Description

Data sharing refers to a general distributed architecture for high
performance transaction processing. The nodes of a data sharing
system are locally coupled via a high-speed interconnect and can
directly access all disks and thus the entire database. While concurrency
and coherence control protocols for data sharing have
been discussed in previous work, the important area of recovery
has mostly been ignored. This paper discusses the new problems
for crash and media recovery that have to be addressed in data sharing
systems. Recovery is complicated by dependencies on other
functions such as buffer management and concurrency control.
Furthermore, a global log file is to be constructed where the modifications
of committed transactions are reflected in chronological
order. New logging and recovery protocols are proposed for loosely
coupled data sharing systems that employ the primary copy approach
for concurrency/coherence control. A comparison with existing
data sharing systems shows that our protocols support high
performance during normal processing as well as efficient recovery
to provide high availability.