|
Dynamic voting is considered a promising technique for achieving
high availability in distributed systems with data replication.
To date, stochastic analysis of dynamic voting algorithms is restricted
to either site or link Markov models, but not both, possibly because of
the large state space which grows exponentially as the number of sites
increases. Furthermore, to reduce the state space, the assumption of
``frequent updates'' was normally made, which results in an overestimation
of the availability.
In this thesis, we develop a Petri net model that considers both site
and link failures and also relaxes the modeling assumption of frequent updates
We test our Petri net model on some network topologies
to (a) analyze if data availability under dynamic voting
can be seriously degraded if updates are not frequent enough under
various site and link failure/repair situations
and (b) determine the maximum achievable improvement in data
availability when null updates are introduced to augment regular updates
to keep the status of availability up-to-date.
Th
is thesis also investigates the effect of re
pair dependency which occurs when sites and links may have to share the sa
me repairman. Four re
pairman models are examined in the thesis: (a) independent re
pairman with one repairman assigned to each link and each node; (b
) dependent repairman with FIFO servicing discipline; (c
) dependent repairman with linear-order servicing discipline; an
d (d) dependent repairman with best-first servicing discipline. Us
ing dynamic voting as a case study, we compare and contrast the resulting av
ailabilities due to the use of these four different repairman models an
d give a physical interpretation of the differences.
|