A stochastic model for performance evaluation of main memory resident database systems

With the rapid evolution in the computer industry where memory chips of increasing capacity are introduced every year at lower costs coupled with the increasing demand for massive on-line databases, a great deal of research has been done in the area of main memory resident databases. The use of these systems can achieve significant performance improvements in environments with very high volume of transactions by eliminating the need for most I/O functions while performing basic database operations. A primary concern in such systems is how to develop efficient recovery mechanisms to cope with system failures that can erase the contents of the main memory. This paper presents a stochastic model to evaluate the performance of the database system in terms of normal operation, checkpointing and recovery from system failures. Through a multi-variate model important transform results of effective residence times in each of the database states are derived, which allows the computation of first moments of interest, ergodic state probabilities, including the system availability, and reward measures. The model developed in the paper is significantly more general than previous stochastic models for traditional databases thus capturing the special features of recovery of main memory systems. The applicability of the model is supported and verified through numerical experiments.