Some performance issues for transactions with firm deadlines

Abstract

We present a performance model for transactions with firm deadlines on a database system that uses locking, but without priority scheduling. Such a system may be a legacy, or bought off-the-shelf. Excluding priority scheduling is also a way of determining how resource and data contention affect deadline misses. The model is used to (a) define a workload number that helps the evaluation of a system design by predicting the stress on it; (b) show that performance is proportional to the cube of transaction length, so transactions should request a minimal number of locks; (c) examine how deadlines should vary with transaction length, thus demonstrating the crucial role of resource contention; and (d) show that execution times and multiprogramming levels can cause a bias only though priority scheduling. We also offer an interpretation of "missed deadlines must be rare" in terms of abort cost.