, 1997).
In this second case, when a particular replica fails, a working replica takes over the client
connection reusing the same address that was using the failed replica, without needing the
client to reconnect.
The single address approach is mainly used in cluster environments (see Figure 2), and
provides a great failure transparency. However, the cluster front-end becomes a single point
of failure. Moreover, since replicas will have to be placed into the same physical location,
and will be sharing the same local area network, the local area network itself also becomes
a single point of failure. The HA-module we propose could be deployed over cluster installations.
In this case, since just one replica would be interacting with clients, the most convenient
replication model would be the passive one, and with this setting, reliable invocation
protocols are not strictly required. However, if the Web application is replicated over the
Internet, a reliable invocation protocol must be implemented.
Reliable invocation protocols depend on the chosen replication model: active (Schneider,
1990), passive (Budhiraja et al., 1993), semi-active (Powell, 1991), or semi-passive (Defago,
Schiper, & Sergent, 1998).
Pages:
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570