The scale and complexity of highly distributed data intensive systems is approaching a point where traditional performance evaluation techniques are becoming difficult to apply. Specifically, use of traditional stochastic performance evaluation methods encounters difficulties in (1) complexity (i.e., scale of the models and intractability of corresponding solution techniques) and (2) parameter estimation (i.e., needed by the models).
In this project we seek to address these two challenges through the use of machine learning techniques. Such techniques have not been traditionally employed in this area, but have emerged recently as a possible direction. We envision that this will lead us not only to better machine learning approaches but will also facilitate merging of machine learning-based techniques with more traditional approaches to performance evaluation, where we anticipate obtaining better results than can be obtained through either approach alone.
The broader impacts of this work will be to enable a deeper understanding of the role, advantages, and limitations of machine learning approaches in performance evaluation of large-scale systems as well as their relationship with more traditional approaches. Broader impact also includes improved interdisciplinary education at the graduate and undergraduate levels and diversity efforts.