Technical Presentations

  • Kristal Sauer and Archana Ganapathi

Using Machine Learning to Characterize, Predict and Optimize System Behavior – Presented by Archana Ganapathi

Abstract: Systems have become very complex and it is often difficult to maintain an accurate model of various system components and their interactions. We use Statistical Machine Learning algorithms to automatically extract relationships between system workload and behavioral metrics. Using Kernel Canonical Correlation Analysis (KCCA) to extract correlations between system workload and observed system behavior, we determine clusters of similar requests that, with high confidence, cause a particular system behavior (e.g. 80% CPU utilization). We can leverage the KCCA-produced model for performance prediction, performance optimization and workload management.

In this talk, I will present results from two use cases of the above methodology. The first set of results are for predicting performance in a commercial decision support database system. The second set of results are for optimizing performance of scientific kernels on a multi-core processor. In both scenarios, we were able to outperform the state of the art technology.

Automatic Workload Evaluation (AWE): Predicting Web 2.0 Workload Behavior – Presented by Kristal Sauer

Abstract: The aim of this project is to use statistical machine learning to predict a system’s performance and resource utilization under changes to the workload or underlying hardware. This could be useful in many scenarios. For instance, in the web service domain, when a company wants to promote a feature of their application and thereby shift the distribution of requests towards a certain type, they would like to have a sense for the resulting performance. Also, when it comes time to upgrade their servers, they would like to predict the system’s behavior on the new hardware. This work is inspired by a study done by Ganapathi et al. (ICDE ’09) in which this approach was utilized for predicting query runtime and resource consumption; the contribution of this work is the application of the predictive framework to the web service domain via analysis of a new Web 2.0 social networking benchmark application called Cloudstone. In previous experiments, we validated that our representation of the input allowed us to recreate the CPU utilization of the system; in this phase of the project, we focus on predicting the performance and resource utilization of a given workload. In this talk, I will describe our framework and present experimental results.

Speaker Details

Kristal Sauer is a second-year graduate student working with David Patterson, Armando Fox, and Michael Jordan in the RAD Lab at UC Berkeley, where she has been awarded the NSF fellowship as well as the Berkeley Chancellor’s Fellowship. She earned BS degrees in Electrical Engineering and Computer Science at the University of Nevada, Las Vegas in 2007. Her research interests include applying statistical machine learning to tackling hard systems problems, especially those related to datacenters. She also enjoys cooking and traveling overseas.

Watch Next