Summary: | 博士 === 國立成功大學 === 資訊工程學系碩博士班 === 91 === Data replication is a proven technique for improving
data availability of distributed systems. Historically the past research
focused mainly on the development of replicated data management
algorithms that can be proven correct and result in
improved data availability, with the performance issues associated
with data maintenance largely ignored. In this thesis, we
analyze both dependability and performance characteristics
of distributed algorithms for managing replicated data
by developing generic modeling techniques based on Petri nets, with the
goal to identify environmental conditions under which these
replicated data management algorithms can be used to satisfy
system dependability and performance requirements.
First, we investigate an effective technique
for calculating the access time distribution for requests that access replicated
data maintained by the distributed system using the majority voting as a case.
The technique can be used to estimate the reliability of real-time
applications which must access replicated data with a deadline requirement.
Then we enhance this technique to analyze user-perceived
dependability and performance properties of quorum-based algorithms.
User-perceived dependability and performance metrics
are very different from conventional ones in that the dependability and
performance properties must be assessed from the perspective of users
accessing the system.
A feature of the enhanced techniques is that no assumption is made regarding
the interconnection topology, the number of replicas, or the quorum definition
used by the replicated system, thus making it applicable to a wide class of
quorum-based algorithms. Our analysis shows that when the user-perceiveness is
taken into consideration, the effect of increasing the network connectivity
and number of replicas on the availability and dependability properties
perceived by users is very different from that under conventional metrics.
Thus, unlike conventional metrics, user-perceived metrics allow a tradeoff
to be exploited between the hardware invested, i.e., higher network
connectivity and number of replicas, and the performance and dependability
properties perceived by users.
Next we analyze reconfigurable algorithms to determine how often
the system should detect and
react to failure conditions so that reorganization operations can be
performed by the system at the appropriate time to improve the
availability of replicated data without adversely compromising the
performance of the system.
We use dynamic voting as a case study to reveal design trade-offs
for designing such reconfigurable algorithms and illustrate how often
failure detection and reconfiguration activities should be performed, by means
of using dummy updates,
so as to maximize data availability. Dummy updates are system-initiated
maintenance updates that will only update the state
of the system regarding the availability of replicated data without
actually changing the value of replicated data. However,
because of using locks,
dummy updates can hinder normal user-initiated updates during the execution
of the conventional 2-phase commitment (2PC) protocol. We develop a modified
2PC protocol to be used by dummy updates
and show that the modified 2PC protocol
greatly improves the availability of replicated data compared to the
conventional 2PC protocol.
Lastly, we examine the availability and performance characteristics
of replicated data in wireless cellular environments in which
users access replicated data through base stations of the network
as they roam in and out of those base stations.
We address the issues of when, where and how
to place replicas on the base stations by developing
a performance model to analyze periodic
maintenance strategies for managing replicated objects in mobile wireless
client-server environments. Under a periodical maintenance strategy, the
system periodically checks local cells to determine if a replicated object
should be allocated or deallocated in a cell to reduce the access cost. Our
performance model considers the missing-read cost, write-propagation cost and
the periodic maintenance cost with the objective to identify optimal periodic
maintenance intervals to minimize the overall cost. Our analysis results show
that the overall cost is high when the user arrival-departure ratio and the
read-write ratio work against each other and is low otherwise.
Under the fixed periodic maintenance strategy, i.e., the maintenance
interval is a constant, there exists an optimal
periodic maintenance interval that would yield the minimum cost.
Further, the optimal
periodic maintenance interval increases as the arrival-departure ratio and the
read-write ratio work in harmony. We also discover that by
adjusting the periodic intervals dynamically in response to
state changes of the
system at run time, it can further reduce the overall cost
obtainable by the fixed periodic maintenance strategy at
optimizing conditions.
|