ARAGÃO, M. J.; http://lattes.cnpq.br/6287951084083955; ARAGÃO, Marcelo Jorge.
Resumo:
The fast dissemination of microcomputers and workstations, together with iheir increasing
processing capacity and the advent of communication networks with higii bandwidth.
have promoted a progressive developineut of distributed applications. Followiug this trend,
severa] enterprises became entirely dependent on their computational systems to provide
their services. This dependence requires that the computational systems remain ahvays in
operation.
The lack of availability of services are usually caused by the shortage of resources or the
occurrence of failure. A strategy to avoid resource shortage can be accomplished through
an adaptive priority-based policy of resource allocation. Failures are inevitable. howevcr,
their consequences can be avoided with pvoper use of fault tolevance techniques. The
combination of these two strategies, use of priorities and use of fault tolerauce technicjues,
enables the development of applications with high availability.
Active replication is a teclmique for fault tolerance that when combined with adaptive
priority-based resource allocation, needs a special treatment if the arrival and the processing
of request occurs asynchronously in the replicas of the server. This treatment consists of
avoidiug, within the context of a group of replicas, cases of Priority Inversion, i.e to avuid
that a high priority request that lias just arrived in a server waits for a low priority request
to release a resource. As the replicas of a replicated server might not observe the same
cases of Priority Inversion, the solution to this problem must be based in the global state
of the replicated server in order to guarantee consistency among ali replicas.
We propose in this work the design and implementation of a middleware that enables
the construction of fault tolerant e-commerce applications that use an adaptive prioritybased
policy for resource allocation. Active replication is implemented t.o tolcrate íaihm>
and Group Priority Inversions are avoided.