of command A that is
sent by client C1 to server S1 which is one of five
servers in a cell. S1 then sends a preAccept message
to the majority of servers which represents the
quorum servers to S1. These servers incorporate S1
itself, S2, and S3. After that, S2 and S3 send
acknowledgment messages to S1. Thus, S1 can
commit locally and asynchronously notify the rest of
the servers. In this case, S1 is the leader of command
A.
Figure 3: Executing command A and B in Egalitarian
Paxos (Moraru et al., 2013).
Egalitarian Paxos has an input into eliminating
some of the challenges that classic replication
techniques have. First, as Mencius, it decentralizes
the responsibilities between servers by having an
inherent load balancing technique. This approach
involves having clients choose any server as the
command leader, which eliminates the need for
electing a new leader. Second, it has latency-
awareness between servers in the cell. This is done
by taking into consideration the latency between
servers in a cell so each server knows its quorum
when it needs the majority of servers to execute a
command. Third, it smoothes the reconfiguration
process because the execution of commands is not
disrupted when pulling away a server from a cell
since clients can choose any other server in the cell.
The challenge that this protocol has is that it has to
go through two round trips instead of one when there
are interfering or concurrent commands.
CLOSER 2016 - 6th International Conference on Cloud Computing and Services Science
150
3.3 The Object Ownership Distribution
The Object Ownership Distribution Replication
(Mostafa and Youssef, 2014a) aims to improve
scalability and availability in Primary-Backup
Replication systems. It decentralizes the exclusive
management role which is the responsibility of the
primary server in the Primary-Backup Replication.
This approach utilizes the concept "object
ownership", which can be defined as the exclusive
right a server can have over an object to permit
update transaction. This approach depends mainly
on ownership distribution between servers, so the
server that created the object has the right to own it.
Figure 4 portrays the system architecture for the
Object Ownership Distribution Replication Protocol.
This protocol has also eliminated some of the
challenges faced by classic replication techniques.
First, we can see that it solves one of the
challenges that we have discussed earlier, which is
the lack of an inherent load balancing technique by
having the ownership distribution mechanism that
makes each server responsible for managing some of
the objects in the system so all servers are involved
in the management job. Second, when a server needs
to be reconfigured, all the objects that it owns are
owned by another server and there is no need for an
election process that can affect the performance to
such an extent. The challenge is that the ownership
distribution is static in a dynamic environment,
which is impractical and does not reflect the reality
of the problem.
Table 2 shows a comparison between the modern
replication techniques.
Figure 4: Object Ownership Distribution Replication
system architecture (Mostafa and Youssef, 2014a).
4 OPPORTUNITIES FOR
IMPROVEMENTS IN
REPLICATION TECHNIQUES
The system to which we are trying to provide a
replication technique can have several characteristics
added to either the classic or modern replication
techniques to improve their performance. One way
to improve classic replication techniques is by
injecting an inherent load balancing technique to
them. This contributes to eliminating the bottleneck
issue caused by having a single point of failure.
Partitioning is one technique that has been utilized to
solve scalability challenges; this technique started
the problem of having imbalanced distributed
systems. More information about partitioning is
available in (Quamar et al., 2013; Ishikawa, 2013).
Table 2: Comparison between Mencius, Egalitarian Paxos, and Object Ownership Distribution Replication.
Type of
Replication
Important Strengths Important Weaknesses
Mencius
Replication
- No single point of failure since there is no leading server.
- Higher throughput than Paxos since it uses a partitioned leader
scheme.
- Full utilization of bandwidth.
- Lower latency than Paxos when there is a small number of
clients since they can use a local server to be the leader of their
requests.
- There is only one leader at any point in
time.
- A server must hear from all other servers
prior to committing any command.
Egalitarian Paxos
- No single point of failure since there is no leading server.
- Improving load balance which increases throughput and
scalability.
- It takes two round trips to deal with
interfering or concurrent commands.
Object Ownership
Distribution (OOD)
Replication
- No single point of failure since there is no primary server.
- Improving load balance which increases throughput and
scalability.
- Static ownership.
Challenges and New Avenues in Existing Replication Techniques
151
So, when using replication techniques, each data
partition needs to be replicated into a number of
servers (normally 3 or 5 servers) which form a
replication cell. In cells where there is a high load
and great pressure is placed on the leader or primary
server, this definitely impairs load balancing within
a cell since this server has to process all clients’
requests (Mao et al., 2008).
Latency-awareness can contribute in lowering
latency which is an issue that is barely investigated
by researchers when designing a replication
protocol. It means that taking into consideration the
latency between a client and server, a client can be
aware of the closest server to it and then sends it a
request. In Mencius, latency is reduced by using a
local server to be the request leader. Latency-
awareness can be improved between servers
themselves, as in Egalitarian Paxos. Each server in a
cell chooses its quorum based on the lowest latency
between itself and the rest of the servers.
Reconfiguration (Aguilera et al., 2010) is an
important aspect to which some replication protocols
pay little attention. Whether planned or unplanned, it
needs to be dealt with practically without having to
stop accessing the data partition, when dealing with
clients’ requests, as this slows its performance.
Classic replication techniques struggle with this
issue. However, Modern replication techniques ease
the process of reconfiguration by eliminating the
need for electing a new leader or primary server.
One aspect that the available replication
protocols can take into consideration is the process
of testing the resources capabilities of each server
and giving them responsibilities depending on the
results of these tests, thereby improving performance
in a cell.
Another aspect to consider is assigning
responsibilities to servers based on network changes
or the continuously changing number of clients’
requests that each server receives.
Table 3 summarizes issues that we have
discussed in this section, whether addressed or not,
by the six common replication techniques that we
have reviewed earlier in this paper.
5 THE DYNAMIC OBJECT
OWNERSHIP DISTRIBUTION
PROTOCOL
The Dynamic Object Ownership Distribution
Protocol is a fault-tolerant protocol that aims to
improve the operations load balance in scalable
distributed systems for the purpose of increasing
throughput and lowering latency. While the Object
Ownership Distribution Protocol has a static
ownership mechanism to own objects, this protocol
has a dynamic way to do that. In an environment
where the behavior of clients is extremely dynamic,
it is only convenient to have a dynamic solution to a
problem in a continuously changing environment.
The protocol gains its dynamicity from the fact
that it allows object ownership to change from one
server to another. The change depends on the
number of update operations that are performed by
each server on a certain object. The server that
performs the highest number of operations on this
object has the right to be its owner. The proposed
protocol is also able to handle failures. Once a server
fails, its objects are owned automatically by the
server that has the highest ratio of operations after
the failed server. When a server recovers, a
Table 3: Addressed and unaddressed issues in classic and modern replication techniques.
Classic Replication Techniques Modern Replication Techniques
Issues Paxos
Primary-
Backup
Replication
Chain
Replication
Mencius
Egalitarian
Paxos
Object
Ownership
Distribution
Inherent load balancing techniques No No No Yes Yes Yes
Latency awareness (between a client and a
server)
No No No Yes No No
Latency awareness (between servers themselves) No No No No Yes No
Smooth reconfiguration process No No No Yes Yes Yes
Awareness of the capabilities of servers’
resources
No No No No No No
Reactive nature to network changes No No No Yes No No
Reactive nature to the amount of clients’
requests
No No No No No No
CLOSER 2016 - 6th International Conference on Cloud Computing and Services Science
152
comparison operation must be performed to check
which server has the highest ratio of update
operations between the recovered server and the new
owner.
Figure 5: The throughput graph in the Object Ownership
Distribution Protocol.
Figure 6: The throughput graph in the Dynamic Object
Ownership Distribution Protocol.
In an experiment where the replication factor is
three, each object is owned by a different server, and
each client updates an object that did not create, the
results of the throughput of the Object Ownership
Distribution Protocol and the Dynamic Object
Ownership Distribution Protocol are shown in
Figure 5 and 6. As we can see, the result is in favor
of the protocol that can perform the largest number
of local transactions which is, in this experiment, the
Dynamic Object Ownership Distribution Protocol.
6 CONCLUSIONS
Classic replication techniques suffer mainly from
having the bottleneck issue that makes one server
take on all the management responsibilities, which
lowers throughput and increases latency.
Furthermore, they struggle with performing effective
reconfiguration operations. On the other hand,
modern replication techniques decentralize the
management responsibilities among servers in a cell.
Numerous approaches to improve current replication
techniques have been discussed in this paper. In
addition, the Dynamic Object Ownership
Distribution protocol is briefly discussed.
REFERENCES
Aguilera, M. K., Keidar, I., Malkhi, D., Martin, J.-P. &
Shraer, A. 2010. Reconfiguring Replicated Atomic
Storage: A Tutorial. Bulletin Of The Eatcs, 84-108.
Bolosky, W. J., Bradshaw, D., Haagens, R. B., Kusters, N.
P. & Li, P. Paxos Replicated State Machines As The
Basis Of A High-Performance Data Store. Nsdi, 2011.
Budhiraja, N., Marzullo, K., Schneider, F. B. & Toueg, S.
1993. The Primary-Backup Approach. Distributed
Systems, 2, 199-216.
Burrows, M. The Chubby Lock Service For Loosely-
Coupled Distributed Systems. Proceedings Of The 7th
Symposium On Operating Systems Design And
Implementation, 2006. Usenix Association, 335-350.
Cecchet, E., Candea, G. & Ailamaki, A. Middleware-
Based Database Replication: The Gaps Between
Theory And Practice. Proceedings Of The 2008 Acm
Sigmod International Conference On Management Of
Data, 2008. Acm, 739-752.
Charron-Bost, B., Pedone, F. & Schiper, A. 2010.
Replication: Theory And Practice, Springer.
Corbett, J. C., Dean, J., Epstein, M., Fikes, A., Frost, C.,
Furman, J. J., Ghemawat, S., Gubarev, A., Heiser, C.
& Hochschild, P. 2013. Spanner: Google’s Globally
Distributed Database. Acm Transactions On Computer
Systems (Tocs), 31, 8.
Dettoni, F., Lung, L. C., Correia, M. & Luiz, A. F.
Byzantine Fault-Tolerant State Machine Replication
With Twin Virtual Machines. Computers And
Communications (Iscc), 2013 Ieee Symposium On,
2013. Ieee, 000398-000403.
Dobre, D., Majuntke, M. & Suri, N. 2006. Corefp:
Contention-Resistant Fast Paxos For Wans. Technical
Report, Tu Darmstadt, Germany.
Effatparvar, M., Yazdani, N., Effatparvar, M., Dadlani, A.
& Khonsari, A. Improved Algorithms For Leader
Election In Distributed Systems. Computer
Engineering And Technology (Iccet), 2010 2nd
International Conference On, 2010. Ieee, V2-6-V2-10.
Fritchie, S. L. Chain Replication In Theory And In
Practice. Proceedings Of The 9th Acm Sigplan
Workshop On Erlang, 2010. Acm, 33-44.
Hunt, P., Konar, M., Junqueira, F. P. & Reed, B.
Zookeeper: Wait-Free Coordination For Internet-Scale
Systems. Usenix Annual Technical Conference, 2010.
9.
Ishikawa, K.-I. 2013. Asura: Scalable And Uniform Data
Distribution Algorithm For Storage Clusters. Arxiv
Preprint Arxiv:1309.7720.
Lamport, L. 1998. The Part-Time Parliament. Acm
Transactions On Computer Systems (Tocs), 16, 133-
169.
Challenges and New Avenues in Existing Replication Techniques
153
Lamport, L. 2001. Paxos Made Simple. Acm Sigact News,
32, 18-25.
Lamport, L. 2006. Fast Paxos. Distributed Computing, 19,
79-103.
Lampson, B. The Abcd's Of Paxos. Podc, 2001. 13.
Lang, W., Patel, J. M. & Naughton, J. F. 2010. On Energy
Management, Load Balancing And Replication. Acm
Sigmod Record, 38, 35-42.
Mao, Y., Junqueira, F. P. & Marzullo, K. Mencius:
Building Efficient Replicated State Machines For
Wans. Osdi, 2008. 369-384.
Moraru, I., Andersen, D. G. & Kaminsky, M. There Is
More Consensus In Egalitarian Parliaments.
Proceedings Of The Twenty-Fourth Acm Symposium
On Operating Systems Principles, 2013. Acm, 358-
372.
Mostafa, A. M. & Youssef, A. E. 2013. A Primary Shift
Protocol For Improving Availability In Replication
Systems. International Journal Of Computer
Applications, 72, 37-44.
Mostafa, A. M. & Youssef, A. E. 2014a. Improving
Resource Utilization, Scalability, And Availability In
Replication Systems Using Object Ownership
Distribution. Arabian Journal For Science And
Engineering, 39, 8731-8741.
Mostafa, A. M. & Youssef, A. E. 2014b. Prp: A Primary
Replacement Protocol Based On Early Discovery Of
Battery Power Failure In Manets. Multimedia Tools
And Applications, 1-12.
Özsu, M. T. & Valduriez, P. 2011. Principles Of
Distributed Database Systems, Springer Science &
Business Media.
Quamar, A., Kumar, K. A. & Deshpande, A. Sword:
Scalable Workload-Aware Data Placement For
Transactional Workloads. Proceedings Of The 16th
International Conference On Extending Database
Technology, 2013. Acm, 430-441.
Rao, S. 2008. Distributed Systems: An Algorithmic
Approach. Ieee Distributed Systems Online, 11, 3.
Schneider, F. B. & Zhou, L. 2005. Implementing
Trustworthy Services Using Replicated State
Machines. Security & Privacy, Ieee, 3, 34-43.
Sousa, J. & Bessani, A. From Byzantine Consensus To Bft
State Machine Replication: A Latency-Optimal
Transformation. Dependable Computing Conference
(Edcc), 2012 Ninth European, 2012. Ieee, 37-48.
Tan, Z., Dang, Y., Sun, J., Zhou, W. & Feng, D. 2014.
Paxstore: A Distributed Key Value Storage System.
Network And Parallel Computing. Springer.
Van Renesse, R. & Schneider, F. B. Chain Replication For
Supporting High Throughput And Availability. Osdi,
2004. 91-104.
Wei, W., Tian, H., Fengyuan, G. & Li, X. Q. Fast
Mencius: Mencius With Low Commit Latency.
Infocom, 2013 Proceedings Ieee, 2013. Ieee, 881-889.
CLOSER 2016 - 6th International Conference on Cloud Computing and Services Science
154