Performance of transaction scheduling policies for parallel database systems

The impact of transaction scheduling in the shared-nothing type of architecture is studied. An abstract model of the shared-nothing type of architecture is used. Four transaction scheduling policies are considered. These policies can be broadly divided into two classes: policies that work independent of the current system state and policies that use the current system state information. The authors consider two policies belonging to each category-one policy works independent of the transaction characteristics and the other policy requires transaction size information. The results, obtained via simulation, indicate that system-state-dependent policies provide substantial performance advantages over the other group of policies.<
>