Predicting transaction theorughput in a nuix~(TM) computer system: a case study

摘要

This paper describes (1) assessing the cost of delay inherent in processing orders for telephone service and (2) predicting the number of orders (transactions) that could pass through the system before it shut down for they day. Working strictly from the transactions' statistical characteristics, we discovered that predicting throughput of transactions was equivalent to predicting the wallclock time corresponding to the completion of the day's orders. The prediction depended heavily on a data table of measured delay versus the number of orders waiting for service in the system. The prediction model was flexible but difficult to validate because of the natural dependency between order delay and the number of orders waiting for service. We eventually struck an empirical balance between underfitting the data (to minimize prediction error) and overfitting it (to minimize experimental error). We found that predicting throughput was useful, but there was an intense hidden demand among our colleagues for ways to make the prediction more favorable. We discuss our plans for satisfying that demand.