Abstract
Often, many software systems fail to meet requirements because of a lack of performance. A proven method for preventing or for diagnosing performance problems is through modeling. Layered Queueing Networks (LQN) are one popular technique for solving performance models. However, if a LQN is solved through decomposition and Mean Value Analysis (MVA), erroneous results can arise because of traffic dependencies in the decomposed models. This paper addresses one traffic dependency, called sub-chains, where customers from one chain "bleed into" another chain causing "extraneous" queueing delays. The new approach described here changes approximate MVA by adjusting the population in a routing chain depending on the originating sub-chain. This new approach substantially reduces, or even eliminates, the extraneous queueing delay caused by the sub-chain dependent traffic. The approach was applied to a substantial model of an on-line bookstore, and reduced the overall error in queueing time by a factor of 20 times, when compared to simulation. The more accurate queueing estimates yield better results for the other outputs of the LQN solver.