Machine Learning for the System Administrator

In this paper, we explore how some of the principal ideas in machine learning can be applied to the specific task of monitoring computer systems. We show that machine learning can be a valuable tool for accomplishing everyday administrative tasks, and a worthy addition to the arsenal of tools utilized by the system administrator. Difficulties in utilizing machine learning do not arise from the science or mathematics of the discipline, but rather, from the need to adopt a new world-view and practice in order to use the ideas effectively. We show some examples of the kind of thought and practice necessary to apply machine learning, and discuss its strengths, its weaknesses, and its future potential in the context of system administration. [Note to the Program Committee: This work was inspired by a final paper written by John Orthoefer (Akamai) for a class in machine learning, about the potential for use of machine learning in monitoring. John was planning to be a co-author of this paper, but a recent change in job has left him too busy and his time commitment too uncertain to commit to working on the paper at this time. He says he will be in a better position to perhaps contribute in a month or two. Because involving him in the draft-writing process at this time is impractical, we are submitting this paper ourselves, but if the paper is accepted and John gets more time to work upon it in the summer, we will gladly add him as co-author of this paper.]