Combining Simulation and Machine Learning for Response Time Prediction for the Shortest Remaining Processing Time Discipline

Data centers have become a large part of our world and infrastructure and now many of them want to provided information to their customers about the response times of their jobs. To this end, it is important to be able to predict what the response times of jobs might be when a data center implements a specific queueing discipline. In this paper, we investigate the feasibility of reliably predicting response times in real time of the single server shortest remaining processing time queue or the G/G/1/SRPT queue. Our proposed prediction methodology is to combine stochastic simulation with supervised learning machine learning algorithms. We consider several forms of state information like the job size or the number of other jobs in the system at the time of arrival. We hope that our methodology can be replicated for prediction purposes in other queueing systems.