Improving replication protocols through priorities

Among other replication control schemes, grid based protocols (S.Y. Cheung et al., 1990; A. Kumar and S.Y. Cheung, 1991) generally exhibit a property called true distribution. This term was first introduced by M. Maekawa (1985) in the scope of solving the mutual exclusion problem within a distributed networking environment. Informally, a protocol is truly distributed if no site participating in the execution of the protocol bears more responsibility than another one. Through a truly distributed protocol, it is therefore guaranteed that no site can constitute an availability or performance bottleneck for the entire system. We present two grid based replication control schemes which sacrifice true distribution. The motivation for doing so stems from the following observation: distributed networking environments of today do not consist of homogeneous components with identical characteristics. The opposite is the case: high performance multiprocessor workstations coexist with ordinary workstations and even low budget personal computers. In such an environment, having a proper system administration, high performance server machines can process a bigger portion of the workload than others without becoming a performance bottleneck. Furthermore, server machines can be regarded being higher available than other machines: either they are equipped with special components like mirror disks and uninterruptible power supplies, or their down times are shorter while their up times are longer due to better administration and maintenance.